diff --git a/.github/workflows/pycodestyle.yml b/.github/workflows/pycodestyle.yml new file mode 100644 index 0000000..5fc2b6b --- /dev/null +++ b/.github/workflows/pycodestyle.yml @@ -0,0 +1,23 @@ +name: Pycodestyle + +on: [pull_request, workflow_dispatch, push] + +jobs: + build: + runs-on: ubuntu-latest + strategy: + matrix: + python-version: ["3.8"] + steps: + - uses: actions/checkout@v2 + - name: Set up Python ${{ matrix.python-version }} + uses: actions/setup-python@v2 + with: + python-version: ${{ matrix.python-version }} + - name: Install dependencies + run: | + python -m pip install --upgrade pip + pip install pycodestyle + - name: Analysing the code with pylint + run: | + pycodestyle $(git ls-files '*.py') diff --git a/.github/workflows/pydocstyle.yml b/.github/workflows/pydocstyle.yml new file mode 100644 index 0000000..ccb8ad1 --- /dev/null +++ b/.github/workflows/pydocstyle.yml @@ -0,0 +1,23 @@ +name: Pydocstyle + +on: [pull_request, workflow_dispatch, push] + +jobs: + build: + runs-on: ubuntu-latest + strategy: + matrix: + python-version: ["3.8"] + steps: + - uses: actions/checkout@v2 + - name: Set up Python ${{ matrix.python-version }} + uses: actions/setup-python@v2 + with: + python-version: ${{ matrix.python-version }} + - name: Install dependencies + run: | + python -m pip install --upgrade pip + pip install pydocstyle + - name: Analysing the code with pydocstyle + run: | + pydocstyle $(git ls-files '*.py') diff --git a/.github/workflows/pylint.yml b/.github/workflows/pylint.yml new file mode 100644 index 0000000..9639509 --- /dev/null +++ b/.github/workflows/pylint.yml @@ -0,0 +1,24 @@ +name: Pylint + +on: [pull_request, workflow_dispatch, push] + +jobs: + build: + runs-on: ubuntu-latest + strategy: + matrix: + python-version: ["3.8"] + steps: + - uses: actions/checkout@v2 + - name: Set up Python ${{ matrix.python-version }} + uses: actions/setup-python@v2 + with: + python-version: ${{ matrix.python-version }} + - name: Install dependencies + run: | + python -m pip install --upgrade pip + pip install numpy torch scipy + pip install --force-reinstall pylint==2.17.5 + - name: Analysing the code with pylint + run: | + pylint $(git ls-files '*.py') --rcfile .pylintrc --recursive y diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000..359ec78 --- /dev/null +++ b/.gitignore @@ -0,0 +1,174 @@ +# Raw data +raw_data/ +.DS_store +outputs/ +data/massw_* + +# Dev notebooks +dev.ipynb + +# Logging +**/oaib.txt + +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +share/python-wheels/ +*.egg-info/ +.installed.cfg +*.egg +MANIFEST + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.nox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +*.py,cover +.hypothesis/ +.pytest_cache/ +cover/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py +db.sqlite3 +db.sqlite3-journal + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +.pybuilder/ +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# IPython +profile_default/ +ipython_config.py + +# pyenv +# For a library or package, you might want to ignore these files since the code is +# intended to run in multiple environments; otherwise, check them in: +# .python-version + +# pipenv +# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. +# However, in case of collaboration, if having platform-specific dependencies or dependencies +# having no cross-platform support, pipenv may install dependencies that don't work, or not +# install all needed dependencies. +#Pipfile.lock + +# poetry +# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. +# This is especially recommended for binary packages to ensure reproducibility, and is more +# commonly ignored for libraries. +# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control +#poetry.lock + +# pdm +# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. +#pdm.lock +# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it +# in version control. +# https://pdm.fming.dev/latest/usage/project/#working-with-version-control +.pdm.toml +.pdm-python +.pdm-build/ + +# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm +__pypackages__/ + +# Celery stuff +celerybeat-schedule +celerybeat.pid + +# SageMath parsed files +*.sage.py + +# Environments +.env +.venv +env/ +venv/ +ENV/ +env.bak/ +venv.bak/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ +.dmypy.json +dmypy.json + +# Pyre type checker +.pyre/ + +# pytype static type analyzer +.pytype/ + +# Cython debug symbols +cython_debug/ + +# PyCharm +# JetBrains specific template is maintained in a separate JetBrains.gitignore that can +# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore +# and can be added to the global gitignore or merged into this file. For a more nuclear +# option (not recommended) you can uncomment the following to ignore the entire idea folder. +#.idea/ diff --git a/.pylintrc b/.pylintrc new file mode 100644 index 0000000..78d2dee --- /dev/null +++ b/.pylintrc @@ -0,0 +1,2 @@ +[MESSAGES CONTROL] +disable=wrong-import-position,import-error,wrong-import-order,protected-access,too-many-instance-attributes,too-many-arguments,broad-exception-caught,attribute-defined-outside-init,dangerous-default-value,invalid-name diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000..0e259d4 --- /dev/null +++ b/LICENSE @@ -0,0 +1,121 @@ +Creative Commons Legal Code + +CC0 1.0 Universal + + CREATIVE COMMONS CORPORATION IS NOT A LAW FIRM AND DOES NOT PROVIDE + LEGAL SERVICES. DISTRIBUTION OF THIS DOCUMENT DOES NOT CREATE AN + ATTORNEY-CLIENT RELATIONSHIP. CREATIVE COMMONS PROVIDES THIS + INFORMATION ON AN "AS-IS" BASIS. CREATIVE COMMONS MAKES NO WARRANTIES + REGARDING THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS + PROVIDED HEREUNDER, AND DISCLAIMS LIABILITY FOR DAMAGES RESULTING FROM + THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS PROVIDED + HEREUNDER. + +Statement of Purpose + +The laws of most jurisdictions throughout the world automatically confer +exclusive Copyright and Related Rights (defined below) upon the creator +and subsequent owner(s) (each and all, an "owner") of an original work of +authorship and/or a database (each, a "Work"). + +Certain owners wish to permanently relinquish those rights to a Work for +the purpose of contributing to a commons of creative, cultural and +scientific works ("Commons") that the public can reliably and without fear +of later claims of infringement build upon, modify, incorporate in other +works, reuse and redistribute as freely as possible in any form whatsoever +and for any purposes, including without limitation commercial purposes. +These owners may contribute to the Commons to promote the ideal of a free +culture and the further production of creative, cultural and scientific +works, or to gain reputation or greater distribution for their Work in +part through the use and efforts of others. + +For these and/or other purposes and motivations, and without any +expectation of additional consideration or compensation, the person +associating CC0 with a Work (the "Affirmer"), to the extent that he or she +is an owner of Copyright and Related Rights in the Work, voluntarily +elects to apply CC0 to the Work and publicly distribute the Work under its +terms, with knowledge of his or her Copyright and Related Rights in the +Work and the meaning and intended legal effect of CC0 on those rights. + +1. Copyright and Related Rights. A Work made available under CC0 may be +protected by copyright and related or neighboring rights ("Copyright and +Related Rights"). Copyright and Related Rights include, but are not +limited to, the following: + + i. the right to reproduce, adapt, distribute, perform, display, + communicate, and translate a Work; + ii. moral rights retained by the original author(s) and/or performer(s); +iii. publicity and privacy rights pertaining to a person's image or + likeness depicted in a Work; + iv. rights protecting against unfair competition in regards to a Work, + subject to the limitations in paragraph 4(a), below; + v. rights protecting the extraction, dissemination, use and reuse of data + in a Work; + vi. database rights (such as those arising under Directive 96/9/EC of the + European Parliament and of the Council of 11 March 1996 on the legal + protection of databases, and under any national implementation + thereof, including any amended or successor version of such + directive); and +vii. other similar, equivalent or corresponding rights throughout the + world based on applicable law or treaty, and any national + implementations thereof. + +2. Waiver. To the greatest extent permitted by, but not in contravention +of, applicable law, Affirmer hereby overtly, fully, permanently, +irrevocably and unconditionally waives, abandons, and surrenders all of +Affirmer's Copyright and Related Rights and associated claims and causes +of action, whether now known or unknown (including existing as well as +future claims and causes of action), in the Work (i) in all territories +worldwide, (ii) for the maximum duration provided by applicable law or +treaty (including future time extensions), (iii) in any current or future +medium and for any number of copies, and (iv) for any purpose whatsoever, +including without limitation commercial, advertising or promotional +purposes (the "Waiver"). Affirmer makes the Waiver for the benefit of each +member of the public at large and to the detriment of Affirmer's heirs and +successors, fully intending that such Waiver shall not be subject to +revocation, rescission, cancellation, termination, or any other legal or +equitable action to disrupt the quiet enjoyment of the Work by the public +as contemplated by Affirmer's express Statement of Purpose. + +3. Public License Fallback. Should any part of the Waiver for any reason +be judged legally invalid or ineffective under applicable law, then the +Waiver shall be preserved to the maximum extent permitted taking into +account Affirmer's express Statement of Purpose. In addition, to the +extent the Waiver is so judged Affirmer hereby grants to each affected +person a royalty-free, non transferable, non sublicensable, non exclusive, +irrevocable and unconditional license to exercise Affirmer's Copyright and +Related Rights in the Work (i) in all territories worldwide, (ii) for the +maximum duration provided by applicable law or treaty (including future +time extensions), (iii) in any current or future medium and for any number +of copies, and (iv) for any purpose whatsoever, including without +limitation commercial, advertising or promotional purposes (the +"License"). The License shall be deemed effective as of the date CC0 was +applied by Affirmer to the Work. Should any part of the License for any +reason be judged legally invalid or ineffective under applicable law, such +partial invalidity or ineffectiveness shall not invalidate the remainder +of the License, and in such case Affirmer hereby affirms that he or she +will not (i) exercise any of his or her remaining Copyright and Related +Rights in the Work or (ii) assert any associated claims and causes of +action with respect to the Work, in either case contrary to Affirmer's +express Statement of Purpose. + +4. Limitations and Disclaimers. + + a. No trademark or patent rights held by Affirmer are waived, abandoned, + surrendered, licensed or otherwise affected by this document. + b. Affirmer offers the Work as-is and makes no representations or + warranties of any kind concerning the Work, express, implied, + statutory or otherwise, including without limitation warranties of + title, merchantability, fitness for a particular purpose, non + infringement, or the absence of latent or other defects, accuracy, or + the present or absence of errors, whether or not discoverable, all to + the greatest extent permissible under applicable law. + c. Affirmer disclaims responsibility for clearing rights of other persons + that may apply to the Work or any use thereof, including without + limitation any person's Copyright and Related Rights in the Work. + Further, Affirmer disclaims responsibility for obtaining any necessary + consents, permissions or other rights required for any use of the + Work. + d. Affirmer understands and acknowledges that Creative Commons is not a + party to this document and has no duty or obligation with respect to + this CC0 or use of the Work. diff --git a/README.md b/README.md new file mode 100644 index 0000000..347e218 --- /dev/null +++ b/README.md @@ -0,0 +1,193 @@ +![Visualization of Key Ideas through Time](assets/MASSW.svg) +

+ Figure: A visualization of key ideas' evolution through time. Darker points indicate more recent work. +

+ +

+ + arXiv + + + License: CC0-1.0 + + + Pycodestyle + + + Pydocstyle + + + Pylint + +

+ + + +MASSW is a comprehensive text dataset on **M**ulti-**A**spect **S**ummarization of **S**cientific **W**orkflows. +MASSW includes more than 152,000 peer-reviewed publications from 17 leading +computer science conferences spanning the past 50 years. + +## Key features + +- **Structured scientific workflows**. MASSW defines [five core aspects of a scientific workflow](#core-aspects-of-a-scientific-workflow) -- *context*, *key idea*, *method*, *outcome*, and *projected impact*. These aspects align with the typical stages in scientific workflows identified in recent literature. Utilizing LLMs, we systematically extract and structure these five aspects from each publication. +- **Large scale**. MASSW contains the structured scientific workflows and meta-information from over 152,000 peer-reviewed publications, across 17 leading computer science conferences, and spanning the past 50 years. +- **Accuracy**. The coverage and accuracy of MASSW have been validated through comprehensive inspections and comparisons with human annotations and alternative methods. +- **Rich benchmark tasks**. MASSW facilitates multiple novel and benchmarkable machine learning tasks, such as idea generation and outcome prediction. It supports diverse tasks centered on predicting, recommending, and expanding key elements of a scientific workflow, serving as a benchmark for evaluating LLM agents' ability to navigate scientific research. + +## Core aspects of a scientific workflow + +| Aspect | Definition | Example | +|---------------|------------|---------| +| **Context**
*Ask questions, review literature (prior to study)* | The status quo of related literature or reality which motivated this study. This could normally be a problem, a research question, or a research gap that has not been successfully addressed by previous work. | *Making language models bigger does not inherently make them better at following a user's intent, as large models can generate outputs that are untruthful, toxic, or not helpful.* | +| **Key Idea**
*Construct hypothesis (proposed in this study)* | The main intellectual merit of this paper, often in comparison to the context. This could normally be a novel idea or solution proposed in this paper that distinguishes it from what’s already done in literature. | *The authors propose InstructGPT, a method to align language models with user intent by fine-tuning GPT-3 using a combination of supervised learning with labeler demonstrations and reinforcement learning from human feedback.* | +| **Method**
*Test hypothesis (after hypothesis construction)* | The specific research method that investigates and validates the key idea. This could be an experimental setup, a theoretical framework, or other necessary validation methodology to implement and/or evaluate the key idea. | *The authors evaluate the performance of InstructGPT by humans on a given prompt distribution and compare it with a much larger model GPT-3.* | +| **Outcome**
*Interpret results, draw conclusion (after testing hypothesis)* | The factual statement about the study output. This could be the experiment results and any other measurable outcome that has occurred. It marks whether the key hypothesis is testified or not. | *InstructGPT, even with 100x fewer parameters, is preferred over GPT-3 in human evaluations. It shows improvements in truthfulness and reductions in toxic outputs with minimal performance regressions on public NLP datasets.* | +| **Projected Impact**
*Future work (anticipated but not yet done)* | The author-anticipated impact of the work on the field, and potential further research identified by the author that may improve or extend this study. | *Fine-tuning with human feedback is a promising direction for aligning language models with human intent.* | + +## Coverage + +MASSW covers 17 leading computer science conferences, including: + +- Artificial Intelligence: AAAI, IJCAI; +- Computer Vision: CVPR, ECCV, ICCV; +- Machine Learning: ICLR, ICML, NeurIPS, KDD; +- Natural Language Processing: ACL, EMNLP, NAACL; +- The Web & Information Retrieval: SIGIR, WWW; +- Databases: SIGMOD, VLDB; +- Interdisciplinary Areas: CHI. + +## Installation + +```bash +pip install -r requirements.txt +pip install -e . +``` + +## Download MASSW dataset + +To download the dataset, you can use the provided script: + +```bash +python massw/download.py +``` + +Or download the dataset manually through Dropbox links: +[[MASSW dataset (150.6MB)](https://www.dropbox.com/scl/fi/ykkrpf269fikuchy429l7/massw_v1.tsv?rlkey=mssrbgz3k8adij1moxqtj34ie&dl=1)] +[[MASSW metadata +(896.2MB)](https://www.dropbox.com/scl/fi/r2jlil9lj0ypo2fpl3fxa/massw_metadata_v1.jsonl?rlkey=ohnriak63x4ekyli25naajp0q&dl=1)]. + +## Basic Usage + +Search for a specific publication by title: + +```python +from massw.data import load_massw + +massw = load_massw() +massw.search("attention is all you need") +``` + +Output: + +```python +[Workflow(id='618cfba991e011333c932618', title='Cross-Attention is All You Need - Adapting Pretrained Transformers for Machine Translation.', abstract='We study the power of cross-attention in the Transformer architecture within the context of transfer learning for machine translation, and extend the findings of studies into cross-attention when training from scratch. We conduct a series of experiments through fine-tuning a translation model on data where either the source or target language has changed. These experiments reveal that fine-tuning only the cross-attention parameters is nearly as effective as fine-tuning all parameters (i.e., the entire translation model). We provide insights into why this is the case and observe that limiting fine-tuning in this manner yields cross-lingually aligned embeddings. The implications of this finding for researchers and practitioners include a mitigation of catastrophic forgetting, the potential for zero-shot translation, and the ability to extend machine translation models to several new language pairs with reduced parameter storage overhead.', year=2021.0, venue='EMNLP', context='Cross-attention mechanism in the Transformer architecture has been of interest in the context of transfer learning for machine translation, and there exists previous studies of cross-attention when training from scratch.', key_idea='The authors explore the efficiency of fine-tuning only the cross-attention parameters in a translation model where either the source or target language has changed, as compared to fine-tuning all parameters (the entire translation model).', method='A series of experiments were conducted through fine-tuning a translation model on data where either the source or target language has changed.', outcome='The results revealed that fine-tuning only the cross-attention parameters is nearly as effective as fine-tuning all parameters. The process leads to cross-lingually aligned embeddings.', projected_impact='The findings can mitigate catastrophic forgetting, facilitate zero-shot translation, and allow extension of machine translation models to several new language pairs with reduced parameter storage overhead.', data={'id': '618cfba991e011333c932618', 'title': 'Cross-Attention is All You Need - Adapting Pretrained Transformers for Machine Translation.', 'abstract': 'We study the power of cross-attention in the Transformer architecture within the context of transfer learning for machine translation, and extend the findings of studies into cross-attention when training from scratch. We conduct a series of experiments through fine-tuning a translation model on data where either the source or target language has changed. These experiments reveal that fine-tuning only the cross-attention parameters is nearly as effective as fine-tuning all parameters (i.e., the entire translation model). We provide insights into why this is the case and observe that limiting fine-tuning in this manner yields cross-lingually aligned embeddings. The implications of this finding for researchers and practitioners include a mitigation of catastrophic forgetting, the potential for zero-shot translation, and the ability to extend machine translation models to several new language pairs with reduced parameter storage overhead.', 'keywords': [], 'year': 2021, 'authors': [{'id': '64c1ec89688ef0ffd383e8b3', 'name': 'Mozhdeh Gheini', 'org': 'Univ Southern Calif, Informat Sci Inst, Los Angeles, CA 90007 USA', 'org_id': ''}, {'id': '5631ff8c45cedb3399f8b414', 'name': 'Xiang Ren', 'org': 'Univ Southern Calif, Informat Sci Inst, Los Angeles, CA 90007 USA', 'org_id': ''}, {'id': '5616d61c45ce1e5963b0b70e', 'name': 'Jonathan May', 'org': 'Univ Southern Calif, Informat Sci Inst, Los Angeles, CA 90007 USA', 'org_id': ''}], 'references': ['59ae3c152bbe271c4c71e900', '5e296ede3a55ac6ad1e44fd9', '5de240bd3a55ace1893738ef'], 'doi': '', 'venue_id': '5eba7087edb6e7d53c1009a5', 'n_citation': 38, 'venue': 'EMNLP', 'normalized_venue_name': 'EMNLP'}), + Workflow(id='5e5e190393d709897ce45fda', title='Channel Attention Is All You Need for Video Frame Interpolation', abstract='Prevailing video frame interpolation techniques rely heavily on optical flow estimation and require additional model complexity and computational cost; it is also susceptible to error propagation in challenging scenarios with large motion and heavy occlusion. To alleviate the limitation, we propose a simple but effective deep neural network for video frame interpolation, which is end-to-end trainable and is free from a motion estimation network component. Our algorithm employs a special feature reshaping operation, referred to as PixelShuffle, with a channel attention, which replaces the optical flow computation module. The main idea behind the design is to distribute the information in a feature map into multiple channels and extract motion information by attending the channels for pixel-level frame synthesis. The model given by this principle turns out to be effective in the presence of challenging motion and occlusion. We construct a comprehensive evaluation benchmark and demonstrate that the proposed approach achieves outstanding performance compared to the existing models with a component for optical flow computation.', year=2020.0, venue='AAAI', context='Existing video frame interpolation techniques heavily rely on optical flow estimation, which increases model complexity and computational cost and is prone to error propagation in scenarios with large motion and heavy occlusion.', key_idea='The authors propose an end-to-end trainable deep neural network for video frame interpolation that does not require a motion estimation network component. They introduce a feature reshaping operation known as PixelShuffle, combined with channel attention, which replaces the optical flow computation module.', method='The authors assess their new model in conditions with challenging motion and occlusion.', outcome='The newly proposed model demonstrates outstanding performance in comparison with existing models that use an optical flow computation component.', projected_impact=None, data={'id': '5e5e190393d709897ce45fda', 'title': 'Channel Attention Is All You Need for Video Frame Interpolation', 'abstract': 'Prevailing video frame interpolation techniques rely heavily on optical flow estimation and require additional model complexity and computational cost; it is also susceptible to error propagation in challenging scenarios with large motion and heavy occlusion. To alleviate the limitation, we propose a simple but effective deep neural network for video frame interpolation, which is end-to-end trainable and is free from a motion estimation network component. Our algorithm employs a special feature reshaping operation, referred to as PixelShuffle, with a channel attention, which replaces the optical flow computation module. The main idea behind the design is to distribute the information in a feature map into multiple channels and extract motion information by attending the channels for pixel-level frame synthesis. The model given by this principle turns out to be effective in the presence of challenging motion and occlusion. We construct a comprehensive evaluation benchmark and demonstrate that the proposed approach achieves outstanding performance compared to the existing models with a component for optical flow computation.', 'keywords': [], 'year': 2020, 'authors': [{'id': '635100b2ac95af67f73c20e3', 'name': 'Myungsub Choi', 'org': '(Seoul National University)', 'org_id': '62331e330a6eb147dca8a707'}, {'id': '645320c8ca4e0609eedd4813', 'name': 'Heewon Kim', 'org': '(Seoul National University)', 'org_id': '62331e330a6eb147dca8a707'}, {'id': '53f43026dabfaee2a1c9f879', 'name': 'Bohyung Han', 'org': '(Seoul National University)', 'org_id': '62331e330a6eb147dca8a707'}, {'id': '53f7c3f6dabfae938c6d9624', 'name': 'Ning Xu', 'org': 'Amazon Go, Seattle, WA USA', 'org_id': ''}, {'id': '53fd9092dabfae8faa5a9792', 'name': 'Kyoung Mu Lee', 'org': '(Seoul National University)', 'org_id': '62331e330a6eb147dca8a707'}], 'references': [], 'doi': '', 'venue_id': '5eba67f5edb6e7d53c100585', 'n_citation': 221, 'venue': 'AAAI', 'normalized_venue_name': 'AAAI'}), + Workflow(id='5dbc024a3a55ac76d0e3516c', title='Attention Is All You Need for Chinese Word Segmentation', abstract='Taking greedy decoding algorithm as it should be, this work focuses on further strengthening the model itself for Chinese word segmentation (CWS), which results in an even more fast and more accurate CWS model. Our model consists of an attention only stacked encoder and a light enough decoder for the greedy segmentation plus two highway connections for smoother training, in which the encoder is composed of a newly proposed Transformer variant, Gaussian-masked Directional (GD) Transformer, and a biaffine attention scorer. With the effective encoder design, our model only needs to take unigram features for scoring. Our model is evaluated on SIGHAN Bakeoff benchmark datasets. The experimental results show that with the highest segmentation speed, the proposed model achieves new state-of-the-art or comparable performance against strong baselines in terms of strict closed test setting.', year=2020.0, venue='EMNLP', context='Current Chinese word segmentation (CWS) models, which typically employ a greedy decoding algorithm, could be improved in terms of speed and accuracy.', key_idea='This study proposes an attention-only model for CWS, encompassing an attention only stacked encoder, a light decoder, two highway connections for smoother training, and a Transformer variant called Gaussian-masked Directional (GD) Transformer. The encoder uses unigram features for scoring.', method='The proposed model is evaluated using the SIGHAN Bakeoff benchmark datasets.', outcome='The proposed model achieves new state-of-the-art or comparable performance against strong baselines in terms of strict closed test setting, whilst maintaining the highest segmentation speed.', projected_impact=None, data={'id': '5dbc024a3a55ac76d0e3516c', 'title': 'Attention Is All You Need for Chinese Word Segmentation', 'abstract': 'Taking greedy decoding algorithm as it should be, this work focuses on further strengthening the model itself for Chinese word segmentation (CWS), which results in an even more fast and more accurate CWS model. Our model consists of an attention only stacked encoder and a light enough decoder for the greedy segmentation plus two highway connections for smoother training, in which the encoder is composed of a newly proposed Transformer variant, Gaussian-masked Directional (GD) Transformer, and a biaffine attention scorer. With the effective encoder design, our model only needs to take unigram features for scoring. Our model is evaluated on SIGHAN Bakeoff benchmark datasets. The experimental results show that with the highest segmentation speed, the proposed model achieves new state-of-the-art or comparable performance against strong baselines in terms of strict closed test setting.', 'keywords': [], 'year': 2020, 'authors': [{'id': '542e1ff4dabfae48d12468b4', 'name': 'Duan Sufeng', 'org': 'Shanghai Jiao Tong University', 'org_id': '5f71b54b1c455f439fe502b0'}, {'id': '56cb1892c35f4f3c65651db6', 'name': 'Zhao Hai', 'org': 'Shanghai Jiao Tong University', 'org_id': '5f71b54b1c455f439fe502b0'}], 'references': ['53e9acc4b7602d97036a1037', '5c5ce50d17c44a400fc390b8', '53e9b93eb7602d9704528a0e', '53e9b3bcb7602d9703ea4029', '5736974d6e3b12023e63899e', '5cede103da562983788e2880', '5aed14d617c44a4438159341', '5b1642388fbcbf6e5a9b578c', '57a4e91dac44365e35c981c6', '58d82fcbd649053542fd5fde', '5ce2d0bcced107d4c63afa41', '599c7987601a182cd2648373', '53e9bb02b7602d970473d988', '53e9b74ab7602d97042e6482', '5550415745ce0a409eb3a739', '573698636e3b12023e72937b', '59ae3c262bbe271c4c71ef86', '5550446445ce0a409eb4d42e', '53e99cc9b7602d970257bebe', '53e99960b7602d970219efd5', '5a260c2817c44a4ba8a23741', '5a260c8617c44a4ba8a3278b', '53e99a5cb7602d97022c46f3', '53e9a981b7602d97032d97fb', '599c7988601a182cd2648ecb', '5bdc315017c44a1f58a05d1e', '599c795f601a182cd2635762', '53e9b607b7602d970415d4ba', '599c7972601a182cd263e9a2'], 'doi': '10.18653/V1/2020.EMNLP-MAIN.317', 'venue_id': '5eba7087edb6e7d53c1009a5', 'n_citation': 20, 'venue': 'Conference on Empirical Methods in Natural Language Processing', 'normalized_venue_name': 'EMNLP'}), + Workflow(id='15a601e1-fd02-492c-8864-ee8e1f14fd55', title='Attention is All You Need', abstract='The dominant sequence transduction models are based on complex recurrent orconvolutional neural networks in an encoder and decoder configuration. The best performing such models also connect the encoder and decoder through an attentionm echanisms. We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train. Our single model with 165 million parameters, achieves 27.5 BLEU onEnglish-to-German translation, improving over the existing best ensemble result by over 1 BLEU. On English-to-French translation, we outperform the previoussingle state-of-the-art with model by 0.7 BLEU, achieving a BLEU score of 41.1.', year=2017.0, venue='NeurIPS', context='Prior to this work, sequence transduction models were dominated by complex recurrent and convolutional neural networks in an encoder-decoder configuration, with optimal results found through the use of attention mechanisms connecting the encoder and decoder.', key_idea='The paper introduces a new network architecture, the Transformer, that exclusively relies on an attention mechanism, foregoing recurrence and convolutions entirely.', method='Experiments were conducted on two machine translation tasks to validate the proposed network architecture, comparing its quality, parallelism, and training speed against the existing models.', outcome='The Transformer model outperformed existing models in terms of quality and training efficiency, achieving a BLEU score of 27.5 on English to German translation, which improved over the current best ensemble result by over 1 BLEU, and a BLEU score of 41.1 on English to French translation, surpassing the previous single state-of-the-art model by 0.7 BLEU.', projected_impact=None, data={'id': '15a601e1-fd02-492c-8864-ee8e1f14fd55', 'area': 'ml', 'venue': 'neurips', 'title': 'Attention is All You Need', 'year': 2017, 'data': {'id': 2963403868, 'title': 'Attention is All You Need', 'authors': [{'name': 'Ashish Vaswani', 'id': 2171687631, 'org_id': 1291425158}, {'name': 'Noam Shazeer', 'id': 2496873187, 'org_id': 1291425158}, {'name': 'Niki Parmar', 'id': 2625834147, 'org_id': 1174212}, {'name': 'Jakob Uszkoreit', 'id': 2226984371, 'org_id': 1291425158}, {'name': 'Llion Jones', 'id': 2624808693, 'org_id': 1291425158}, {'name': 'Aidan N. Gomez', 'id': 2626656444}, {'name': 'Lukasz Kaiser', 'id': 2948063405, 'org_id': 1291425158}, {'name': 'Illia Polosukhin', 'id': 2519781307, 'org_id': 1291425158}], 'venue': {'name': 'neural information processing systems', 'id': 1127325140}, 'year': 2017, 'n_citation': 4375, 'page_start': '5998', 'page_end': '6008', 'doc_type': 'Conference', 'publisher': 'Curran Associates Inc.', 'volume': '', 'issue': '', 'indexed_abstract': '{"IndexLength":116,"InvertedIndex":{"The":[0,19],"dominant":[1],"sequence":[2],"transduction":[3],"models":[4,23,59],"are":[5],"based":[6,41],"on":[7,52],"complex":[8],"recurrent":[9],"orconvolutional":[10],"neural":[11],"networks":[12],"in":[13],"an":[14,31],"encoder":[15,27],"and":[16,28,49,68],"decoder":[17,29],"configuration.":[18],"best":[20,90],"performing":[21],"such":[22],"also":[24],"connect":[25],"the":[26,88,102],"through":[30],"attentionm":[32],"echanisms.":[33],"We":[34],"propose":[35],"a":[36,111],"novel,":[37],"simple":[38],"network":[39],"architecture":[40],"solely":[42],"onan":[43],"attention":[44],"mechanism,":[45],"dispensing":[46],"with":[47,77,105],"recurrence":[48],"convolutions":[50],"entirely.Experiments":[51],"two":[53],"machine":[54],"translation":[55],"tasks":[56],"show":[57],"these":[58],"to":[60],"be":[61],"superiorin":[62],"quality":[63],"while":[64],"being":[65],"more":[66],"parallelizable":[67],"requiring":[69],"significantly":[70],"less":[71],"timeto":[72],"train.":[73],"Our":[74],"single":[75],"model":[76,106],"165":[78],"million":[79],"parameters,":[80],"achieves":[81],"27.5":[82],"BLEU":[83,112],"onEnglish-to-German":[84],"translation,":[85,99],"improving":[86],"over":[87,94],"existing":[89],"ensemble":[91],"result":[92],"by":[93,107],"1":[95],"BLEU.":[96],"On":[97],"English-to-French":[98],"we":[100],"outperform":[101],"previoussingle":[103],"state-of-the-art":[104],"0.7":[108],"BLEU,":[109],"achieving":[110],"score":[113],"of":[114],"41.1.":[115]}}', 'fos': [{'name': 'artificial neural network', 'w': 0.47565}, {'name': 'bleu', 'w': 0.50846}, {'name': 'parallelizable manifold', 'w': 0.42665}, {'name': 'convolution', 'w': 0.41377}, {'name': 'encoder', 'w': 0.51848}, {'name': 'computer science', 'w': 0.43723}, {'name': 'network architecture', 'w': 0.45739}, {'name': 'artificial intelligence', 'w': 0.0}, {'name': 'machine learning', 'w': 0.45043}, {'name': 'machine translation', 'w': 0.57648}, {'name': 'single model', 'w': 0.0}], 'url': ['https://arxiv.org/abs/1706.03762', 'https://papers.nips.cc/paper/7181-attention-is-all-you-need.pdf', 'http://papers.nips.cc/paper/7181-attention-is-all-you-need', 'https://arxiv.org/pdf/1706.03762', 'https://ui.adsabs.harvard.edu/abs/2017arXiv170603762V/abstract', 'https://dblp.uni-trier.de/db/conf/nips/nips2017.html#VaswaniSPUJGKP17', 'https://www.arxiv-vanity.com/papers/1706.03762/', 'https://ai.google/research/pubs/pub46201', 'https://research.google/pubs/pub46201/', 'https://dl.acm.org/citation.cfm?id=3295349', 'https://research.google.com/pubs/pub46201.html']}, 'abstract': 'The dominant sequence transduction models are based on complex recurrent orconvolutional neural networks in an encoder and decoder configuration. The best performing such models also connect the encoder and decoder through an attentionm echanisms. We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train. Our single model with 165 million parameters, achieves 27.5 BLEU onEnglish-to-German translation, improving over the existing best ensemble result by over 1 BLEU. On English-to-French translation, we outperform the previoussingle state-of-the-art with model by 0.7 BLEU, achieving a BLEU score of 41.1.'})] +``` + +Search by unique ID and show all fields: + +```python +vars(massw["15a601e1-fd02-492c-8864-ee8e1f14fd55"]) +``` + +Output: + +```python +{'id': '15a601e1-fd02-492c-8864-ee8e1f14fd55', + 'title': 'Attention is All You Need', + 'abstract': 'The dominant sequence transduction models are based on complex recurrent orconvolutional neural networks in an encoder and decoder configuration. The best performing such models also connect the encoder and decoder through an attentionm echanisms. We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train. Our single model with 165 million parameters, achieves 27.5 BLEU onEnglish-to-German translation, improving over the existing best ensemble result by over 1 BLEU. On English-to-French translation, we outperform the previoussingle state-of-the-art with model by 0.7 BLEU, achieving a BLEU score of 41.1.', + 'year': 2017, + 'venue': 'NeurIPS', + 'context': 'Prior to this work, sequence transduction models were dominated by complex recurrent and convolutional neural networks in an encoder-decoder configuration, with optimal results found through the use of attention mechanisms connecting the encoder and decoder.', + 'key_idea': 'The paper introduces a new network architecture, the Transformer, that exclusively relies on an attention mechanism, foregoing recurrence and convolutions entirely.', + 'method': 'Experiments were conducted on two machine translation tasks to validate the proposed network architecture, comparing its quality, parallelism, and training speed against the existing models.', + 'outcome': 'The Transformer model outperformed existing models in terms of quality and training efficiency, achieving a BLEU score of 27.5 on English to German translation, which improved over the current best ensemble result by over 1 BLEU, and a BLEU score of 41.1 on English to French translation, surpassing the previous single state-of-the-art model by 0.7 BLEU.', + 'projected_impact': None, + 'data': {'id': '15a601e1-fd02-492c-8864-ee8e1f14fd55', + 'area': 'ml', + 'venue': 'neurips', + 'title': 'Attention is All You Need', + 'year': 2017, + 'data': {'id': 2963403868, + 'title': 'Attention is All You Need', + 'authors': [{'name': 'Ashish Vaswani', + 'id': 2171687631, + 'org_id': 1291425158}, + {'name': 'Noam Shazeer', 'id': 2496873187, 'org_id': 1291425158}, + {'name': 'Niki Parmar', 'id': 2625834147, 'org_id': 1174212}, + {'name': 'Jakob Uszkoreit', 'id': 2226984371, 'org_id': 1291425158}, + {'name': 'Llion Jones', 'id': 2624808693, 'org_id': 1291425158}, + {'name': 'Aidan N. Gomez', 'id': 2626656444}, + {'name': 'Lukasz Kaiser', 'id': 2948063405, 'org_id': 1291425158}, + {'name': 'Illia Polosukhin', 'id': 2519781307, 'org_id': 1291425158}], + 'venue': {'name': 'neural information processing systems', + 'id': 1127325140}, + 'year': 2017, + 'n_citation': 4375, + 'page_start': '5998', + 'page_end': '6008', + 'doc_type': 'Conference', + 'publisher': 'Curran Associates Inc.', + 'volume': '', + 'issue': '', + 'indexed_abstract': '{"IndexLength":116,"InvertedIndex":{"The":[0,19],"dominant":[1],"sequence":[2],"transduction":[3],"models":[4,23,59],"are":[5],"based":[6,41],"on":[7,52],"complex":[8],"recurrent":[9],"orconvolutional":[10],"neural":[11],"networks":[12],"in":[13],"an":[14,31],"encoder":[15,27],"and":[16,28,49,68],"decoder":[17,29],"configuration.":[18],"best":[20,90],"performing":[21],"such":[22],"also":[24],"connect":[25],"the":[26,88,102],"through":[30],"attentionm":[32],"echanisms.":[33],"We":[34],"propose":[35],"a":[36,111],"novel,":[37],"simple":[38],"network":[39],"architecture":[40],"solely":[42],"onan":[43],"attention":[44],"mechanism,":[45],"dispensing":[46],"with":[47,77,105],"recurrence":[48],"convolutions":[50],"entirely.Experiments":[51],"two":[53],"machine":[54],"translation":[55],"tasks":[56],"show":[57],"these":[58],"to":[60],"be":[61],"superiorin":[62],"quality":[63],"while":[64],"being":[65],"more":[66],"parallelizable":[67],"requiring":[69],"significantly":[70],"less":[71],"timeto":[72],"train.":[73],"Our":[74],"single":[75],"model":[76,106],"165":[78],"million":[79],"parameters,":[80],"achieves":[81],"27.5":[82],"BLEU":[83,112],"onEnglish-to-German":[84],"translation,":[85,99],"improving":[86],"over":[87,94],"existing":[89],"ensemble":[91],"result":[92],"by":[93,107],"1":[95],"BLEU.":[96],"On":[97],"English-to-French":[98],"we":[100],"outperform":[101],"previoussingle":[103],"state-of-the-art":[104],"0.7":[108],"BLEU,":[109],"achieving":[110],"score":[113],"of":[114],"41.1.":[115]}}', + 'fos': [{'name': 'artificial neural network', 'w': 0.47565}, + {'name': 'bleu', 'w': 0.50846}, + {'name': 'parallelizable manifold', 'w': 0.42665}, + {'name': 'convolution', 'w': 0.41377}, + {'name': 'encoder', 'w': 0.51848}, + {'name': 'computer science', 'w': 0.43723}, + {'name': 'network architecture', 'w': 0.45739}, + {'name': 'artificial intelligence', 'w': 0.0}, + {'name': 'machine learning', 'w': 0.45043}, + {'name': 'machine translation', 'w': 0.57648}, + {'name': 'single model', 'w': 0.0}], + 'url': ['https://arxiv.org/abs/1706.03762', + 'https://papers.nips.cc/paper/7181-attention-is-all-you-need.pdf', + 'http://papers.nips.cc/paper/7181-attention-is-all-you-need', + 'https://arxiv.org/pdf/1706.03762', + 'https://ui.adsabs.harvard.edu/abs/2017arXiv170603762V/abstract', + 'https://dblp.uni-trier.de/db/conf/nips/nips2017.html#VaswaniSPUJGKP17', + 'https://www.arxiv-vanity.com/papers/1706.03762/', + 'https://ai.google/research/pubs/pub46201', + 'https://research.google/pubs/pub46201/', + 'https://dl.acm.org/citation.cfm?id=3295349', + 'https://research.google.com/pubs/pub46201.html']}, + 'abstract': 'The dominant sequence transduction models are based on complex recurrent orconvolutional neural networks in an encoder and decoder configuration. The best performing such models also connect the encoder and decoder through an attentionm echanisms. We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train. Our single model with 165 million parameters, achieves 27.5 BLEU onEnglish-to-German translation, improving over the existing best ensemble result by over 1 BLEU. On English-to-French translation, we outperform the previoussingle state-of-the-art with model by 0.7 BLEU, achieving a BLEU score of 41.1.'}} +``` + +The dataset is iterable and ready for loop processing: + +```python +for workflow in massw: + print(workflow.title) +``` + +## Benchmark Tasks + +See [here](benchmark/aspect_prediction/README.md) for more details. + +## Citation + +If you find MASSW useful, please cite the following paper: + +```bibtex +@article{zhang2024massw, + title={MASSW: A New Dataset and Benchmark Tasks for AI-Assisted Scientific Workflows}, + author={Zhang, Xingjian and Xie, Yutong and Huang, Jin and Ma, Jinge and Pan, Zhaoying and Liu, Qijia and Xiong, Ziyang and Ergen, Tolga and Shim, Dongsub and Lee, Honglak and others}, + journal={arXiv preprint arXiv:2406.06357}, + year={2024} +} +``` diff --git a/assets/MASSW.svg b/assets/MASSW.svg new file mode 100644 index 0000000..abe0ca8 --- /dev/null +++ b/assets/MASSW.svg @@ -0,0 +1 @@ + \ No newline at end of file diff --git a/benchmark/aspect_prediction/README.md b/benchmark/aspect_prediction/README.md new file mode 100644 index 0000000..5a6e3b3 --- /dev/null +++ b/benchmark/aspect_prediction/README.md @@ -0,0 +1,88 @@ +# Benchmark + +## Reproduce Results in the Paper + +To reproduce the benchmark results across different models and prompt types, run + +```bash +python benchmark/aspect_prediction/task.py --model --prompt --num_samples 1020 +``` + +where: + +- `` is chosen from `gpt-35-turbo`, `gpt-4`, `mistral-8x7b`. +- `` is chosen from `zero-shot`, `few-shot`, `chain-of-thought`, `few-shot-cot`. + +> We provide the benchmark output through a Dropbox link +> [here](https://www.dropbox.com/scl/fi/nap87vh9s2mc7v3daql5u/results_v1.zip?rlkey=m1n5vck90quwhqygiq1otn2zp&dl=0). +> You could download the results and unzip them to the +> `benchmark/aspect_prediction/outputs` directory through: +> +> ```bash +> wget "https://www.dropbox.com/scl/fi/nap87vh9s2mc7v3daql5u/results_v1.zip?rlkey=m1n5vck90quwhqygiq1otn2zp&dl=1" -O results_v1.zip +> unzip results_v1.zip -d benchmark/aspect_prediction +> rm results_v1.zip +> mv benchmark/aspect_prediction/results benchmark/aspect_prediction/outputs +> ``` + +After running the tasks, evaluate the outcomes by running: + +```bash +python benchmark/aspect_prediction/eval.py --model_output_dir benchmark/aspect_prediction/outputs/gpt-35-turbo_zero-shot +``` + +--- + + + +## Adding a Custom Model to MASSW/API + +To extend the functionality of MASSW by adding custom API scripts for additional models, follow these guidelines. This will allow your model to integrate seamlessly with the existing framework used for aspect prediction and evaluation. + +#### 1. **Location for API Scripts** + +Place your custom API scripts in the `massw/api` directory. This should be similar in structure and design to the existing scripts: + +- `massw/api/api_gpt.py` +- `massw/api/api_mistral.py` + +#### 2. **Required Functions** + +Each API script must include two essential functions: + +- **`prompts_to_raw_output_`**: This function processes prompts and generates raw outputs. + +```python +def prompts_to_raw_output_(messages: List[Tuple[str, str]], **other_arguments) -> pd.DataFrame: + """ + Process prompts to generate raw outputs. + + Parameters: + - messages (List[Tuple[str, str]], str]]): A list of tuples containing paper IDs and messages. + 'pid' is the paper ID, and 'message' is the text of the conversation or prompt. + + Returns: + - pd.DataFrame: A DataFrame containing the processed outputs with paper IDs. + """ + pass +``` + +- **`raw_output_to_dict_`**: This function parses raw outputs into a dictionary format. + + ```python + def raw_output_to_dict_(output_path: str) -> Dict[str, str]: + """ + Convert raw outputs into a dictionary mapping from paper ID to output. + + Parameters: + - output_path (str): The file path to the output directory where the results are stored. + + Returns: + - Dict[str, str]: A dictionary mapping each paper ID to its corresponding output. + """ + pass + ``` + +#### 3. **Modify the Task Processing Function** + +Update the `process_task` function in `benchmark/aspect_prediction/task.py` to handle your custom model by calling your new API functions. Additionally, adapt the `postprocess_output` function in `benchmark/aspect_observer/eval.py` to support the evaluation of your model's outputs. \ No newline at end of file diff --git a/benchmark/aspect_prediction/eval.py b/benchmark/aspect_prediction/eval.py new file mode 100644 index 0000000..665ec73 --- /dev/null +++ b/benchmark/aspect_prediction/eval.py @@ -0,0 +1,115 @@ +"""Evaluate experiment results based on the model generated output (file).""" +import pandas as pd +from massw.metrics import compute_metrics, flatten_metrics +from massw.api.api_gpt import raw_output_to_dict_gpt +from massw.api.api_mistral import raw_output_to_dict_mistral +from utils import postprocess_cot, TASK_NAMES, TASK2GT + +import argparse +import json +import sys +import nest_asyncio + +sys.path.append("..") +nest_asyncio.apply() + + +def postprocess_output(model_output_dir, + reference_path, + used_cot=False, + model_type="gpt"): + """ + Process model output files to match predictions with references. + + Args: + model_output_dir (str): Directory containing the output files. + reference_path (str): Path to the file containing reference data. + used_cot (bool): Flag to determine if COT processing is needed. + model_type (str): Type of model used to adjust processing logic. + + Returns: + dict: A dictionary containing predictions and references by task. + """ + results = {} + with open(reference_path, "r", encoding="utf-8") as f: + references = [json.loads(line) for line in f] + id2ref = {r["pid"]: r for r in references} + + for task_name in TASK_NAMES: + gt_name = TASK2GT[task_name] + model_path = f"{model_output_dir}/{task_name}.tsv" + + if model_type == "gpt": + id2predictions = raw_output_to_dict_gpt(model_path) + elif model_type == "mistral": + id2predictions = raw_output_to_dict_mistral(model_path) + else: + raise ValueError(f"Model type {model_type} not supported.") + + if used_cot: + for pid in id2predictions: + id2predictions[pid] = postprocess_cot(id2predictions[pid]) + + results[task_name] = { + "predictions": list(id2predictions.values()), + "references": [id2ref[pid][gt_name] for pid in id2ref.keys() + if pid in id2predictions] + } + + return results + + +def main(): + """Run main function to process and compute evaluation metrics.""" + parser = argparse.ArgumentParser() + parser.add_argument( + "--model_output_dir", + type=str, + help="Path to the model output dir.", + default="benchmark/aspect_prediction/outputs/gpt-35-turbo_zero-shot", + ) + parser.add_argument( + "--reference_path", + type=str, + help="Path to the references file.", + default="data/benchmark_0531.jsonl", + ) + parser.add_argument( + "--used_cot", + action="store_true", + help="Used COT.", + ) + args = parser.parse_args() + model_type = "gpt" if "gpt" in args.model_output_dir else "mistral" + + results = postprocess_output( + args.model_output_dir, + args.reference_path, + args.used_cot, + model_type=model_type, + ) + metrics_output_path = f"{args.model_output_dir}/metrics.tsv" + + metrics = {} + + for task_name, task_results in results.items(): + print(f"Processing task: {task_name}") + predictions = task_results["predictions"] + references = task_results["references"] + metrics[task_name] = flatten_metrics( + compute_metrics( + predictions, + references, + metric_names=[ + "bleu", "rouge", "cosine", "bertscore", "bleurt" + ], + )) + print(f"Processed task: {task_name}") + print(metrics[task_name]) + + df = pd.DataFrame(metrics) + df.to_csv(metrics_output_path, index=True, sep="\t") + + +if __name__ == "__main__": + main() diff --git a/benchmark/aspect_prediction/prompts.py b/benchmark/aspect_prediction/prompts.py new file mode 100644 index 0000000..989360c --- /dev/null +++ b/benchmark/aspect_prediction/prompts.py @@ -0,0 +1,156 @@ +""" +This program defines a series of functions that automate the generation. + +Each function corresponds to a specific aspect of research paper generation, \ +such as idea generation, method\ +recommendation, outcome prediction, future work recommendation, \ +and title prediction. +""" + +SYSTEM_PROMPT = """ +You are an expert in research tasked with generating detailed prompts for \ + various aspects of academic research papers. +Each task involves creating a specific type of prompt \ +based on the provided information. Here are the definitions of \ +each part you will work with: + +- Concept + - Definition + - Relative Time + +- Context: The status quo of related literature or reality \ +which motivated this study. +This could normally be a problem, a research question, or a research gap \ +that has not been successfully addressed by previous work. \ +This is anything that happened before this study. + +- Key Idea: The main intellectual merit of this paper, \ +often in comparison to the context. +This could normally be a novel idea or solution proposed in this paper \ +that distinguishes it from what's already done in literature. +This is proposed in this study. + +- Method: The specific research method that investigates \ +and validates the key idea. +This could be an experimental setup, a theoretical framework, or \ +other necessary methodology to implement and/or evaluate the key idea. +This is performed in this study. + +- Outcome: The factual statement about the study output. +This could be the experiment results and any other measurable \ +outcome that has occurred. +It marks whether the key hypothesis is testified or not. \ +This is produced in this study. + +- Projected Impact: The author-anticipated impact of the work on the field, \ +and potential further research identified by the author \ +that may improve or extend this study. +This is anything being anticipated but has not happened yet. +""" + + +def idea_generation(data): + """ + Generate a prompt for idea generation based on the provided context. + + Args: + data (dict): Contains 'context' and 'key_idea' from the research data. + + Returns: + tuple: A tuple containing the prompt and the ground truth for + idea generation. + """ + context = data['context'] + prompt = f"Given the context: '{context}', generate key ideas \ + that could advance this area of study. " + ground_truth = data['key_idea'] + return prompt, ground_truth + + +def method_recommendation(data): + """ + Recommend a method to validate a key idea. + + Args: + data (dict): Contains 'context', 'key_idea', and 'method' + from the research data. + + Returns: + tuple: A tuple containing the prompt and the ground truth + for method recommendation. + """ + context = data['context'] + key_idea = data['key_idea'] + prompt = f"Given the context: '{context}' and the key idea: '{key_idea}', \ + recommend the most suitable method to validate this idea. " + ground_truth = data['method'] + return prompt, ground_truth + + +def outcome_prediction(data): + """ + Predict the potential outcome of a research. + + Args: + data (dict): Contains 'context', 'key_idea', 'method', and 'outcome'. + + Returns: + tuple: A tuple containing the prompt and the ground truth + for outcome prediction. + """ + context = data['context'] + key_idea = data['key_idea'] + method = data['method'] + prompt = f"Based on the context: '{context}', the key idea: '{key_idea}', \ + and the recommended method: '{method}', \ + predict the potential outcome of this research. " + ground_truth = data['outcome'] + return prompt, ground_truth + + +def future_work_recommendation(data): + """ + Suggest projected impact for the research. + + Args: + data (dict): Contains 'context', 'key_idea', 'method', 'outcome', \ + and 'future_impact' from the research data. + + Returns: + tuple: A tuple containing the prompt and the ground truth + for future work. + """ + context = data['context'] + key_idea = data['key_idea'] + method = data['method'] + outcome = data['outcome'] + prompt = f"Based on the context: '{context}', the key idea: '{key_idea}', \ + the method: '{method}', and the outcome: '{outcome}', \ + suggest projected impact for this research." + ground_truth = data.get('future_impact', '') + return prompt, ground_truth + + +def predict_title(data): + """ + Predict the title of a research paper. + + Args: + data (dict): Contains all necessary information from the research data. + + Returns: + tuple: A tuple containing the prompt and the ground trut + for title prediction. + """ + context = data['context'] + key_idea = data['key_idea'] + method = data['method'] + outcome = data['outcome'] + future_impact = data['future_impact'] + prompt = f"Given the context: '{context}', the key idea: '{key_idea}', \ + the method: '{method}', the outcome: '{outcome}', \ + and the future impact: '{future_impact}', \ + predict the title of this research paper. \ + The title should be concise and reflective of the core aspects." + ground_truth = data.get('title', '') + return prompt, ground_truth diff --git a/benchmark/aspect_prediction/task.py b/benchmark/aspect_prediction/task.py new file mode 100644 index 0000000..42237f5 --- /dev/null +++ b/benchmark/aspect_prediction/task.py @@ -0,0 +1,158 @@ +""" +This script facilitates the automated benchmarking of research prompts. + +The script supports multiple models and prompt types +and is designed to work with large sets +of test data asynchronously. +""" + +from massw.api.api_gpt import prompts_to_raw_output_gpt +from massw.api.api_mistral import prompts_to_raw_output_mistral + +from prompts import ( + future_work_recommendation, idea_generation, method_recommendation, + outcome_prediction, predict_title, SYSTEM_PROMPT +) +from utils import ( + allow_self_signed_https, load_examples, + MODEL_CHOICES, PROMPT_CHOICES, save_results +) + +import argparse +import os +import sys +import jsonlines as jl + +sys.path.append("../..") + +allow_self_signed_https(True) + +few_shot_examples, cot_examples = load_examples() + + +def prepare_messages(model, task_name, prompt_type, main_prompt): + """Prepare the messages based on the task and prompt type.""" + messages = [{"role": "system", "content": SYSTEM_PROMPT}] + if model == "mistral-8x7b": + format_instruction = "\nNever use double quotes in your output. \ + Use single quotes instead.\n" + messages = [ + {"role": "user", "content": SYSTEM_PROMPT + format_instruction}, + {"role": "assistant", "content": "I got it. \ + Please give me further instructions!"} + ] + + if prompt_type in ["few-shot", "few-shot-cot"]: + examples = few_shot_examples if prompt_type == "few-shot"\ + else cot_examples + for example in examples.get(task_name, []): + messages.extend([ + {"role": "user", "content": example["user"]}, + {"role": "assistant", "content": example["assistant"]} + ]) + + if prompt_type == "chain-of-thought": + main_prompt += "Let's think step by step. \ + You should first present you reasoning. \ + After that, the final prediction should start after \ + the marker 'Prediction:'." + + messages.append({"role": "user", "content": main_prompt}) + + return messages + + +def process_task(generate_prompt_fn, test_cases, task_name, **kwargs): + """Process the tasks and retrieve chat completions.""" + messages = [] + for entry in test_cases: + main_prompt, _ = generate_prompt_fn(entry) + message = prepare_messages(kwargs['model'], + task_name, + kwargs['prompt_type'], + main_prompt) + messages.append((entry['pid'], message)) + + model = kwargs['model'] + if model == "mistral-8x7b": + chat_results = prompts_to_raw_output_mistral(messages) + elif model in ["gpt-35-turbo", "gpt-4"]: + chat_results = prompts_to_raw_output_gpt(messages, + model, + kwargs.get('tpm')) + else: + raise ValueError(f"Model {model} not supported. \ + You can modify the code here \ + to support custom models.") + + return chat_results + + +def main(): + """Execute main function to process tasks.""" + parser = argparse.ArgumentParser(description="Process benchmarking \ + of academic paper prompts.") + parser.add_argument("--test_data", + type=str, + default="data/benchmark_0531.jsonl") + parser.add_argument("--output_dir", + type=str, + default=False) + parser.add_argument("--model", + type=str, + default="gpt-35-turbo") + parser.add_argument("--prompt", + type=str, + default="zero-shot") + parser.add_argument("--num_samples", + type=int, + default=5) + args = parser.parse_args() + + if args.model not in MODEL_CHOICES: + raise ValueError(f"Model {args.model} not supported. \ + Choose from {MODEL_CHOICES}") + + if args.prompt not in PROMPT_CHOICES: + raise ValueError(f"Prompt type {args.prompt} not supported. \ + Choose from {PROMPT_CHOICES}") + + if not args.output_dir: + args.output_dir = os.path.join("benchmark", + "aspect_prediction", + "outputs", + f"{args.model}_{args.prompt}") + + # Load test data + with jl.open(args.test_data) as file: + test_data = [record for record, _ in + zip(file, range(args.num_samples))] + + tasks = [ + ("idea_generation", idea_generation), + ("method_recommendation", method_recommendation), + ("outcome_prediction", outcome_prediction), + ("future_work_recommendation", future_work_recommendation), + ("title_prediction", predict_title) + ] + + tokens_per_minute = {"gpt-35-turbo": 40000, + "gpt-4": 10000, + "mistral-8x7b": None} + + for task_name, generate_prompt_fn in tasks: + print(f"Processing task: {task_name}") + chat_results = process_task( + generate_prompt_fn, + test_data, + task_name, + model=args.model, + prompt_type=args.prompt, + tpm=tokens_per_minute[args.model] + ) + print(f"{chat_results = }") + save_results(chat_results, args.output_dir, task_name) + + +if __name__ == "__main__": + main() diff --git a/benchmark/aspect_prediction/utils.py b/benchmark/aspect_prediction/utils.py new file mode 100644 index 0000000..ee8e36d --- /dev/null +++ b/benchmark/aspect_prediction/utils.py @@ -0,0 +1,80 @@ +"""This script includes utility functions for benchmarking scripts.""" +import ssl +import json +import os + +MODEL_CHOICES = ["gpt-35-turbo", "gpt-4", "mistral-8x7b"] +PROMPT_CHOICES = ["zero-shot", "few-shot", "chain-of-thought", "few-shot-cot"] + + +def allow_self_signed_https(allowed): + """Config SSL settings to allow self-signed certificates.""" + if allowed and not os.environ.get('PYTHONHTTPSVERIFY', '') and getattr( + ssl, '_create_unverified_context', None): + ssl._create_default_https_context = ssl._create_unverified_context + + +def load_examples(): + """Load few-shot and chain-of-thought examples from files.""" + with open("data/few_shot_examples.json", "r", encoding="utf-8") as f: + few_shot_examples = json.load(f) + with open("data/cot_examples.json", "r", encoding="utf-8") as f: + cot_examples = json.load(f) + return few_shot_examples, cot_examples + + +def save_results(chat_results, output_dir, task_name): + """ + Save the chat results to a TSV file. + + Args: + chat_results (DataFrame): DataFrame containing the chat results. + output_dir (str): Directory path to save the result files. + task_name (str): Name of task which will be used to name the output. + """ + os.makedirs(output_dir, exist_ok=True) + + # Serializing dictionary entries in 'output' column, if present + if 'output' in chat_results.columns: + chat_results['output'] = chat_results['output'].apply( + lambda x: json.dumps(x) if isinstance(x, dict) else x + ) + + output_path = os.path.join(output_dir, f"{task_name}.tsv") + chat_results.to_csv(output_path, sep="\t", index=False) + + +TASK_NAMES = [ + "idea_generation", + "method_recommendation", + "outcome_prediction", + "future_work_recommendation", + "title_prediction", +] + +TASK2GT = { + "idea_generation": "key_idea", + "method_recommendation": "method", + "outcome_prediction": "outcome", + "future_work_recommendation": "future_impact", + "title_prediction": "title", +} + + +def postprocess_cot(output: str): + """ + Extract the actual prediction from the output string. + + Args: + output (str): The output string containing the prediction. + + Returns: + str: The extracted prediction or the original output + if no marker is found. + """ + marker_index = output.find("Prediction:") + if marker_index != -1: + actual_prediction = output[marker_index + len("Prediction:"):].strip() + return actual_prediction + + return output diff --git a/data/annotation_0531.jsonl b/data/annotation_0531.jsonl new file mode 100644 index 0000000..4f5dc27 --- /dev/null +++ b/data/annotation_0531.jsonl @@ -0,0 +1,240 @@ +{"id": "01f161fe-dd40-45dd-89bd-fb1562771d73", "displayed_text": "Title: Learning Dynamic Context Graphs for Predicting Social Events\n\nAbstract: Event forecasting with an aim at modeling contextual information is an important task for applications such as automated analysis generation and resource allocation. Captured contextual information for an event of interest can aid human analysts in understanding the factors associated with that event. However, capturing contextual information within event forecasting is challenging due to several factors: (i) uncertainty of context structure and formulation, (ii) high dimensional features, and (iii) adaptation of features over time. Recently, graph representations have demonstrated success in applications such as traffic forecasting, social influence prediction, and visual question answering systems. In this paper, we study graph representations in modeling social events to identify dynamic properties of event contexts as social indicators. Inspired by graph neural networks, we propose a novel graph convolutional network for predicting future events (e.g., civil unrest movements). We extract and learn graph representations from historical/prior event documents. By employing the hidden word graph features, our proposed model predicts the occurrence of future events and identifies sequences of dynamic graphs as event context. Experimental results on multiple real-world data sets show that the proposed method is competitive against various state-of-the-art methods for social event prediction.", "label_annotations": {"Multi-aspect Summary": {"Context": "Capturing contextual information for an event can aid analysts in understanding the factors associated with the event, but it is challenging due to uncertainty of context structure, high dimensional features and adaptation of features over time. Recently, graph representations have rendered helpful on capturing contextual information.", "Key idea": "The author proposed a graph convolutional network for predicting future events. The network model predicts the future events and identifies sequences of dynamic graphs as event context by learning graph representations from historical event documents, employing the hidden word graph features.", "Method": "The author utilizes real-world data sets to benchmark the proposed model and compare against various state-of-the-art methods for social event prediction.", "Outcome": "The results show that the proposed model is competitive against various state-of-the-art methods for social event prediction.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 51s "}} +{"id": "01f161fe-dd40-45dd-89bd-fb1562771d73", "displayed_text": "Title: Learning Dynamic Context Graphs for Predicting Social Events\n\nAbstract: Event forecasting with an aim at modeling contextual information is an important task for applications such as automated analysis generation and resource allocation. Captured contextual information for an event of interest can aid human analysts in understanding the factors associated with that event. However, capturing contextual information within event forecasting is challenging due to several factors: (i) uncertainty of context structure and formulation, (ii) high dimensional features, and (iii) adaptation of features over time. Recently, graph representations have demonstrated success in applications such as traffic forecasting, social influence prediction, and visual question answering systems. In this paper, we study graph representations in modeling social events to identify dynamic properties of event contexts as social indicators. Inspired by graph neural networks, we propose a novel graph convolutional network for predicting future events (e.g., civil unrest movements). We extract and learn graph representations from historical/prior event documents. By employing the hidden word graph features, our proposed model predicts the occurrence of future events and identifies sequences of dynamic graphs as event context. Experimental results on multiple real-world data sets show that the proposed method is competitive against various state-of-the-art methods for social event prediction.", "label_annotations": {"Multi-aspect Summary": {"Context": "Event forecasting with an aim at modeling contextual information is a challenging but important task for applications such as automated analysis generation and resource allocation. Recently, graph representations have demonstrated success in applications such as traffic forecasting, social influence prediction, and visual question answering systems.", "Key idea": "The authors propose a novel graph convolutional network for predicting future events.", "Method": "The authors evaluate their approach on multiple real-world data sets.", "Outcome": "Experimental results on multiple real-world data sets show that the proposed method is competitive against various state-of-the-art methods for social event prediction.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 6m 10s "}} +{"id": "07c3daea-a88c-4a67-9aac-20ef0ec62e79", "displayed_text": "Title: Enabling Biographical Cognitive Stimulation for People with Dementia\n\nAbstract: Non-pharmacological interventions are the most common and arguably most effective for people with dementia. Some of these approaches have been proven to benefit from the usage of biographical or personalized materials. These contents are not always easy to obtain. Alongside, it is a challenge to maintain awareness of what is meaningful for a certain person. Faced with an absence of tools to collect and manage biographical materials, we created a web platform that supports the work of psychologists, streamlining the collection of relevant information about people with dementia. This knowledge is then used as a starting point to perform reminiscence and other biographical cognitive stimulation practices. In this paper, we present the design of our platform and results from a case study with one psychologist and three patients, across a period of two weeks that showed improvements in the collection of meaningful data about a person, and on maintaining awareness of the therapy as a whole.", "label_annotations": {"Multi-aspect Summary": {"Context": "It is a challenge to maintain awareness of what is meaningful for a certain person, and absence of tools to collect and manage biographical materials happens to biographical for people with Dementia.", "Key idea": "The authors created a web platform that supports the work of psychologists, streamlining the collection of relevant information about people with dementia.", "Method": "The authors use a case study with one psychologist and three patients, across a period of two weeks to show the effectiveness of new web platform.", "Outcome": "The case study show that web platform make improvements in the collection of meaningful data about a person, and on maintaining awareness of the therapy as a whole.\r\n", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 14s "}} +{"id": "07c3daea-a88c-4a67-9aac-20ef0ec62e79", "displayed_text": "Title: Enabling Biographical Cognitive Stimulation for People with Dementia\n\nAbstract: Non-pharmacological interventions are the most common and arguably most effective for people with dementia. Some of these approaches have been proven to benefit from the usage of biographical or personalized materials. These contents are not always easy to obtain. Alongside, it is a challenge to maintain awareness of what is meaningful for a certain person. Faced with an absence of tools to collect and manage biographical materials, we created a web platform that supports the work of psychologists, streamlining the collection of relevant information about people with dementia. This knowledge is then used as a starting point to perform reminiscence and other biographical cognitive stimulation practices. In this paper, we present the design of our platform and results from a case study with one psychologist and three patients, across a period of two weeks that showed improvements in the collection of meaningful data about a person, and on maintaining awareness of the therapy as a whole.", "label_annotations": {"Multi-aspect Summary": {"Context": "Non-pharmacological interventions are the most common and arguably most effective for people with dementia. However, some of these approaches have been proven to benefit from the usage of biographical or personalized materials, which are not always easy to obtain and challenging to distinguish what are meaningful for a certain person.", "Key idea": "The authors create a web platform that supports the work of psychologists, streamlining the collection of relevant information about people with dementia.", "Method": "The authors conduct a case study with one psychologist and three patients, across a period of two weeks.", "Outcome": "Results from a case study with one psychologist and three patients, across a period of two weeks show improvements of the proposed platform in the collection of meaningful data about a person, and on maintaining awareness of the therapy as a whole.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 7m 45s "}} +{"id": "081d6673-3c7c-4aec-b101-cf55d75ac718", "displayed_text": "Title: Fast, Accurate and Memory-Efficient Partial Permutation Synchronization\n\nAbstract: Previous partial permutation synchronization (PPS) algorithms, which are commonly used for multi-object matching, often involve computation-intensive and memory-demanding matrix operations. These operations become intractable for large scale structure-from-motion datasets. For pure permutation synchronization, the recent Cycle-Edge Message Passing (CEMP) framework suggests a memory-efficient and fast solution. Here we overcome the restriction of CEMP to compact groups and propose an improved algorithm, CEMP-Partial, for estimating the corruption levels of the observed partial permutations. It allows us to subsequently implement a nonconvex weighted projected power method without the need of spectral initialization. The resulting new PPS algorithm, MatchFAME (Fast, Accurate and Memory-Efficient Matching), only involves sparse matrix operations, and thus enjoys lower time and space complexities in comparison to previous PPS algorithms. We prove that under adversarial corruption, though without additive noise and with certain assumptions, CEMP-Partial is able to exactly classify corrupted and clean partial permutations. We demonstrate the state-of-the-art accuracy, speed and memory efficiency of our method on both synthetic and real datasets.", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous partial permutation synchronization (PPS) algorithms commonly used for multi-object matching is intractable for large scale structure-from-motion datasets. For pure permutation synchronization, the recent Cycle-Edge Message Passing (CEMP) framework suggests a memory-efficient and fast solution.", "Key idea": "The authors propose an improved algorithm named CEMP-Partial, which can overcome the restriction of CEMP to compact groups, used to estimate the corruption levels of the observed partial permutations. CEMP-Partial allows people to subsequently implement a nonconvex weighted projected power method without the need of spectral initialization. ", "Method": "The authors apply CEMP-Partial algorithms on adversarial corruption and on both synthetic and real datasets.", "Outcome": "Compared to previous PPS algorithms, CEMP-Partial enjoys lower time and space complexities. Under adversarial corruption, CEMP-Partial is able to exactly classify corrupted and clean partial permutations without additive noise and with certain assumptions. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 51s "}} +{"id": "081d6673-3c7c-4aec-b101-cf55d75ac718", "displayed_text": "Title: Fast, Accurate and Memory-Efficient Partial Permutation Synchronization\n\nAbstract: Previous partial permutation synchronization (PPS) algorithms, which are commonly used for multi-object matching, often involve computation-intensive and memory-demanding matrix operations. These operations become intractable for large scale structure-from-motion datasets. For pure permutation synchronization, the recent Cycle-Edge Message Passing (CEMP) framework suggests a memory-efficient and fast solution. Here we overcome the restriction of CEMP to compact groups and propose an improved algorithm, CEMP-Partial, for estimating the corruption levels of the observed partial permutations. It allows us to subsequently implement a nonconvex weighted projected power method without the need of spectral initialization. The resulting new PPS algorithm, MatchFAME (Fast, Accurate and Memory-Efficient Matching), only involves sparse matrix operations, and thus enjoys lower time and space complexities in comparison to previous PPS algorithms. We prove that under adversarial corruption, though without additive noise and with certain assumptions, CEMP-Partial is able to exactly classify corrupted and clean partial permutations. We demonstrate the state-of-the-art accuracy, speed and memory efficiency of our method on both synthetic and real datasets.", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous partial permutation synchronization (PPS) algorithms often involve computation-intensive and memory-demanding matrix operations, which are intractable for large scale structure-from-motion datasets. Pure permutation synchronization, such as Cycle-Edge Message Passing (CEMP) framework suggests a memory-efficient and fast solution.", "Key idea": "The authors propose an improved algorithm, CEMP-Partial, for estimating the corruption levels of the observed partial permutations, as well as the resulting new PPS algorithm, MatchFAME (Fast, Accurate and Memory-Efficient Matching), which only involves sparse matrix operations, and thus enjoys lower time and space complexities in comparison to previous PPS algorithms..", "Method": "The authors examine the performance of CEMP-Partial under adversarial corruption but without additive noise and with certain assumptions.", "Outcome": "The authors prove that CEMP-Partial is able to exactly classify corrupted and clean partial permutations and demonstrate the state-of-the-art accuracy, speed and memory efficiency of our method on both synthetic and real datasets.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 9m 48s "}} +{"id": "0b31e456-4944-47e5-80ed-deaf6421c375", "displayed_text": "Title: Directed dialogue protocols: verbal data for user interface design\n\nAbstract: The development of an interface design tool called \u201cdirected dialogue protocols\u201d is discussed. The tool is based upon Katou0027s (1986) method of verbal data collection, \u201cquestion-asking protocols.\u201d Three extensions to the question-asking method are detailed: 1) an experimental procedure of atomic tasks which facilitate the quantization of verbal data; 2) interventions by the experimenter that probe the subjectu0027s expectations and prompt verbalizations; and 3) a technique for answering subject queries called sequential disclosure. Also discussed are applications of the directed dialogue that have identified design choices which build learnability and usability into a productu0027s user-interface.", "label_annotations": {"Multi-aspect Summary": {"Context": "An interface design tool called \u201cdirected dialogue protocols\u201d is developed, based on Katou0027s (1986) method of verbal data collection, \u201cquestion-asking protocols.\u201d.", "Key idea": "Three extensions to the question-asking method are detailed: 1) an experimental procedure of atomic tasks which facilitate the quantization of verbal data; 2) interventions by the experimenter that probe the subjectu0027s expectations and prompt verbalizations; and 3) a technique for answering subject queries called sequential disclosure.", "Method": "N/A", "Outcome": "Authors develop applications of the directed dialogue that have identified design choices which build learnability and usability into a productu0027s user-interface.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 22s "}} +{"id": "0b31e456-4944-47e5-80ed-deaf6421c375", "displayed_text": "Title: Directed dialogue protocols: verbal data for user interface design\n\nAbstract: The development of an interface design tool called \u201cdirected dialogue protocols\u201d is discussed. The tool is based upon Katou0027s (1986) method of verbal data collection, \u201cquestion-asking protocols.\u201d Three extensions to the question-asking method are detailed: 1) an experimental procedure of atomic tasks which facilitate the quantization of verbal data; 2) interventions by the experimenter that probe the subjectu0027s expectations and prompt verbalizations; and 3) a technique for answering subject queries called sequential disclosure. Also discussed are applications of the directed dialogue that have identified design choices which build learnability and usability into a productu0027s user-interface.", "label_annotations": {"Multi-aspect Summary": {"Context": "In 1986, Kato propose a method of verbal data collection, \u201cquestion-asking protocols.\u201d", "Key idea": "The authors discuss the development of an interface design tool called \u201cdirected dialogue protocols\u201d, which has 3 extensions to Kato's question-asking method. They also discuss the applications of the directed dialogue that have identified design choices which build learnability and usability into a product's user-interface.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 22s "}} +{"id": "10c15fe5-c315-4b6d-8910-e6bc3279c817", "displayed_text": "Title: Knowledge-Preserving Incremental Social Event Detection via Heterogeneous GNNs\n\nAbstract: ABSTRACT Social events provide valuable insights into group social behaviors and public concerns and therefore have many applications in fields such as product recommendation and crisis management. The complexity and streaming nature of social messages make it appealing to address social event detection in an incremental learning setting, where acquiring, preserving, and extending knowledge are major concerns. Most existing methods, including those based on incremental clustering and community detection, learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data. Moreover, they cannot memorize previously acquired knowledge. In this paper, we propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection. To acquire more knowledge, KPGNN models complex social messages into unified social graphs to facilitate data utilization and explores the expressive power of GNNs for knowledge extraction. To continuously adapt to the incoming data, KPGNN adopts contrastive loss terms that cope with a changing number of event classes. It also leverages the inductive learning ability of GNNs to efficiently detect events and extends its knowledge from previously unseen data. To deal with large social streams, KPGNN adopts a mini-batch subgraph sampling strategy for scalable training, and periodically removes obsolete data to maintain a dynamic embedding space. KPGNN requires no feature engineering and has few hyperparameters to tune. Extensive experiment results demonstrate the superiority of KPGNN over various baselines.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most existing methods, including those based on incremental clustering and community detection, learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data. Moreover, they cannot memorize previously acquired knowledge.", "Key idea": "The authors propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection. ", "Method": "The authors conduct experiment to evaluate Knowledge-Preserving Incremental Heterogeneous Graph Neural Network(KPGNN)'s performance on incremental social event detection compared to baseline method's performance.", "Outcome": "Extensive experiment results demonstrate the superiority of KPGNN's performance on incremental social event detection over various baselines.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "10c15fe5-c315-4b6d-8910-e6bc3279c817", "displayed_text": "Title: Knowledge-Preserving Incremental Social Event Detection via Heterogeneous GNNs\n\nAbstract: ABSTRACT Social events provide valuable insights into group social behaviors and public concerns and therefore have many applications in fields such as product recommendation and crisis management. The complexity and streaming nature of social messages make it appealing to address social event detection in an incremental learning setting, where acquiring, preserving, and extending knowledge are major concerns. Most existing methods, including those based on incremental clustering and community detection, learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data. Moreover, they cannot memorize previously acquired knowledge. In this paper, we propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection. To acquire more knowledge, KPGNN models complex social messages into unified social graphs to facilitate data utilization and explores the expressive power of GNNs for knowledge extraction. To continuously adapt to the incoming data, KPGNN adopts contrastive loss terms that cope with a changing number of event classes. It also leverages the inductive learning ability of GNNs to efficiently detect events and extends its knowledge from previously unseen data. To deal with large social streams, KPGNN adopts a mini-batch subgraph sampling strategy for scalable training, and periodically removes obsolete data to maintain a dynamic embedding space. KPGNN requires no feature engineering and has few hyperparameters to tune. Extensive experiment results demonstrate the superiority of KPGNN over various baselines.", "label_annotations": {"Multi-aspect Summary": {"Context": "The complexity and streaming nature of social messages make it appealing to address social event detection in an incremental learning setting, where acquiring, preserving, and extending knowledge are major concerns. Most existing methods, including those based on incremental clustering and community detection, learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data. Moreover, they cannot memorize previously acquired knowledge.", "Key idea": "Authors propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection. To acquire more knowledge, KPGNN models complex social messages into unified social graphs to facilitate data utilization and explores the expressive power of GNNs for knowledge extraction. To continuously adapt to the incoming data, KPGNN adopts contrastive loss terms that cope with a changing number of event classes. It also leverages the inductive learning ability of GNNs to efficiently detect events and extends its knowledge from previously unseen data. To deal with large social streams, KPGNN adopts a mini-batch subgraph sampling strategy for scalable training, and periodically removes obsolete data to maintain a dynamic embedding space. ", "Method": "Authors setup experiments to compare KPGNN over various baselines.", "Outcome": "Extensive experiment results demonstrate the superiority of KPGNN over various baselines.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 28s "}} +{"id": "14b0ebd1-b654-4eed-bdd8-ebeb74250b15", "displayed_text": "Title: Relation-Guided Few-Shot Relational Triple Extraction\n\nAbstract: In few-shot relational triple extraction (FS-RTE), one seeks to extract relational triples from plain texts by utilizing only few annotated samples. Recent work first extracts all entities and then classifies their relations. Such an entity-then-relation paradigm ignores the entity discrepancy between relations. To address it, we propose a novel task decomposition strategy, Relation-then-Entity, for FS-RTE. It first detects relations occurred in a sentence and then extracts the corresponding head/tail entities of the detected relations. To instantiate this strategy, we further propose a model, RelATE, which builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities. Experimental results show that our model outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).", "label_annotations": {"Multi-aspect Summary": {"Context": "In few-shot relational triple extraction, one seeks to extract relational triples from plain texts by utilizing few annotated samples. Recent work first extracts all entities and then classifies their relations, which ignores the entity discrepancy between relations.", "Key idea": "The author first proposes a novel task decomposition strategy for FS-RTE, which detects relations occurred in a sentence and then extracts the corresponding head/tail entities of the detected relations. To instantiate this strategy, the author also proposes a model that builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities.", "Method": "The author assesses the performance of the proposed strategy against previous works.", "Outcome": "Experimental results show that the proposed model outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 55s "}} +{"id": "14b0ebd1-b654-4eed-bdd8-ebeb74250b15", "displayed_text": "Title: Relation-Guided Few-Shot Relational Triple Extraction\n\nAbstract: In few-shot relational triple extraction (FS-RTE), one seeks to extract relational triples from plain texts by utilizing only few annotated samples. Recent work first extracts all entities and then classifies their relations. Such an entity-then-relation paradigm ignores the entity discrepancy between relations. To address it, we propose a novel task decomposition strategy, Relation-then-Entity, for FS-RTE. It first detects relations occurred in a sentence and then extracts the corresponding head/tail entities of the detected relations. To instantiate this strategy, we further propose a model, RelATE, which builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities. Experimental results show that our model outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).", "label_annotations": {"Multi-aspect Summary": {"Context": "Recent work in few-shot relational triple extraction (FS-RTE) first extracts all entities and then classifies their relations. Such an entity-then-relation paradigm ignores the entity discrepancy between relations.", "Key idea": "The authors propose a novel task decomposition strategy, Relation-then-Entity, for FS-RTE, which first detects relations occurred in a sentence and then extracts the corresponding head/tail entities of the detected relations. They further propose a model, RelATE, which builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities.", "Method": "N/A", "Outcome": "Experimental results show that our model outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "16c1b4ae-73f8-4c23-8bdb-b931ade1baa5", "displayed_text": "Title: Computing optimal subsets\n\nAbstract: Various tasks in decision making and decision support require selecting a preferred subset of items from a given set of feasible items. Recent work in this area considered methods for specifying such preferences based on the attribute values of individual elements within the set. Of these, the approach of (Brafman et al. 2006) appears to be the most general. In this paper, we consider the problem of computing an optimal subset given such a specification. The problem is shown to be NP-hard in the general case, necessitating heuristic search methods. We consider two algorithm classes for this problem: direct set construction, and implicit enumeration as solutions to appropriate CSPs. New algorithms are presented in each class and compared empirically against previous results.", "label_annotations": {"Multi-aspect Summary": {"Context": "Various tasks in decision making and decision support require selecting a preferred subset of items from a given set of feasible items. Recent work in this area considered methods for specifying such preferences based on the attribute values of individual elements within the set. Of these, the approach of (Brafman et al. 2006) appears to be the most general.", "Key idea": "Authors consider the problem of computing an optimal subset given such a specification. Authors consider two algorithm classes for this problem: direct set construction, and implicit enumeration as solutions to appropriate CSPs. New algorithms are presented in each class and compared empirically against previous results.", "Method": "Authors compare proposed method with previous methods empirically .", "Outcome": "Authors present new algorithms in each class and compared empirically against previous results.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 52s "}} +{"id": "16c1b4ae-73f8-4c23-8bdb-b931ade1baa5", "displayed_text": "Title: Computing optimal subsets\n\nAbstract: Various tasks in decision making and decision support require selecting a preferred subset of items from a given set of feasible items. Recent work in this area considered methods for specifying such preferences based on the attribute values of individual elements within the set. Of these, the approach of (Brafman et al. 2006) appears to be the most general. In this paper, we consider the problem of computing an optimal subset given such a specification. The problem is shown to be NP-hard in the general case, necessitating heuristic search methods. We consider two algorithm classes for this problem: direct set construction, and implicit enumeration as solutions to appropriate CSPs. New algorithms are presented in each class and compared empirically against previous results.", "label_annotations": {"Multi-aspect Summary": {"Context": "Various tasks in decision making and decision support require selecting a preferred subset of items from a given set of feasible items. Recent work in this area considered the general methods for specifying such preferences based on the attribute values of individual elements within the set.", "Key idea": "The authors contribute to two algorithm classes for the problem of computing an optimal subset: direct set construction, and implicit enumeration as solutions to appropriate CSPs.", "Method": "The authors compare their proposed algorithms empirically against previous results.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 37m 39s "}} +{"id": "18f27ced-9f37-45d7-9b76-6663c349d408", "displayed_text": "Title: Scalable Deep Multimodal Learning for Cross-Modal Retrieval\n\nAbstract: Cross-modal retrieval takes one type of data as the query to retrieve relevant data of another type. Most of existing cross-modal retrieval approaches were proposed to learn a common subspace in a joint manner, where the data from all modalities have to be involved during the whole training process. For these approaches, the optimal parameters of different modality-specific transformations are dependent on each other and the whole model has to be retrained when handling samples from new modalities. In this paper, we present a novel cross-modal retrieval method, called Scalable Deep Multimodal Learning (SDML). It proposes to predefine a common subspace, in which the between-class variation is maximized while the within-class variation is minimized. Then, it trains m modality-specific networks for m modalities (one network for each modality) to transform the multimodal data into the predefined common subspace to achieve multimodal learning. Unlike many of the existing methods, our method can train different modality-specific networks independently and thus be scalable to the number of modalities. To the best of our knowledge, the proposed SDML could be one of the first works to independently project data of an unfixed number of modalities into a predefined common subspace. Comprehensive experimental results on four widely-used benchmark datasets demonstrate that the proposed method is effective and efficient in multimodal learning and outperforms the state-of-the-art methods in cross-modal retrieval.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most of existing cross-modal retrieval approaches were proposed to learn a common subspace in a joint manner, where the data from all modalities have to be involved during the whole training process. \r\nIn these approaches, the whole model has to be retrained when handling samples from new modalities.", "Key idea": "The paper introduces a novel cross-modal retrieval method namedScalable Deep Multimodal Learning (SDML), a method that predefines a common subspace to maximize between-class variation and minimize within-class variation. ", "Method": "The authors applied SDML on four widely-used benchmark datasets to figure its performance.", "Outcome": "SDML could be one of the first works to independently project data of an unfixed number of modalities into a predefined common subspace.\r\nExperimental results on four widely-used benchmark datasets demonstrate that the SDML is effective and efficient in multimodal learning and outperforms the state-of-the-art methods in cross-modal retrieval.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 23s "}} +{"id": "18f27ced-9f37-45d7-9b76-6663c349d408", "displayed_text": "Title: Scalable Deep Multimodal Learning for Cross-Modal Retrieval\n\nAbstract: Cross-modal retrieval takes one type of data as the query to retrieve relevant data of another type. Most of existing cross-modal retrieval approaches were proposed to learn a common subspace in a joint manner, where the data from all modalities have to be involved during the whole training process. For these approaches, the optimal parameters of different modality-specific transformations are dependent on each other and the whole model has to be retrained when handling samples from new modalities. In this paper, we present a novel cross-modal retrieval method, called Scalable Deep Multimodal Learning (SDML). It proposes to predefine a common subspace, in which the between-class variation is maximized while the within-class variation is minimized. Then, it trains m modality-specific networks for m modalities (one network for each modality) to transform the multimodal data into the predefined common subspace to achieve multimodal learning. Unlike many of the existing methods, our method can train different modality-specific networks independently and thus be scalable to the number of modalities. To the best of our knowledge, the proposed SDML could be one of the first works to independently project data of an unfixed number of modalities into a predefined common subspace. Comprehensive experimental results on four widely-used benchmark datasets demonstrate that the proposed method is effective and efficient in multimodal learning and outperforms the state-of-the-art methods in cross-modal retrieval.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most of existing cross-modal retrieval approaches were proposed to learn a common subspace in a joint manner, where the data from all modalities have to be involved during the whole training process. The optimal parameters of different modality-specific transformations are dependent on each other and the whole model has to be retrained when handling samples from new modalities.", "Key idea": "The author proposes a cross-modal retrieval method that predefines a common subspace, maximizing between-class and minimizing within-class variation, and trains m modality-specific networks for m modalities (one network for each modality) to transform the multimodal data into the predefined common subspace to achieve multimodal learning. The model can train different modality-specific networks independently.", "Method": "The author tests the model against four widely-used benchmark datasets for effectiveness and efficiency in multimodal learning.", "Outcome": "Experimental results on four widely-used benchmark datasets demonstrate that the proposed method is effective and efficient in multimodal learning and outperforms the state-of-the-art methods in cross-modal retrieval.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 29s "}} +{"id": "192f7803-df4d-40c0-b816-ba34339026b3", "displayed_text": "Title: Amortised MAP Inference for Image Super-resolution\n\nAbstract: Image super-resolution (SR) is an underdetermined inverse problem, where a large number of plausible high resolution images can explain the same downsampled image. Most current single image SR methods use empirical risk minimisation, often with a pixel-wise mean squared error (MSE) loss.\nHowever, the outputs from such methods tend to be blurry, over-smoothed and generally appear implausible. A more desirable approach would employ Maximum a Posteriori (MAP) inference, preferring solutions that always have a high probability under the image prior, and thus appear more plausible. Direct MAP estimation for SR is non-trivial, as it requires us to build a model for the image prior from samples. Here we introduce new methods for \\emph{amortised MAP inference} whereby we calculate the MAP estimate directly using a convolutional neural network. We first introduce a novel neural network architecture that performs a projection to the affine subspace of valid SR solutions ensuring that the high resolution output of the network is always consistent with the low resolution input. We show that, using this architecture, the amortised MAP inference problem reduces to minimising the cross-entropy between two distributions, similar to training generative models. We propose three methods to solve this optimisation problem: (1) Generative Adversarial Networks (GAN) (2) denoiser-guided SR which backpropagates gradient-estimates from denoising to train the network, and (3) a baseline method using a maximum-likelihood-trained image prior. Our experiments show that the GAN based approach performs best on real image data. Lastly, we establish a connection between GANs and amortised variational inference as in e.g. variational autoencoders.", "label_annotations": {"Multi-aspect Summary": {"Context": "Image super-resolution (SR) is an underdetermined inverse problem, where a large number of plausible high resolution images can explain the same downsampled image. Most current single image SR methods use empirical risk minimisation, often with a pixel-wise mean squared error (MSE) loss, but the outputs from such methods tend to be blurry, over-smoothed and generally appear implausible.", "Key idea": "The authors introduce new methods for amortised MAP inference whereby they calculate the MAP estimate directly using a convolutional neural network. They also propose three methods to solve this optimisation problem and establish a connection between GANs and amortised variational inference.", "Method": "N/A", "Outcome": "The authors show that, using their proposed architecture, the amortised MAP inference problem reduces to minimising the cross-entropy between two distributions, similar to training generative models. The experiments also show that the GAN based approach performs best on real image data.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 4s "}} +{"id": "192f7803-df4d-40c0-b816-ba34339026b3", "displayed_text": "Title: Amortised MAP Inference for Image Super-resolution\n\nAbstract: Image super-resolution (SR) is an underdetermined inverse problem, where a large number of plausible high resolution images can explain the same downsampled image. Most current single image SR methods use empirical risk minimisation, often with a pixel-wise mean squared error (MSE) loss.\nHowever, the outputs from such methods tend to be blurry, over-smoothed and generally appear implausible. A more desirable approach would employ Maximum a Posteriori (MAP) inference, preferring solutions that always have a high probability under the image prior, and thus appear more plausible. Direct MAP estimation for SR is non-trivial, as it requires us to build a model for the image prior from samples. Here we introduce new methods for \\emph{amortised MAP inference} whereby we calculate the MAP estimate directly using a convolutional neural network. We first introduce a novel neural network architecture that performs a projection to the affine subspace of valid SR solutions ensuring that the high resolution output of the network is always consistent with the low resolution input. We show that, using this architecture, the amortised MAP inference problem reduces to minimising the cross-entropy between two distributions, similar to training generative models. We propose three methods to solve this optimisation problem: (1) Generative Adversarial Networks (GAN) (2) denoiser-guided SR which backpropagates gradient-estimates from denoising to train the network, and (3) a baseline method using a maximum-likelihood-trained image prior. Our experiments show that the GAN based approach performs best on real image data. Lastly, we establish a connection between GANs and amortised variational inference as in e.g. variational autoencoders.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most current single image SR methods use empirical risk minimisation, often with a pixel-wise mean squared error (MSE) loss, but the outputs from such methods tend to be blurry, over-smoothed and generally appear implausible. A more desirable approach employs Maximum a Posteriori (MAP) inference, preferring solutions that always have a high probability under the image prior, and thus appear more plausible, which is not non-trivial as it requires us to build a model for the image prior from samples. ", "Key idea": "The author introduces new methods for amortised MAP inference by reducing the problem to minimising the cross-entropy between two distributions after using a novel neural network architecture that performs a projection to the affine subspace of valid SR solutions, and then proposes three methods to solve the reduced problem: (1) Generative Adversarial Networks (GAN) (2) denoiser-guided SR which backpropagates gradient-estimates from denoising to train the network, and (3) a baseline method using a maximum-likelihood-trained image prior.", "Method": "The author tests the proposed three reduced optimisation problem on real image data.", "Outcome": "The experiments show that the GAN based approach performs best on real image data.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 20s "}} +{"id": "1946f496-f6cd-4736-8c30-a6ae70baa8b2", "displayed_text": "Title: Content-aware click modeling\n\nAbstract: Click models aim at extracting intrinsic relevance of documents to queries from biased user clicks. One basic modeling assumption made in existing work is to treat such intrinsic relevance as an atomic query-document-specific parameter, which is solely estimated from historical clicks without using any content information about a document or relationship among the clicked/skipped documents under the same query. Due to this overly simplified assumption, existing click models can neither fully explore the information about a document's relevance quality nor make predictions of relevance for any unseen documents. In this work, we proposed a novel Bayesian Sequential State model for modeling the user click behaviors, where the document content and dependencies among the sequential click events within a query are characterized by a set of descriptive features via a probabilistic graphical model. By applying the posterior regularized Expectation Maximization algorithm for parameter learning, we tailor the model to meet specific ranking-oriented properties, e.g., pairwise click preferences, so as to exploit richer information buried in the user clicks. Experiment results on a large set of real click logs demonstrate the effectiveness of the proposed model compared with several state-of-the-art click models.", "label_annotations": {"Multi-aspect Summary": {"Context": "Due to overly simplified assumption that treat such intrinsic relevance as an atomic query-document-specific parameter, which is solely estimated from historical clicks without using any content information about a document or relationship among the clicked/skipped documents under the same query, existing click models can neither fully explore the information about a document's relevance quality nor make predictions of relevance for any unseen documents.", "Key idea": "The authors proposed a novel Bayesian Sequential State model for modeling the user click behaviors, where the document content and dependencies among the sequential click events within a query are characterized by a set of descriptive features via a probabilistic graphical model.", "Method": "The authors conduct experiments on a large set of real click logs to evaluate effectiveness of novel Bayesian Sequential State model ", "Outcome": "Experiment results on a large set of real click logs demonstrate the effectiveness of novel Bayesian Sequential State model compared with several state-of-the-art click models.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 37s "}} +{"id": "1946f496-f6cd-4736-8c30-a6ae70baa8b2", "displayed_text": "Title: Content-aware click modeling\n\nAbstract: Click models aim at extracting intrinsic relevance of documents to queries from biased user clicks. One basic modeling assumption made in existing work is to treat such intrinsic relevance as an atomic query-document-specific parameter, which is solely estimated from historical clicks without using any content information about a document or relationship among the clicked/skipped documents under the same query. Due to this overly simplified assumption, existing click models can neither fully explore the information about a document's relevance quality nor make predictions of relevance for any unseen documents. In this work, we proposed a novel Bayesian Sequential State model for modeling the user click behaviors, where the document content and dependencies among the sequential click events within a query are characterized by a set of descriptive features via a probabilistic graphical model. By applying the posterior regularized Expectation Maximization algorithm for parameter learning, we tailor the model to meet specific ranking-oriented properties, e.g., pairwise click preferences, so as to exploit richer information buried in the user clicks. Experiment results on a large set of real click logs demonstrate the effectiveness of the proposed model compared with several state-of-the-art click models.", "label_annotations": {"Multi-aspect Summary": {"Context": "One basic modeling assumption of Click models made in existing work is to treat the intrinsic relevance of documents to queries as an atomic query-document-specific parameter. Due to this overly simplified assumption, existing click models can neither fully explore the information about a document's relevance quality nor make predictions of relevance for any unseen documents.", "Key idea": "The authors propose a novel Bayesian Sequential State model for modeling the user click behaviors, where the document content and dependencies among the sequential click events within a query are characterized by a set of descriptive features via a probabilistic graphical model.", "Method": "The authors evaluate their model on a large set of real click logs.", "Outcome": "Experiment results on a large set of real click logs demonstrate the effectiveness of the proposed model compared with several state-of-the-art click models.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 11s "}} +{"id": "1b0e4045-d39b-4bea-8dec-e747f5c674f5", "displayed_text": "Title: Mining uncertain data with probabilistic guarantees\n\nAbstract: Data uncertainty is inherent in applications such as sensor monitoring systems, location-based services, and biological databases. To manage this vast amount of imprecise information, probabilistic databases have been recently developed. In this paper, we study the discovery of frequent patterns and association rules from probabilistic data under the Possible World Semantics. This is technically challenging, since a probabilistic database can have an exponential number of possible worlds. We propose two effcient algorithms, which discover frequent patterns in bottom-up and top-down manners. Both algorithms can be easily extended to discover maximal frequent patterns. We also explain how to use these patterns to generate association rules. Extensive experiments, using real and synthetic datasets, were conducted to validate the performance of our methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Probabilistic databases have been recently developed to manage data uncertainty such as sensor monitoring systems, location-based services, and biological databases. The difficulty of a probabilistic database is that it can have an exponential number of possible worlds.", "Key idea": "The author studies the discovery of frequent patterns and association rules from probabilistic data under the Possible World Semantics. The author proposes two efficient algorithms, which discover frequent patterns in bottom-up and top-down manners, both of which can be easily extended to discover maximal frequent patterns.", "Method": "The author uses real and synthetic datasets to test the performance of their methods.", "Outcome": "The performance of their methods is validated through real and synthetic datasets.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 31s "}} +{"id": "1b0e4045-d39b-4bea-8dec-e747f5c674f5", "displayed_text": "Title: Mining uncertain data with probabilistic guarantees\n\nAbstract: Data uncertainty is inherent in applications such as sensor monitoring systems, location-based services, and biological databases. To manage this vast amount of imprecise information, probabilistic databases have been recently developed. In this paper, we study the discovery of frequent patterns and association rules from probabilistic data under the Possible World Semantics. This is technically challenging, since a probabilistic database can have an exponential number of possible worlds. We propose two effcient algorithms, which discover frequent patterns in bottom-up and top-down manners. Both algorithms can be easily extended to discover maximal frequent patterns. We also explain how to use these patterns to generate association rules. Extensive experiments, using real and synthetic datasets, were conducted to validate the performance of our methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "To manage this vast amount of imprecise information, probabilistic databases have been recently developed.", "Key idea": "The authors propose two efficient algorithms, which discover frequent patterns in bottom-up and top-down manners.", "Method": "The authors perform extensive experiments, using real and synthetic datasets to evaluate two algorithms, which discover frequent patterns in bottom-up and top-down manners.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 4s "}} +{"id": "1dea5ec2-d311-4c03-bba5-e38d7a62fbd4", "displayed_text": "Title: Sign Spotting Using Hierarchical Sequential Patterns with Temporal Intervals\n\nAbstract: This paper tackles the problem of spotting a set of signs occuring in videos with sequences of signs. To achieve this, we propose to model the spatio-temporal signatures of a sign using an extension of sequential patterns that contain temporal intervals called Sequential Interval Patterns (SIP). We then propose a novel multi-class classifier that organises different sequential interval patterns in a hierarchical tree structure called a Hierarchical SIP Tree (HSP-Tree). This allows one to exploit any subsequence sharing that exists between different SIPs of different classes. Multiple trees are then combined together into a forest of HSP-Trees resulting in a strong classifier that can be used to spot signs. We then show how the HSP-Forest can be used to spot sequences of signs that occur in an input video. We have evaluated the method on both concatenated sequences of isolated signs and continuous sign sequences. We also show that the proposed method is superior in robustness and accuracy to a state of the art sign recogniser when applied to spotting a sequence of signs.", "label_annotations": {"Multi-aspect Summary": {"Context": "It's challenging to tackles the problem of spotting a set of signs occuring in videos with sequences of signs. ", "Key idea": "The authors propose a novel multi-class classifier that organises different sequential interval patterns in a hierarchical tree structure called a Hierarchical SIP Tree (HSP-Tree) to tackles the problem of spotting a set of signs occuring in videos with sequences of signs.", "Method": "The authors evaluated the HSP-Forest on both concatenated sequences of isolated signs and continuous sign sequences. ", "Outcome": "The authors prove that HSP-Forest can be used to spot sequences of signs that occur in an input video and the proposed method is superior in robustness and accuracy to a state of the art sign recogniser when applied to spotting a sequence of signs.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 58s "}} +{"id": "1dea5ec2-d311-4c03-bba5-e38d7a62fbd4", "displayed_text": "Title: Sign Spotting Using Hierarchical Sequential Patterns with Temporal Intervals\n\nAbstract: This paper tackles the problem of spotting a set of signs occuring in videos with sequences of signs. To achieve this, we propose to model the spatio-temporal signatures of a sign using an extension of sequential patterns that contain temporal intervals called Sequential Interval Patterns (SIP). We then propose a novel multi-class classifier that organises different sequential interval patterns in a hierarchical tree structure called a Hierarchical SIP Tree (HSP-Tree). This allows one to exploit any subsequence sharing that exists between different SIPs of different classes. Multiple trees are then combined together into a forest of HSP-Trees resulting in a strong classifier that can be used to spot signs. We then show how the HSP-Forest can be used to spot sequences of signs that occur in an input video. We have evaluated the method on both concatenated sequences of isolated signs and continuous sign sequences. We also show that the proposed method is superior in robustness and accuracy to a state of the art sign recogniser when applied to spotting a sequence of signs.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "To tackle the problem of spotting a set of signs occurring in videos with sequences of signs, the authors propose to model the spatio-temporal signatures of a sign using an extension of sequential patterns that contain temporal intervals called Sequential Interval Patterns (SIP). They further propose a novel multi-class classifier that organises different sequential interval patterns in a hierarchical tree structure called a Hierarchical SIP Tree (HSP-Tree).", "Method": "The authors evaluate the method on both concatenated sequences of isolated signs and continuous sign sequences. ", "Outcome": "The authors show that the proposed method is superior in robustness and accuracy to a state of the art sign recogniser when applied to spotting a sequence of signs.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 9s "}} +{"id": "1e396f93-a73e-4d33-9a8e-56097a8c3c28", "displayed_text": "Title: Deep Active Learning for Named Entity Recognition\n\nAbstract: Deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER). However, this typically requires large amounts of labeled data. In this work, we demonstrate that the amount of labeled training data can be drastically reduced when deep learning is combined with active learning. While active learning is sample-efficient, it can be computationally expensive since it requires iterative retraining. To speed this up, we introduce a lightweight architecture for NER, viz., the CNN-CNN-LSTM model consisting of convolutional character and word encoders and a long short term memory (LSTM) tag decoder. The model achieves nearly state-of-the-art performance on standard datasets for the task while being computationally much more efficient than best performing models. We carry out incremental active learning, during the training process, and are able to nearly match state-of-the-art performance with just 25\\% of the original training data.", "label_annotations": {"Multi-aspect Summary": {"Context": "Although deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER), it typically requires large amounts of labeled data. ", "Key idea": "The authors introduce a lightweight architecture for NER, viz., the CNN-CNN-LSTM model consisting of convolutional character and word encoders and a long short term memory (LSTM) tag decoder to speed up active learning.", "Method": "The authors evaluate CNN-CNN-LSTM model performance on standard datasets.", "Outcome": "CNN-CNN-LSTM model achieves nearly state-of-the-art performance on standard datasets for the task while being computationally much more efficient than best performing models. \r\nDuring the training process, incremental active learning has been carried out, CNN-CNN-LSTM mode are able to nearly match state-of-the-art performance with just 25% of the original training data.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 49m 13s "}} +{"id": "1e396f93-a73e-4d33-9a8e-56097a8c3c28", "displayed_text": "Title: Deep Active Learning for Named Entity Recognition\n\nAbstract: Deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER). However, this typically requires large amounts of labeled data. In this work, we demonstrate that the amount of labeled training data can be drastically reduced when deep learning is combined with active learning. While active learning is sample-efficient, it can be computationally expensive since it requires iterative retraining. To speed this up, we introduce a lightweight architecture for NER, viz., the CNN-CNN-LSTM model consisting of convolutional character and word encoders and a long short term memory (LSTM) tag decoder. The model achieves nearly state-of-the-art performance on standard datasets for the task while being computationally much more efficient than best performing models. We carry out incremental active learning, during the training process, and are able to nearly match state-of-the-art performance with just 25\\% of the original training data.", "label_annotations": {"Multi-aspect Summary": {"Context": "Deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER). However, this typically requires large amounts of labeled data.", "Key idea": "The authors demonstrate that the amount of labeled training data can be drastically reduced when deep learning is combined with active learning, and introduce a lightweight architecture for NER, viz., the CNN-CNN-LSTM model consisting of convolutional character and word encoders and a long short term memory (LSTM) tag decoder to speed up the process.", "Method": "The authors evaluate their method on standard datasets.", "Outcome": "The proposed model achieves nearly state-of-the-art performance on standard datasets for the task while being computationally much more efficient than best performing models. The authors carry out incremental active learning, during the training process, and are able to nearly match state-of-the-art performance with just 25% of the original training data.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 45s "}} +{"id": "1e5571af-71cf-40b8-ba2a-18b42cae5b42", "displayed_text": "Title: Knowledge-based sequence mining with ASP\n\nAbstract: We introduce a framework for knowledge-based sequence mining, based on Answer Set Programming (ASP). We begin by modeling the basic task and refine it in the sequel in several ways. First, we show how easily condensed patterns can be extracted by modular extensions of the basic approach. Second, we illustrate how ASPu0027s preference handling capacities can be exploited for mining patterns of interest. In doing so, we demonstrate the ease of incorporating knowledge into the ASP-based mining process. To assess the trade-off in effectiveness, we provide an empirical study comparing our approach with a related sequence mining mechanism.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce a framework for knowledge-based sequence mining, based on Answer Set Programming (ASP). ", "Method": "The authors provide an empirical study comparing new approach with a related sequence mining mechanism to assess the trade-off in effectiveness.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 18s "}} +{"id": "1e5571af-71cf-40b8-ba2a-18b42cae5b42", "displayed_text": "Title: Knowledge-based sequence mining with ASP\n\nAbstract: We introduce a framework for knowledge-based sequence mining, based on Answer Set Programming (ASP). We begin by modeling the basic task and refine it in the sequel in several ways. First, we show how easily condensed patterns can be extracted by modular extensions of the basic approach. Second, we illustrate how ASPu0027s preference handling capacities can be exploited for mining patterns of interest. In doing so, we demonstrate the ease of incorporating knowledge into the ASP-based mining process. To assess the trade-off in effectiveness, we provide an empirical study comparing our approach with a related sequence mining mechanism.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce a framework for knowledge-based sequence mining, based on Answer Set Programming (ASP).", "Method": "The authors demonstrate the ease of incorporating knowledge into the ASP-based mining process, and further provide an empirical study comparing the proposed approach with a related sequence mining mechanism.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 59s "}} +{"id": "1ef9b762-e9be-46c5-ad19-090fe16200c4", "displayed_text": "Title: Visual estimation of pointed targets for robot guidance via fusion of face pose and hand orientation\n\nAbstract: In this paper we address an important issue in human-robot interaction, that of accurately deriving pointing information from a corresponding gesture. Based on the fact that in most applications it is the pointed object rather than the actual pointing direction which is important, we formulate a novel approach which takes into account prior information about the location of possible pointing targets. To decide about the pointed object, the proposed approach uses the Dempster-Shafer theory of evidence to fuse information from two different input streams: head pose, estimated by visually tracking the off-plane rotations of the face, and hand pointing orientation. Detailed experimental results are presented that validate the effectiveness of the method in realistic application setups.", "label_annotations": {"Multi-aspect Summary": {"Context": "Accurately deriving pointing information from a corresponding gesture is an important issue in human-robot interaction.", "Key idea": "Based on the fact that in most applications it is the pointed object rather than the actual pointing direction which is important, authors formulate a novel approach which takes into account prior information about the location of possible pointing targets. \r\nTo decide about the pointed object, the proposed approach uses the Dempster-Shafer theory of evidence to fuse information from two different input streams: head pose, estimated by visually tracking the off-plane rotations of the face, and hand pointing orientation.", "Method": "Authors design detailed experiments to validate the propose method in realistic application setups..", "Outcome": "Authors present detailed experimental results that validate the effectiveness of the method in realistic application setups.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 59s "}} +{"id": "1ef9b762-e9be-46c5-ad19-090fe16200c4", "displayed_text": "Title: Visual estimation of pointed targets for robot guidance via fusion of face pose and hand orientation\n\nAbstract: In this paper we address an important issue in human-robot interaction, that of accurately deriving pointing information from a corresponding gesture. Based on the fact that in most applications it is the pointed object rather than the actual pointing direction which is important, we formulate a novel approach which takes into account prior information about the location of possible pointing targets. To decide about the pointed object, the proposed approach uses the Dempster-Shafer theory of evidence to fuse information from two different input streams: head pose, estimated by visually tracking the off-plane rotations of the face, and hand pointing orientation. Detailed experimental results are presented that validate the effectiveness of the method in realistic application setups.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors address the important issue of accurately deriving pointing information from a corresponding gesture in human-robot interaction. They formulate a novel approach which takes into account prior information about the location of possible pointing targets.", "Method": "The authors evaluate the proposed method in realistic application setups.", "Outcome": "Detailed experimental results are presented that validate the effectiveness of the method in realistic application setups.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 4s "}} +{"id": "2065b977-7782-4981-ad70-3121a2315687", "displayed_text": "Title: Correlated Bigram LSA for Unsupervised Language Model Adaptation\n\nAbstract: We present a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition. The model is trained using efficient variational EM and smoothed using the proposed fractional Kneser-Ney smoothing which handles fractional counts. We address the scalability issue to large training corpora via bootstrapping of bigram LSA from unigram LSA. For LM adaptation, unigram and bigram LSA are integrated into the background N-gram LM via marginal adaptation and linear interpolation respectively. Experimental results on the Mandarin RT04 test set show that applying unigram and bigram LSA together yields 6%-8% relative perplexity reduction and 2.5% relative character error rate reduction which is statistically significant compared to applying only unigram LSA. On the large-scale evaluation on Arabic, 3% relative word error rate reduction is achieved which is also statistically significant.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author presents a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition. The model is trained using efficient variational EM and smoothed using the proposed fractional Kneser-Ney smoothing. The scalability is achieved via bootstrapping of bigram LSA from unigram LSA, and LM adaptation is produced by integrating unigram and bigram LSA.", "Method": "The author tests the model against the Mandarin RT04 test set and on the large-scale evaluation on Arabic.", "Outcome": "Applying unigram and bigram LSA together yields 6%-8% relative perplexity reduction and 2.5% relative character error rate reduction on the Mandarin RT04 test set. 3% relative word error rate reduction is achieved on the large-scale evaluation on Arabic.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 28s "}} +{"id": "2065b977-7782-4981-ad70-3121a2315687", "displayed_text": "Title: Correlated Bigram LSA for Unsupervised Language Model Adaptation\n\nAbstract: We present a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition. The model is trained using efficient variational EM and smoothed using the proposed fractional Kneser-Ney smoothing which handles fractional counts. We address the scalability issue to large training corpora via bootstrapping of bigram LSA from unigram LSA. For LM adaptation, unigram and bigram LSA are integrated into the background N-gram LM via marginal adaptation and linear interpolation respectively. Experimental results on the Mandarin RT04 test set show that applying unigram and bigram LSA together yields 6%-8% relative perplexity reduction and 2.5% relative character error rate reduction which is statistically significant compared to applying only unigram LSA. On the large-scale evaluation on Arabic, 3% relative word error rate reduction is achieved which is also statistically significant.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition. The model is trained using efficient variational EM and smoothed using the proposed fractional Kneser-Ney smoothing which handles fractional counts. ", "Method": "The authors compare results of applying unigram and bigram LSA and applying only unigram LSA on the Mandarin RT04 test set and large-scale evaluation on Arabic. ", "Outcome": "On the Mandarin RT04 test set, applying unigram and bigram LSA together yields 6%-8% relative perplexity reduction and 2.5% relative character error rate reduction compared to applying only unigram LSA. On the large-scale evaluation on Arabic, 3% relative word error rate reduction is achieved by applyting unigram and bigram LSA together compared to applying only unigram LSA.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "206d2d53-dbaf-4a2f-810d-856309d8eb83", "displayed_text": "Title: Neural Head Reenactment with Latent Pose Descriptors\n\nAbstract: We propose a neural head reenactment system, which is driven by a latent pose representation and is capable of predicting the foreground segmentation alongside the RGB image. The latent pose representation is learned as a part of the entire reenactment system, and the learning process is based solely on image reconstruction losses. We show that despite its simplicity, with a large and diverse enough training dataset, such learning successfully decomposes pose from identity. The resulting system can then reproduce mimics of the driving person and, furthermore, can perform cross-person reenactment. Additionally, we show that the learned descriptors are useful for other pose-related tasks, such as keypoint prediction and pose-based retrieval. ", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author proposes a neural head reenactment system to predict the foreground segmentation alongside the RGB image. The system is driven by a latent pose representation learned as a part of the entire reenactment system, based solely on image reconstruction losses.", "Method": "The author trains the system with a large and diverse dataset.", "Outcome": "Against a large and diverse training dataset, the learning successfully decomposes pose from identity, and resulting system can then reproduce mimics of the driving person and perform cross-person reenactment. The learned descriptors are useful for other pose-related tasks like keypoint prediction and pose-based retrieval.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 41s "}} +{"id": "206d2d53-dbaf-4a2f-810d-856309d8eb83", "displayed_text": "Title: Neural Head Reenactment with Latent Pose Descriptors\n\nAbstract: We propose a neural head reenactment system, which is driven by a latent pose representation and is capable of predicting the foreground segmentation alongside the RGB image. The latent pose representation is learned as a part of the entire reenactment system, and the learning process is based solely on image reconstruction losses. We show that despite its simplicity, with a large and diverse enough training dataset, such learning successfully decomposes pose from identity. The resulting system can then reproduce mimics of the driving person and, furthermore, can perform cross-person reenactment. Additionally, we show that the learned descriptors are useful for other pose-related tasks, such as keypoint prediction and pose-based retrieval. ", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a neural head reenactment system, which is driven by a latent pose representation and is capable of predicting the foreground segmentation alongside the RGB image.", "Method": "N/A", "Outcome": "The authors show that their proposed method successfully decomposes pose from identity, and that the learned descriptors are useful for other pose-related tasks, such as keypoint prediction and pose-based retrieval.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 44s "}} +{"id": "24e29617-a320-450a-aaa5-19d8700d74b7", "displayed_text": "Title: NOMAD: non-locking, stochastic multi-machine algorithm for asynchronous and decentralized matrix completion\n\nAbstract: We develop an efficient parallel distributed algorithm for matrix completion, named NOMAD (Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion). NOMAD is a decentralized algorithm with non-blocking communication between processors. One of the key features of NOMAD is that the ownership of a variable is asynchronously transferred between processors in a decentralized fashion. As a consequence it is a lock-free parallel algorithm. In spite of being asynchronous, the variable updates of NOMAD are serializable, that is, there is an equivalent update ordering in a serial implementation. NOMAD outperforms synchronous algorithms which require explicit bulk synchronization after every iteration: our extensive empirical evaluation shows that not only does our algorithm perform well in distributed setting on commodity hardware, but also outperforms state-of-the-art algorithms on a HPC cluster both in multi-core and distributed memory settings.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors develop an efficient parallel distributed algorithm for matrix completion, named NOMAD (Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion).", "Method": "The authors evaluate NOMAD algorithm performance on commodity hardware and on a HPC cluster both in multi-core and distributed memory settings.", "Outcome": "Extensive empirical evaluation shows that NOMAD algorithm performs well in distributed setting on commodity hardware, and outperforms state-of-the-art algorithms on a HPC cluster both in multi-core and distributed memory settings.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 1s "}} +{"id": "24e29617-a320-450a-aaa5-19d8700d74b7", "displayed_text": "Title: NOMAD: non-locking, stochastic multi-machine algorithm for asynchronous and decentralized matrix completion\n\nAbstract: We develop an efficient parallel distributed algorithm for matrix completion, named NOMAD (Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion). NOMAD is a decentralized algorithm with non-blocking communication between processors. One of the key features of NOMAD is that the ownership of a variable is asynchronously transferred between processors in a decentralized fashion. As a consequence it is a lock-free parallel algorithm. In spite of being asynchronous, the variable updates of NOMAD are serializable, that is, there is an equivalent update ordering in a serial implementation. NOMAD outperforms synchronous algorithms which require explicit bulk synchronization after every iteration: our extensive empirical evaluation shows that not only does our algorithm perform well in distributed setting on commodity hardware, but also outperforms state-of-the-art algorithms on a HPC cluster both in multi-core and distributed memory settings.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors develop an efficient parallel distributed algorithm for matrix completion, named NOMAD (Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion).", "Method": "The authors evaluate their NOMAD algorithm in distributed setting on commodity hardware and in multi-core and distributed memory settings on a HPC cluster.", "Outcome": "NOMAD outperforms synchronous algorithms which require explicit bulk synchronization after every iteration. The extensive empirical evaluation shows that not only does our algorithm perform well in distributed setting on commodity hardware, but also outperforms state-of-the-art algorithms on a HPC cluster both in multi-core and distributed memory settings.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 36s "}} +{"id": "29dd9fd3-6c98-4e4b-b70c-0474ff361419", "displayed_text": "Title: Birds of a Feather: Capturing Avian Shape Models from Images\n\nAbstract: Animals are diverse in shape, but building a deformable shape model for a new species is not always possible due to the lack of 3D data. We present a method to capture new species using an articulated template and images of that species. In this work, we focus mainly on birds. Although birds represent almost twice the number of species as mammals, no accurate shape model is available. To capture a novel species, we first fit the articulated template to each training sample. By disentangling pose and shape, we learn a shape space that captures variation both among species and within each species from image evidence. We learn models of multiple species from the CUB dataset, and contribute new species-specific and multi-species shape models that are useful for downstream reconstruction tasks. Using a low-dimensional embedding, we show that our learned 3D shape space better reflects the phylogenetic relationships among birds than learned perceptual features.", "label_annotations": {"Multi-aspect Summary": {"Context": "Animals are diverse in shape, but building a deformable shape model for a new species is not always possible due to the lack of 3D data.", "Key idea": "The authors present a method to capture new species using an articulated template and images of that species. They learn models of multiple species from the CUB dataset, and contribute new species-specific and multi-species shape models that are useful for downstream reconstruction tasks.", "Method": "The authors compare their learned 3D shape space with learned perceptual features.", "Outcome": "Using a low-dimensional embedding, the authors show that the learned 3D shape space better reflects the phylogenetic relationships among birds than learned perceptual features.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 17s "}} +{"id": "29dd9fd3-6c98-4e4b-b70c-0474ff361419", "displayed_text": "Title: Birds of a Feather: Capturing Avian Shape Models from Images\n\nAbstract: Animals are diverse in shape, but building a deformable shape model for a new species is not always possible due to the lack of 3D data. We present a method to capture new species using an articulated template and images of that species. In this work, we focus mainly on birds. Although birds represent almost twice the number of species as mammals, no accurate shape model is available. To capture a novel species, we first fit the articulated template to each training sample. By disentangling pose and shape, we learn a shape space that captures variation both among species and within each species from image evidence. We learn models of multiple species from the CUB dataset, and contribute new species-specific and multi-species shape models that are useful for downstream reconstruction tasks. Using a low-dimensional embedding, we show that our learned 3D shape space better reflects the phylogenetic relationships among birds than learned perceptual features.", "label_annotations": {"Multi-aspect Summary": {"Context": "Building a deformable shape model for a new species is not always possible due to the lack of 3D data.", "Key idea": "The authors propose a method to capture new species using an articulated template and images of that species. ", "Method": "The authors learn models of multiple species from the CUB dataset, and contribute new species-specific and multi-species shape models that are useful for downstream reconstruction tasks. Using a low-dimensional embedding, the authors compare learned 3D shape space with learned perceptual features.", "Outcome": "Experiments prove that learned 3D shape space better reflects the phylogenetic relationships among birds than learned perceptual features.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 24s "}} +{"id": "2a220303-8653-497f-b2b5-c829583c2714", "displayed_text": "Title: Polynomial-Time Algorithms For Counting And Sampling Markov Equivalent Dags\n\nAbstract: Counting and sampling directed acyclic graphs from a Markov equivalence class are fundamental tasks in graphical causal analysis. In this paper, we show that these tasks can be performed in polynomial time, solving a long-standing open problem in this area. Our algorithms are effective and easily implementable. Experimental results show that the algorithms significantly outperform state-of-the-art methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Counting and sampling directed acyclic graphs from a Markov equivalence class are fundamental tasks in graphical causal analysis. ", "Key idea": "The authors propose algorithms that can solve the problems that counting and sampling directed acyclic graphs from a Markov equivalence class in polynomial time.", "Method": "N/A", "Outcome": "Experimental results show that the algorithms significantly outperform state-of-the-art methods on tasks that counting and sampling directed acyclic graphs from a Markov equivalence class.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 27s "}} +{"id": "2a220303-8653-497f-b2b5-c829583c2714", "displayed_text": "Title: Polynomial-Time Algorithms For Counting And Sampling Markov Equivalent Dags\n\nAbstract: Counting and sampling directed acyclic graphs from a Markov equivalence class are fundamental tasks in graphical causal analysis. In this paper, we show that these tasks can be performed in polynomial time, solving a long-standing open problem in this area. Our algorithms are effective and easily implementable. Experimental results show that the algorithms significantly outperform state-of-the-art methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Counting and sampling directed acyclic graphs from a Markov equivalence class are fundamental tasks in graphical causal analysis.", "Key idea": "The authors show that counting and sampling directed acyclic graphs from a Markov equivalence class can be performed in polynomial time, solving a long-standing open problem in this area.", "Method": "N/A", "Outcome": "Experimental results show that the proposed algorithms significantly outperform state-of-the-art methods.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 7s "}} +{"id": "2b5cc037-4841-4fb1-85ff-673230198be1", "displayed_text": "Title: Noise-Tolerant Life-Long Matrix Completion via Adaptive Sampling\n\nAbstract: We study the problem of recovering an incomplete $m\\times n$ matrix of rank $r$ with columns arriving online over time. This is known as the problem of life-long matrix completion, and is widely applied to recommendation system, computer vision, system identification, etc. The challenge is to design provable algorithms tolerant to a large amount of noises, with small sample complexity. In this work, we give algorithms achieving strong guarantee under two realistic noise models. In bounded deterministic noise, an adversary can add any bounded yet unstructured noise to each column. For this problem, we present an algorithm that returns a matrix of a small error, with sample complexity almost as small as the best prior results in the noiseless case. For sparse random noise, where the corrupted columns are sparse and drawn randomly, we give an algorithm that exactly recovers an $\\mu_0$-incoherent matrix by probability at least $1-\\delta$ with sample complexity as small as $O(\\mu_0rn\\log(r/\\delta))$. This result advances the state-of-the-art work and matches the lower bound in a worst case. We also study the scenario where the hidden matrix lies on a mixture of subspaces and show that the sample complexity can be even smaller. Our proposed algorithms perform well experimentally in both synthetic and real-world datasets.", "label_annotations": {"Multi-aspect Summary": {"Context": "It's challenging to design provable algorithms that can recover an incomplete m* n matrix of rank r, while being tolerant to a large amount of noises, with small sample complexity. ", "Key idea": "The authors propose algorithms that can recover an incomplete m* n matrix of rank r, providing strong guarantee under two realistic noise models.", "Method": "The authors compare new algorithms performance in both bounded deterministic noise and noiseless case.\r\nThe authors study the new algorithms performance under scenario where the hidden matrix lies on a mixture of subspaces.\r\nThe authors conduct experiments to evaluate new algorithms performance in both synthetic and real-world datasets.", "Outcome": "The authors prove that new algorithms can return a matrix of a small error, with sample complexity almost as small as the best prior results in the noiseless case and new algorithms perform well experimentally in both synthetic and real-world datasets.\r\nThe sample complexity can be even smaller when apply new algorithms under scenario where the hidden matrix lies on a mixture of subspaces .", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 10m 31s "}} +{"id": "2b5cc037-4841-4fb1-85ff-673230198be1", "displayed_text": "Title: Noise-Tolerant Life-Long Matrix Completion via Adaptive Sampling\n\nAbstract: We study the problem of recovering an incomplete $m\\times n$ matrix of rank $r$ with columns arriving online over time. This is known as the problem of life-long matrix completion, and is widely applied to recommendation system, computer vision, system identification, etc. The challenge is to design provable algorithms tolerant to a large amount of noises, with small sample complexity. In this work, we give algorithms achieving strong guarantee under two realistic noise models. In bounded deterministic noise, an adversary can add any bounded yet unstructured noise to each column. For this problem, we present an algorithm that returns a matrix of a small error, with sample complexity almost as small as the best prior results in the noiseless case. For sparse random noise, where the corrupted columns are sparse and drawn randomly, we give an algorithm that exactly recovers an $\\mu_0$-incoherent matrix by probability at least $1-\\delta$ with sample complexity as small as $O(\\mu_0rn\\log(r/\\delta))$. This result advances the state-of-the-art work and matches the lower bound in a worst case. We also study the scenario where the hidden matrix lies on a mixture of subspaces and show that the sample complexity can be even smaller. Our proposed algorithms perform well experimentally in both synthetic and real-world datasets.", "label_annotations": {"Multi-aspect Summary": {"Context": "The problem of life-long matrix completion is widely applied to recommendation system, computer vision, system identification. The challenge is to design provable algorithms tolerant to a large amount of noises, with small sample complexity.", "Key idea": "The authors give algorithms achieving strong guarantee under two realistic noise models.", "Method": "The authors evaluate their method in both synthetic and real-world datasets.\r\n", "Outcome": "The theoretical result advances the state-of-the-art work and matches the lower bound in a worst case.. The proposed algorithms also perform well experimentally in both synthetic and real-world datasets.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 10s "}} +{"id": "2eb66e5a-472c-4db1-b02f-47fe5eb7e71e", "displayed_text": "Title: Exploiting Inductive Bias in Transformers for Unsupervised Disentanglement of Syntax and Semantics with VAEs\n\nAbstract: We propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics. Contrary to previous work, this model does not need syntactic information such as constituency parses, or semantic information such as paraphrase pairs. Our model relies solely on the inductive bias found in attention-based architectures such as Transformers. In the attention of Transformers, keys handle information selection while values specify what information is conveyed. Our model, dubbed QKVAE, uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values. We run experiments on latent representations and experiments on syntax/semantics transfer which show that QKVAE displays clear signs of disentangled syntax and semantics. We also show that our model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer. The code for our experiments is publicly available. ", "label_annotations": {"Multi-aspect Summary": {"Context": "How to use the Inductive Bias in Transformers for Unsupervised Disentanglement is under explored.", "Key idea": "Authors propose a generative model, dubbed QKVAE, that uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values.\r\nThe model relies solely on the inductive bias found in attention-based architectures such as Transformers.", "Method": "Authors run experiments on latent representations and experiments on syntax/semantics transfer.", "Outcome": "Authors show that QKVAE displays clear signs of disentangled syntax and semantics. Authors also show that the model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 50s "}} +{"id": "2eb66e5a-472c-4db1-b02f-47fe5eb7e71e", "displayed_text": "Title: Exploiting Inductive Bias in Transformers for Unsupervised Disentanglement of Syntax and Semantics with VAEs\n\nAbstract: We propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics. Contrary to previous work, this model does not need syntactic information such as constituency parses, or semantic information such as paraphrase pairs. Our model relies solely on the inductive bias found in attention-based architectures such as Transformers. In the attention of Transformers, keys handle information selection while values specify what information is conveyed. Our model, dubbed QKVAE, uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values. We run experiments on latent representations and experiments on syntax/semantics transfer which show that QKVAE displays clear signs of disentangled syntax and semantics. We also show that our model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer. The code for our experiments is publicly available. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous generative model for text generation need syntactic information such as constituency parses, or semantic information such as paraphrase pairs.", "Key idea": "The authors propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics, QKVAE.", "Method": "The authors use their proposed QKVAE run experiments on latent representations and experiments on syntax/semantics transfer", "Outcome": "The experiments on latent representations and experiments on syntax/semantics transfer show that QKVAE displays clear signs of disentangled syntax and semantics. The authors also show that the model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 19s "}} +{"id": "31ab88b5-e66f-4b69-98b6-7a470dce9875", "displayed_text": "Title: Conditioning Sequence-to-sequence Networks with Learned Activations\n\nAbstract: Conditional neural networks play an important role in a number of sequence-to-sequence modeling tasks, including personalized sound enhancement (PSE), speaker dependent automatic speech recognition (ASR), and generative modeling such as text-to-speech synthesis. In conditional neural networks, the output of a model is often influenced by a conditioning vector, in addition to the input. Common approaches of conditioning include input concatenation or modulation with the conditioning vector, which comes at a cost of increased model size. In this work, we introduce a novel approach of neural network conditioning by learning intermediate layer activations based on the conditioning vector. We systematically explore and show that learned activation functions can produce conditional models with comparable or better quality, while decreasing model sizes, thus making them ideal candidates for resource-efficient on-device deployment. As exemplary target use-cases we consider (i) the task of PSE as a pre-processing technique for improving telephony or pre-trained ASR performance under noise, and (ii) personalized ASR in single speaker scenarios. We find that conditioning via activation function learning is an effective modeling strategy, suggesting a broad applicability of the proposed technique across a number of application domains.", "label_annotations": {"Multi-aspect Summary": {"Context": "Common approaches of conditioning include input concatenation or modulation with the conditioning vector, which comes at a cost of increased model size.", "Key idea": "The authors introduce a novel approach of neural network conditioning by learning intermediate layer activations based on the conditioning vector.", "Method": "The authors systematically explore novel neural network conditioning approach and conduct experiment to compare novel neural network conditioning approach's performance to traditional conditioning methods.", "Outcome": "Experiments demonstrate that conditional models using learned activation functions can have comparable or better quality than traditional methods, while decreasing model sizes, thus making them ideal candidates for resource-efficient on-device deployment.", "Future Impact": "This novel neural network conditioning approach may have a broad applicability of the proposed technique across a number of application domains."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "31ab88b5-e66f-4b69-98b6-7a470dce9875", "displayed_text": "Title: Conditioning Sequence-to-sequence Networks with Learned Activations\n\nAbstract: Conditional neural networks play an important role in a number of sequence-to-sequence modeling tasks, including personalized sound enhancement (PSE), speaker dependent automatic speech recognition (ASR), and generative modeling such as text-to-speech synthesis. In conditional neural networks, the output of a model is often influenced by a conditioning vector, in addition to the input. Common approaches of conditioning include input concatenation or modulation with the conditioning vector, which comes at a cost of increased model size. In this work, we introduce a novel approach of neural network conditioning by learning intermediate layer activations based on the conditioning vector. We systematically explore and show that learned activation functions can produce conditional models with comparable or better quality, while decreasing model sizes, thus making them ideal candidates for resource-efficient on-device deployment. As exemplary target use-cases we consider (i) the task of PSE as a pre-processing technique for improving telephony or pre-trained ASR performance under noise, and (ii) personalized ASR in single speaker scenarios. We find that conditioning via activation function learning is an effective modeling strategy, suggesting a broad applicability of the proposed technique across a number of application domains.", "label_annotations": {"Multi-aspect Summary": {"Context": "In conditional neural networks, the output of a model is often influenced by a conditioning vector, in addition to the input. Common approaches of conditioning include input concatenation or modulation with the conditioning vector, which comes at a cost of increased model size.", "Key idea": "The author introduces a novel approach of neural network conditioning by learning intermediate layer activations based on the conditioning vector.", "Method": "As exemplary target use-cases the author considers (i) the task of PSE as a pre-processing technique for improving telephony or pre-trained ASR performance under noise, and (ii) personalized ASR in single speaker scenarios.", "Outcome": "The learned activation functions can produce conditional models with comparable or better quality, while decreasing model sizes, thus making them ideal candidates for resource-efficient on-device deployment. The author finds that conditioning via activation function learning is an effective modeling strategy.", "Future Impact": "Conditioning via activation function learning has a broad applicability of the proposed technique across a number of application domains."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 34s "}} +{"id": "36d7073a-d006-4a4a-850f-ca7a3f0ca19b", "displayed_text": "Title: Learning convolutional neural networks for graphs\n\nAbstract: Numerous important problems can be framed as learning from graph data. We propose a framework for learning convolutional neural networks for arbitrary graphs. These graphs may be undirected, directed, and with both discrete and continuous node and edge attributes. Analogous to image-based convolutional networks that operate on locally connected regions of the input, we present a general approach to extracting locally connected regions from graphs. Using established benchmark data sets, we demonstrate that the learned feature representations are competitive with state of the art graph kernels and that their computation is highly efficient.", "label_annotations": {"Multi-aspect Summary": {"Context": "Numerous important problems can be framed as learning from graph data.", "Key idea": "The author proposes a framework for learning convolutional neural networks for arbitrary graphs, where the graphs may be undirected, directed, and with both discrete and continuous node and edge attributes. The author presents a general approach to extracting locally connected regions from graphs analogous to image-based convolutional networks.", "Method": "The authors test the learned feature representations and computation efficiency proposed framework on established benchmark data sets.", "Outcome": "Using established benchmark data sets, the author demonstrates that the learned feature representations are competitive with state of the art graph kernels and that their computation is highly efficient.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 26s "}} +{"id": "36d7073a-d006-4a4a-850f-ca7a3f0ca19b", "displayed_text": "Title: Learning convolutional neural networks for graphs\n\nAbstract: Numerous important problems can be framed as learning from graph data. We propose a framework for learning convolutional neural networks for arbitrary graphs. These graphs may be undirected, directed, and with both discrete and continuous node and edge attributes. Analogous to image-based convolutional networks that operate on locally connected regions of the input, we present a general approach to extracting locally connected regions from graphs. Using established benchmark data sets, we demonstrate that the learned feature representations are competitive with state of the art graph kernels and that their computation is highly efficient.", "label_annotations": {"Multi-aspect Summary": {"Context": "Numerous important problems can be framed as learning from graph data.", "Key idea": "The authors propose a framework for learning convolutional neural networks for arbitrary graphs.", "Method": "The authors evaluated their proposed framework on established benchmark data sets.", "Outcome": "The authors demonstrate that the learned feature representations are competitive with state of the art graph kernels and that their computation is highly efficient on established benchmark data sets.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 58s "}} +{"id": "370141c7-e1bb-4010-9938-efcad6cf2e62", "displayed_text": "Title: GradInit: Learning to Initialize Neural Networks for Stable and Efficient Training.\n\nAbstract: Innovations in neural architectures have fostered significant breakthroughs in language modeling and computer vision. Unfortunately, novel architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized. A number of architecture-specific initialization schemes have been proposed, but these schemes are not always portable to new architectures. This paper presents GradInit, an automated and architecture agnostic method for initializing neural networks. GradInit is based on a simple heuristic; the norm of each network layer is adjusted so that a single step of SGD or Adam with prescribed hyperparameters results in the smallest possible loss value. This adjustment is done by introducing a scalar multiplier variable in front of each parameter block, and then optimizing these variables using a simple numerical scheme. GradInit accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. It also improves the stability of the original Transformer architecture for machine translation, enabling training it without learning rate warmup using either Adam or SGD under a wide range of learning rates and momentum coefficients. Code is available at https://github.com/zhuchen03/gradinit.", "label_annotations": {"Multi-aspect Summary": {"Context": "Novel architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized. A number of architecture-specific initialization schemes have been proposed, but these schemes are not always portable to new architectures.", "Key idea": "This paper presents GradInit, an automated and architecture agnostic method for initializing neural networks. GradInit is based on a simple heuristic; the norm of each network layer is adjusted so that a single step of SGD or Adam with prescribed hyperparameters results in the smallest possible loss value.", "Method": "The authors test convergence and test performance of many convolutional architectures of Gradlnit. And compare Gradlnit's stability with original Transformer architectures' stability for machine translation.", "Outcome": "GradInit accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. GradInit also improves the stability of the original Transformer architecture for machine translation, enabling training it without learning rate warmup using either Adam or SGD under a wide range of learning rates and momentum coefficients. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "370141c7-e1bb-4010-9938-efcad6cf2e62", "displayed_text": "Title: GradInit: Learning to Initialize Neural Networks for Stable and Efficient Training.\n\nAbstract: Innovations in neural architectures have fostered significant breakthroughs in language modeling and computer vision. Unfortunately, novel architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized. A number of architecture-specific initialization schemes have been proposed, but these schemes are not always portable to new architectures. This paper presents GradInit, an automated and architecture agnostic method for initializing neural networks. GradInit is based on a simple heuristic; the norm of each network layer is adjusted so that a single step of SGD or Adam with prescribed hyperparameters results in the smallest possible loss value. This adjustment is done by introducing a scalar multiplier variable in front of each parameter block, and then optimizing these variables using a simple numerical scheme. GradInit accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. It also improves the stability of the original Transformer architecture for machine translation, enabling training it without learning rate warmup using either Adam or SGD under a wide range of learning rates and momentum coefficients. Code is available at https://github.com/zhuchen03/gradinit.", "label_annotations": {"Multi-aspect Summary": {"Context": "Novel neural architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized. Previously proposed architecture-specific initialization schemes are not always portable to new architectures.", "Key idea": "The author proposes an automated and architecture agnostic method for initializing neural networks, where the norm of each network layer is adjusted so that a single step of SGD or Adam with prescribed hyperparameters results in the smallest possible loss value. The adjustment is done by introducing a scalar multiplier variable in front of each parameter block, and then optimizing these variables using a simple numerical scheme.", "Method": "The author tests the proposed approach\u2019s performance against many convolutional architectures, both with or without skip connections, or without normalization layers. They also test stability of the approach against the original Transformer architecture for machine translation.", "Outcome": "The proposed approach accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. It also improves the stability of the original Transformer architecture for machine translation, enabling training it without learning rate warmup using either Adam or SGD under a wide range of learning rates and momentum coefficients.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 27s "}} +{"id": "37da719b-8db9-4523-8a4a-3f2eef238978", "displayed_text": "Title: Tamper-proof image watermarking using self-embedding\n\nAbstract: We propose a fragile watermarking with self-embedding for recovery of tampered image that does not use authentication bits. We use a robust spread spectrum based watermarking scheme using block based embedding, DCT based compression, and other improvements. Simulation results showing recovery performance are presented.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a fragile watermarking with self-embedding for recovery of tampered image that does not use authentication bits.", "Method": "The authors run simulation results of their proposed method.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 20s "}} +{"id": "37da719b-8db9-4523-8a4a-3f2eef238978", "displayed_text": "Title: Tamper-proof image watermarking using self-embedding\n\nAbstract: We propose a fragile watermarking with self-embedding for recovery of tampered image that does not use authentication bits. We use a robust spread spectrum based watermarking scheme using block based embedding, DCT based compression, and other improvements. Simulation results showing recovery performance are presented.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a fragile watermarking with self-embedding for recovery of tampered image that does not use authentication bits.", "Method": "The authors use simulation to evaluate fragile watermarking.", "Outcome": "Simulation results prove fragile watermarking recovery performance.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 14s "}} +{"id": "3ac464c2-2214-4bf2-a6b6-03da2498cb03", "displayed_text": "Title: Distribution-based query scheduling\n\nAbstract: Query scheduling, a fundamental problem in database management systems, has recently received a renewed attention, perhaps in part due to the rise of the database as a service\" (DaaS) model for database deployment. While there has been a great deal of work investigating different scheduling algorithms, there has been comparatively little work investigating what the scheduling algorithms can or should know about the queries to be scheduled. In this work, we investigate the efficacy of using histograms describing the distribution of likely query execution times as input to the query scheduler. We propose a novel distribution-based scheduling algorithm, Shepherd, and show that Shepherd substantially outperforms state-of-the-art point-based methods through extensive experimentation with both synthetic and TPC workloads.", "label_annotations": {"Multi-aspect Summary": {"Context": "Query scheduling, a fundamental problem in database management systems, has recently received a renewed attention, perhaps in part due to the rise of the database as a service\" (DaaS) model for database deployment. While there has been a great deal of work investigating different scheduling algorithms, there has been comparatively little work investigating what the scheduling algorithms can or should know about the queries to be scheduled.", "Key idea": "The authors investigate the efficacy of using histograms describing the distribution of likely query execution times as input to the query scheduler and propose a novel distribution-based scheduling algorithm, Shepherd.", "Method": "The authors evaluate their Shepherd through extensive experimentation with both synthetic and TPC workloads and compare it with state-of-the-art point-based methods .", "Outcome": "The authors show that their Shepherd substantially outperforms state-of-the-art point-based methods through extensive experimentation with both synthetic and TPC workloads.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 54s "}} +{"id": "3ac464c2-2214-4bf2-a6b6-03da2498cb03", "displayed_text": "Title: Distribution-based query scheduling\n\nAbstract: Query scheduling, a fundamental problem in database management systems, has recently received a renewed attention, perhaps in part due to the rise of the database as a service\" (DaaS) model for database deployment. While there has been a great deal of work investigating different scheduling algorithms, there has been comparatively little work investigating what the scheduling algorithms can or should know about the queries to be scheduled. In this work, we investigate the efficacy of using histograms describing the distribution of likely query execution times as input to the query scheduler. We propose a novel distribution-based scheduling algorithm, Shepherd, and show that Shepherd substantially outperforms state-of-the-art point-based methods through extensive experimentation with both synthetic and TPC workloads.", "label_annotations": {"Multi-aspect Summary": {"Context": "While there has been a great deal of work investigating different scheduling algorithms in database management systems, there has been comparatively little work investigating what the scheduling algorithms can or should know about the queries to be scheduled.", "Key idea": "The author proposes a novel distribution-based scheduling algorithm.", "Method": "The author compares the proposed algorithm with state-of-the-art point-based methods through extensive experimentation with both synthetic and TPC workloads.", "Outcome": "The author shows that the proposed algorithm substantially outperforms state-of-the-art point-based methods through extensive experimentation with both synthetic and TPC workloads.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 40s "}} +{"id": "3b076d81-3ed0-4d57-84e8-4145f67052bd", "displayed_text": "Title: THE COMPUTATIONAL COMPLEXITY OF AVOIDING CONVERSATIONAL IMPLICATURES\n\nAbstract: Referring expressions and other object descriptions should be maximal under the Local Brevity, No Unnecessary Components, and Lexical Preference preference rules; otherwise, they may lead hearers to infer unwanted conversational implicatures. These preference rules can be incorporated into a polynomial time generation algorithm, while some alternative formalizations of conversational implicature make the generation task NP-Hard.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "N/A", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 21s "}} +{"id": "3b076d81-3ed0-4d57-84e8-4145f67052bd", "displayed_text": "Title: THE COMPUTATIONAL COMPLEXITY OF AVOIDING CONVERSATIONAL IMPLICATURES\n\nAbstract: Referring expressions and other object descriptions should be maximal under the Local Brevity, No Unnecessary Components, and Lexical Preference preference rules; otherwise, they may lead hearers to infer unwanted conversational implicatures. These preference rules can be incorporated into a polynomial time generation algorithm, while some alternative formalizations of conversational implicature make the generation task NP-Hard.", "label_annotations": {"Multi-aspect Summary": {"Context": "Referring expressions and other object descriptions should be maximal under the Local Brevity, No Unnecessary Components, and Lexical Preference preference rules, or they may lead hearers to infer unwanted conversational implicatures. ", "Key idea": "N/A", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "3b6e7572-2ef8-4565-a3b7-301a3fd38acd", "displayed_text": "Title: SEDONA: Search for Decoupled Neural Networks toward Greedy Block-wise Learning\n\nAbstract: Backward locking and update locking are well-known sources of inefficiency in backpropagation that prevent from concurrently updating layers. Several works have recently suggested using local error signals to train network blocks asynchronously to overcome these limitations. However, they often require numerous iterations of trial-and-error to find the best configuration for local training, including how to decouple network blocks and which auxiliary networks to use for each block. In this work, we propose a differentiable search algorithm named SEDONA to automate this process. Experimental results show that our algorithm can consistently discover transferable decoupled architectures for VGG and ResNet variants, and significantly outperforms the ones trained with end-to-end backpropagation and other state-of-the-art greedy-leaning methods in CIFAR-10, Tiny-ImageNet and ImageNet. Thanks to improved parallelism by local training, we also report up to 2\u00d7 speedup over backpropagation in total training time.", "label_annotations": {"Multi-aspect Summary": {"Context": "Backward locking and update locking are sources of inefficiency in backpropagation that prevent from concurrently updating layers. Using local error signals to train network blocks asynchronously are suggested to overcome these limitations, but they often require numerous iterations of trial-and-error to find the best configuration for local training.", "Key idea": "The authors propose a differentiable search algorithm named SEDONA to automate process, carrying on numerous iterations of trial-and-error to find the best configuration for local training, including how to decouple network blocks and which auxiliary networks to use for each block.", "Method": "The authors apply SEDONA search algorithm on VGG and ResNet and record the time of backpropagation after using SEDONA search algorithm.", "Outcome": "SEDONA search algorithm consistently discover transferable decoupled architectures for VGG and ResNet variants, and significantly outperforms the ones trained with end-to-end backpropagation and other state-of-the-art greedy-leaning methods in CIFAR-10, Tiny-ImageNet and ImageNet.\r\nThere are 2\u00d7 speedup over backpropagation in total training time due to improved parallelism by local training.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "3b6e7572-2ef8-4565-a3b7-301a3fd38acd", "displayed_text": "Title: SEDONA: Search for Decoupled Neural Networks toward Greedy Block-wise Learning\n\nAbstract: Backward locking and update locking are well-known sources of inefficiency in backpropagation that prevent from concurrently updating layers. Several works have recently suggested using local error signals to train network blocks asynchronously to overcome these limitations. However, they often require numerous iterations of trial-and-error to find the best configuration for local training, including how to decouple network blocks and which auxiliary networks to use for each block. In this work, we propose a differentiable search algorithm named SEDONA to automate this process. Experimental results show that our algorithm can consistently discover transferable decoupled architectures for VGG and ResNet variants, and significantly outperforms the ones trained with end-to-end backpropagation and other state-of-the-art greedy-leaning methods in CIFAR-10, Tiny-ImageNet and ImageNet. Thanks to improved parallelism by local training, we also report up to 2\u00d7 speedup over backpropagation in total training time.", "label_annotations": {"Multi-aspect Summary": {"Context": "Backward locking and update locking are well-known sources of inefficiency in backpropagation that prevent from concurrently updating layers. Several works have recently suggested using local error signals to train network blocks asynchronously to overcome these limitations. However, they often require numerous iterations of trial-and-error to find the best configuration for local training, including how to decouple network blocks and which auxiliary networks to use for each block. ", "Key idea": "Authors propose a differentiable search algorithm named SEDONA to automate this process.", "Method": "Authors test the proposed method against other SOTA methods, in CIFAR-10, Tiny-ImageNet and ImageNet. ", "Outcome": "Experimental results show that our algorithm can consistently discover transferable decoupled architectures for VGG and ResNet variants, and significantly outperforms the ones trained with end-to-end backpropagation and other state-of-the-art greedy-leaning methods in CIFAR-10, Tiny-ImageNet and ImageNet. Authors also report up to 2\u00d7 speedup over backpropagation in total training time.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 20s "}} +{"id": "4164fd9a-7ab8-4013-b416-a205231f10f2", "displayed_text": "Title: Probabilistic plan recognition using off-the-shelf classical planners\n\nAbstract: Plan recognition is the problem of inferring the goals and plans of an agent after observing its behavior. Recently, it has been shown that this problem can be solved efficiently, without the need of a plan library, using slightly modified planning algorithms. In this work, we extend this approach to the more general problem of probabilistic plan recognition where a probability distribution over the set of goals is sought under the assumptions that actions have deterministic effects and both agent and observer have complete information about the initial state. We show that this problem can be solved efficiently using classical planners provided that the probability of a partially observed execution given a goal is defined in terms of the cost difference of achieving the goal under two conditions: complying with the observations, and not complying with them. This cost, and hence the posterior goal probabilities, are computed by means of two calls to a classical planner that no longer has to be modified in any way. A number of examples is considered to illustrate the quality, flexibility, and scalability of the approach.", "label_annotations": {"Multi-aspect Summary": {"Context": "It has been shown that plan recognition, which is the problem of inferring the goals and plans of an agent after observing its behavior, can be solved efficiently, without the need of a plan library, using slightly modified planning algorithms.", "Key idea": "The authors extend the method, using slightly modified planning algorithms, to the more general problem of probabilistic plan recognition where a probability distribution over the set of goals is sought under the assumptions that actions have deterministic effects and both agent and observer have complete information about the initial state. ", "Method": "The authors consider a number of examples to illustrate the quality, flexibility, and scalability of the new approach.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 18s "}} +{"id": "4164fd9a-7ab8-4013-b416-a205231f10f2", "displayed_text": "Title: Probabilistic plan recognition using off-the-shelf classical planners\n\nAbstract: Plan recognition is the problem of inferring the goals and plans of an agent after observing its behavior. Recently, it has been shown that this problem can be solved efficiently, without the need of a plan library, using slightly modified planning algorithms. In this work, we extend this approach to the more general problem of probabilistic plan recognition where a probability distribution over the set of goals is sought under the assumptions that actions have deterministic effects and both agent and observer have complete information about the initial state. We show that this problem can be solved efficiently using classical planners provided that the probability of a partially observed execution given a goal is defined in terms of the cost difference of achieving the goal under two conditions: complying with the observations, and not complying with them. This cost, and hence the posterior goal probabilities, are computed by means of two calls to a classical planner that no longer has to be modified in any way. A number of examples is considered to illustrate the quality, flexibility, and scalability of the approach.", "label_annotations": {"Multi-aspect Summary": {"Context": "Plan recognition is the problem of inferring the goals and plans of an agent after observing its behavior. Recently, it has been shown that this problem can be solved efficiently, without the need of a plan library, using slightly modified planning algorithms.", "Key idea": "The authors extend the slightly modified planning algorithms to the more general problem of probabilistic plan recognition, and show the efficiency of classical planners in solving this problem.", "Method": "N/A", "Outcome": "The authors illustrate the quality, flexibility, and scalability of the approach using a number of examples.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 18s "}} +{"id": "432f5702-7b7d-4995-812e-40925f1a18dd", "displayed_text": "Title: The Landmark Selection Method for Multiple Output Prediction\n\nAbstract: Conditional modeling x \u2192 y is a central problem in machine learning. A substantial research effort is devoted to such modeling when x is high dimensional. We consider, instead, the case of a high dimensional y, where x is either low dimensional or high dimensional. Our approach is based on selecting a small subset yL of the dimensions of y, and proceed by modeling (i) x \u2192 yL and (ii) yL \u2192 y. Composing these two models, we obtain a conditional model x \u2192 y that possesses convenient statistical properties. Multilabel classification and multivariate regression experiments on several datasets show that this method outperforms the one vs. all approach as well as several sophisticated multiple output prediction methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Conditional modeling x \u2192 y is a central problem in machine learning, and a substantial research effort is devoted to such modeling when x is high dimensional.", "Key idea": "The authors consider conditional modeling x \u2192 y with a high dimensional y by selecting a small subset yL of the dimensions of y, modeling (i) x \u2192 yL and (ii) yL \u2192 y and composing these two models.", "Method": "The authors run multilabel classification and multivariate regression experiments on several datasets.", "Outcome": "Multilabel classification and multivariate regression experiments on several datasets show that this method outperforms the one vs. all approach as well as several sophisticated multiple output prediction methods.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 50s "}} +{"id": "432f5702-7b7d-4995-812e-40925f1a18dd", "displayed_text": "Title: The Landmark Selection Method for Multiple Output Prediction\n\nAbstract: Conditional modeling x \u2192 y is a central problem in machine learning. A substantial research effort is devoted to such modeling when x is high dimensional. We consider, instead, the case of a high dimensional y, where x is either low dimensional or high dimensional. Our approach is based on selecting a small subset yL of the dimensions of y, and proceed by modeling (i) x \u2192 yL and (ii) yL \u2192 y. Composing these two models, we obtain a conditional model x \u2192 y that possesses convenient statistical properties. Multilabel classification and multivariate regression experiments on several datasets show that this method outperforms the one vs. all approach as well as several sophisticated multiple output prediction methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Conditional modeling x \u2192 y is a central problem in machine learning.", "Key idea": "The authors propose an method that composing two models, one is modeling (i) x \u2192 yL and the other one is modeling (ii) yL \u2192 y on selecting a small subset yL of the dimensions of y, and obtain a conditional model x \u2192 y that possesses convenient statistical properties in the end.", "Method": "The authors product multilabel classification and multivariate regression experiments on several dataset to evaluate the performance of new method.", "Outcome": "The new method, composing two models, outperforms the one vs. all approach as well as several sophisticated multiple output prediction methods on multilabel classification and multivariate regression tasks.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "438caf1c-5c7e-4283-a3cd-bbab302df185", "displayed_text": "Title: Active Learning by Acquiring Contrastive Examples.\n\nAbstract: Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively. In this work, leveraging the best of both worlds, we propose an acquisition function that opts for selecting \\textit{contrastive examples}, i.e. data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods. We compare our approach, CAL (Contrastive Active Learning), with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. Our experiments show that CAL performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data. We also conduct an extensive ablation study of our method and we further analyze all actively acquired datasets showing that CAL achieves a better trade-off between uncertainty and diversity compared to other strategies. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively.", "Key idea": "Leveraging the best of uncertainty and diversity sampling, the author proposes an acquisition function that opts for selecting contrastive examples, which are data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods.", "Method": "The author compares the proposed approach with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. The author also conducts an extensive ablation study of the method and analyzes all actively acquired datasets.", "Outcome": "The experiments show that proposed approach performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data. The proposed method also achieves a better trade-off between uncertainty and diversity compared to other strategies.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 47s "}} +{"id": "438caf1c-5c7e-4283-a3cd-bbab302df185", "displayed_text": "Title: Active Learning by Acquiring Contrastive Examples.\n\nAbstract: Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively. In this work, leveraging the best of both worlds, we propose an acquisition function that opts for selecting \\textit{contrastive examples}, i.e. data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods. We compare our approach, CAL (Contrastive Active Learning), with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. Our experiments show that CAL performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data. We also conduct an extensive ablation study of our method and we further analyze all actively acquired datasets showing that CAL achieves a better trade-off between uncertainty and diversity compared to other strategies. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively.", "Key idea": "The authors propose an acquisition function that opts for selecting \"contrastive examples\", i.e. data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods.", "Method": "The authors compare our approach, CAL (Contrastive Active Learning), with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. They also conduct an extensive ablation study of the proposed method", "Outcome": "The experiments show that CAL performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data. An extensive ablation study of the proposed method and further analysis show that CAL achieves a better trade-off between uncertainty and diversity compared to other strategies.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 23s "}} +{"id": "45b76955-9670-4664-939c-f5a61eb597df", "displayed_text": "Title: Quantifying the Impact of User Attentionon Fair Group Representation in Ranked Lists\n\nAbstract: In this work, we introduce a novel metric for auditing group fairness in ranked lists. Our approach offers two benefits compared to the state of the art. First, we offer a blueprint for modeling of user attention. Rather than assuming a logarithmic loss in importance as a function of the rank, we can account for varying user behaviors through parametrization. For example, we expect a user to see more items during a viewing of a social media feed than when they inspect the results list of a single web search query. Second, we allow non-binary protected attributes to enable investigating inherently continuous attributes (e.g., political alignment on the liberal to conservative spectrum) as well as to facilitate measurements across aggregated sets of search results, rather than separately for each result list. By combining these two elements into our metric, we are able to better address the human factors inherent in this problem. We measure the whole sociotechnical system, consisting of a ranking algorithm and individuals using it, instead of exclusively focusing on the ranking algorithm. Finally, we use our metric to perform three simulated fairness audits. We show that determining fairness of a ranked output necessitates knowledge (or a model) of the end-users of the particular service. Depending on their attention distribution function, a fixed ranking of results can appear biased both in favor and against a protected group1.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "auditing group fairness in ranked lists is an important problem.", "Key idea": "Authors introduce a novel metric for auditing group fairness in ranked lists. Authors first offer a blueprint for modeling of user attention. Second, authors allow non-binary protected attributes to enable investigating inherently continuous attributes (e.g., political alignment on the liberal to conservative spectrum) as well as to facilitate measurements across aggregated sets of search results, rather than separately for each result list.", "Method": "Authors use the proposed metric to perform three simulated fairness audits.", "Outcome": "The proposed metrics are able to better address the human factors inherent in this problem. Authors show that determining fairness of a ranked output necessitates knowledge (or a model) of the end-users of the particular service. Depending on their attention distribution function, a fixed ranking of results can appear biased both in favor and against a protected group1.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 8s "}} +{"id": "45b76955-9670-4664-939c-f5a61eb597df", "displayed_text": "Title: Quantifying the Impact of User Attentionon Fair Group Representation in Ranked Lists\n\nAbstract: In this work, we introduce a novel metric for auditing group fairness in ranked lists. Our approach offers two benefits compared to the state of the art. First, we offer a blueprint for modeling of user attention. Rather than assuming a logarithmic loss in importance as a function of the rank, we can account for varying user behaviors through parametrization. For example, we expect a user to see more items during a viewing of a social media feed than when they inspect the results list of a single web search query. Second, we allow non-binary protected attributes to enable investigating inherently continuous attributes (e.g., political alignment on the liberal to conservative spectrum) as well as to facilitate measurements across aggregated sets of search results, rather than separately for each result list. By combining these two elements into our metric, we are able to better address the human factors inherent in this problem. We measure the whole sociotechnical system, consisting of a ranking algorithm and individuals using it, instead of exclusively focusing on the ranking algorithm. Finally, we use our metric to perform three simulated fairness audits. We show that determining fairness of a ranked output necessitates knowledge (or a model) of the end-users of the particular service. Depending on their attention distribution function, a fixed ranking of results can appear biased both in favor and against a protected group1.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous metrics for auditing group fairness in ranked lists cannot exclude the inherent human factors.", "Key idea": "The authors introduce a novel metric for auditing group fairness in ranked lists, which can be able to better address the human factors inherent in this problem.", "Method": "The authors use novel metric to perform three simulated fairness audits.", "Outcome": "The authors conclude that determining fairness of a ranked output necessitates knowledge (or a model) of the end-users of the particular service. Depending on attention distribution function, a fixed ranking of results can appear biased both in favor and against a protected group1.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "48bacac1-1ca9-4be8-90e6-470596de0e26", "displayed_text": "Title: Building Lexicon for Sentiment Analysis from Massive Collection of HTML Documents\n\nAbstract: Recognizing polarity requires a list of polar words and phrases. For the purpose of building such lexicon automatically, a lot of studies have investigated (semi-) unsupervised method of learning polarity of words and phrases. In this paper, we explore to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences. The key idea is to develop the structural clues so that it achieves extremely high precision at the cost of recall. In order to compensate for the low recall, we used massive collection of HTML documents. Thus, we could prepare enough polar sentence corpus.", "label_annotations": {"Multi-aspect Summary": {"Context": "Recognizing polarity requires a list of polar words and phrases. For the purpose of building such lexicon automatically, a lot of studies have investigated (semi-) unsupervised method of learning polarity of words and phrases.", "Key idea": "Authors propose to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences. The key idea is to develop the structural clues so that it achieves extremely high precision at the cost of recall.", "Method": "Authors design experiments with massive collection of HTML documents.", "Outcome": "The proposed method can effectively achieve high precision with low recall.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 57s "}} +{"id": "48bacac1-1ca9-4be8-90e6-470596de0e26", "displayed_text": "Title: Building Lexicon for Sentiment Analysis from Massive Collection of HTML Documents\n\nAbstract: Recognizing polarity requires a list of polar words and phrases. For the purpose of building such lexicon automatically, a lot of studies have investigated (semi-) unsupervised method of learning polarity of words and phrases. In this paper, we explore to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences. The key idea is to develop the structural clues so that it achieves extremely high precision at the cost of recall. In order to compensate for the low recall, we used massive collection of HTML documents. Thus, we could prepare enough polar sentence corpus.", "label_annotations": {"Multi-aspect Summary": {"Context": "Recognizing polarity requires a list of polar words and phrases. In order to building such lexicon automatically, many studies have investigated (semi-)unsupervised method of learning polarity of words and phrases. ", "Key idea": "The authors explore to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences and develop the structural clues so that it achieves extremely high precision at the cost of recall.", "Method": "The author use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences.", "Outcome": "The author prepare enough polar sentence corpus used to recognizing polarity.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 7s "}} +{"id": "49309d5a-5959-4f8f-ae30-9fd2350f0cbc", "displayed_text": "Title: Focused Quantization for Sparse CNNs\n\nAbstract: Deep convolutional neural networks (CNNs) are powerful tools for a wide range of vision tasks, but the enormous amount of memory and compute resources required by CNNs poses a challenge in deploying them on constrained devices. Existing compression techniques, while excelling at reducing model sizes, struggle to be computationally friendly. In this paper, we attend to the statistical properties of sparse CNNs and present focused quantization, a novel quantization strategy based on power-of-two values, which exploits the weight distributions after fine-grained pruning. The proposed method dynamically discovers the most effective numerical representation for weights in layers with varying sparsities, significantly reducing model sizes. Multiplications in quantized CNNs are replaced with much cheaper bit-shift operations for efficient inference. Coupled with lossless encoding, we build a compression pipeline that provides CNNs with high compression ratios (CR), low computation cost and minimal loss in accuracies. In ResNet-50, we achieved a 18.08x CR with only 0.24% loss in top-5 accuracy, outperforming existing compression methods. We fully compress a ResNet-18 and found that it is not only higher in CR and top-5 accuracy, but also more hardware efficient as it requires fewer logic gates to implement when compared to other state-of-the-art quantization methods assuming the same throughput.", "label_annotations": {"Multi-aspect Summary": {"Context": "CNNs poses a challenge in deploying them on constrained devices since it requires enormous amount of memory and compute resources. Existing compression techniques, while excelling at reducing model sizes, struggle to be computationally friendly.", "Key idea": "The authors propose a novel quantization strategy based on power-of-two values, which exploits the weight distributions after fine-grained pruning when attending to the statistical properties of sparse CNNs and present focused quantization.", "Method": "The authors evaluate novel quantization strategy CR and top-5 accuracy on ResNet-50 and ResNet-18", "Outcome": "In ResNet-50, novel quantization strategy achieved a 18.08x CR with only 0.24% loss in top-5 accuracy, outperforming existing compression methods.\r\nOn ResNet-18, novel quantization not only achieved higher R and top-5 accuracy, but also achieved more hardware efficient as it requires fewer logic gates to implement when compared to other state-of-the-art quantization methods assuming the same throughput.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 58s "}} +{"id": "49309d5a-5959-4f8f-ae30-9fd2350f0cbc", "displayed_text": "Title: Focused Quantization for Sparse CNNs\n\nAbstract: Deep convolutional neural networks (CNNs) are powerful tools for a wide range of vision tasks, but the enormous amount of memory and compute resources required by CNNs poses a challenge in deploying them on constrained devices. Existing compression techniques, while excelling at reducing model sizes, struggle to be computationally friendly. In this paper, we attend to the statistical properties of sparse CNNs and present focused quantization, a novel quantization strategy based on power-of-two values, which exploits the weight distributions after fine-grained pruning. The proposed method dynamically discovers the most effective numerical representation for weights in layers with varying sparsities, significantly reducing model sizes. Multiplications in quantized CNNs are replaced with much cheaper bit-shift operations for efficient inference. Coupled with lossless encoding, we build a compression pipeline that provides CNNs with high compression ratios (CR), low computation cost and minimal loss in accuracies. In ResNet-50, we achieved a 18.08x CR with only 0.24% loss in top-5 accuracy, outperforming existing compression methods. We fully compress a ResNet-18 and found that it is not only higher in CR and top-5 accuracy, but also more hardware efficient as it requires fewer logic gates to implement when compared to other state-of-the-art quantization methods assuming the same throughput.", "label_annotations": {"Multi-aspect Summary": {"Context": "Deep convolutional neural networks (CNNs) are powerful tools for a wide range of vision tasks, but the enormous amount of memory and compute resources required by CNNs poses a challenge in deploying them on constrained devices. Existing compression techniques, while excelling at reducing model sizes, struggle to be computationally friendly.", "Key idea": "The authors present focused quantization, a novel quantization strategy based on power-of-two values, which exploits the weight distributions after fine-grained pruning, based on the statistical properties of sparse CNNs.", "Method": "The authors apply their proposed compression method on ResNet-50 and ResNet-18.", "Outcome": "In ResNet-50, the proposed compression method achieved a 18.08x CR with only 0.24% loss in top-5 accuracy, outperforming existing compression methods. The authors also fully compress a ResNet-18 and find that it is not only higher in CR and top-5 accuracy, but also more hardware efficient", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 34s "}} +{"id": "4e6f8004-9384-4c5c-8d7f-265410a290df", "displayed_text": "Title: Multimodal Shape Completion via Conditional Generative Adversarial Networks\n\nAbstract: Several deep learning methods have been proposed for completing partial data from shape acquisition setups, i.e., filling the regions that were missing in the shape. These methods, however, only complete the partial shape with a single output, ignoring the ambiguity when reasoning the missing geometry. Hence, we pose a multi-modal shape completion problem, in which we seek to complete the partial shape with multiple outputs by learning a one-to-many mapping. We develop the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. Our approach distills the ambiguity by conditioning the completion on a learned multimodal distribution of possible results. We extensively evaluate the approach on several datasets that contain varying forms of shape incompleteness, and compare among several baseline methods and variants of our methods qualitatively and quantitatively, demonstrating the merit of our method in completing partial shapes with both diversity and quality.", "label_annotations": {"Multi-aspect Summary": {"Context": "Several deep learning methods have been proposed for completing partial data from shape acquisition setups, i.e., filling the regions that were missing in the shape. These methods, however, only complete the partial shape with a single output, ignoring the ambiguity when reasoning the missing geometry. ", "Key idea": "Authors pose a multi-modal shape completion problem, in which authors seek to complete the partial shape with multiple outputs by learning a one-to-many mapping. Authors develop the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. Authors' approach distills the ambiguity by conditioning the completion on a learned multimodal distribution of possible results.", "Method": "Authors build a testbed on several datasets that contain varying forms of shape incompleteness, and compare among several baseline methods and variants of proposed methods qualitatively and quantitatively.", "Outcome": "Authors extensively evaluate the approach on several datasets that contain varying forms of shape incompleteness, and compare among several baseline methods and variants of our methods qualitatively and quantitatively, demonstrating the merit of proposed method in completing partial shapes with both diversity and quality.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 20s "}} +{"id": "4e6f8004-9384-4c5c-8d7f-265410a290df", "displayed_text": "Title: Multimodal Shape Completion via Conditional Generative Adversarial Networks\n\nAbstract: Several deep learning methods have been proposed for completing partial data from shape acquisition setups, i.e., filling the regions that were missing in the shape. These methods, however, only complete the partial shape with a single output, ignoring the ambiguity when reasoning the missing geometry. Hence, we pose a multi-modal shape completion problem, in which we seek to complete the partial shape with multiple outputs by learning a one-to-many mapping. We develop the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. Our approach distills the ambiguity by conditioning the completion on a learned multimodal distribution of possible results. We extensively evaluate the approach on several datasets that contain varying forms of shape incompleteness, and compare among several baseline methods and variants of our methods qualitatively and quantitatively, demonstrating the merit of our method in completing partial shapes with both diversity and quality.", "label_annotations": {"Multi-aspect Summary": {"Context": "Several deep learning methods have been proposed for completing partial data from shape acquisition setups, i.e., filling the regions that were missing in the shape. These methods, however, only complete the partial shape with a single output, ignoring the ambiguity when reasoning the missing geometry.", "Key idea": "The authors propose a multi-modal shape completion problem, in which they seek to complete the partial shape with multiple outputs by learning a one-to-many mapping. They further develop the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data.", "Method": "The authors extensively evaluate the approach on several datasets that contain varying forms of shape incompleteness, and compare among several baseline methods.", "Outcome": "The variants of the proposed methods qualitatively and quantitatively demonstrate the merit in completing partial shapes with both diversity and quality.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "4fdcceeb-f50f-4c4f-8b92-5985498114f8", "displayed_text": "Title: The recovery of non-rigid motion from stereo images\n\nAbstract: The problem of recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images is discussed. The object undergoes uniform expansion and three-dimensional shearing about an unknown point in space, in addition to being subjected to rigid motion. Feature correspondence over multiple frames is assumed. The problem of recovering the three-dimensional motion uniquely is reduced to the (unique) solution of a set of homogeneous polynomial equations using algebraic geometry, the commutative algebra software package, MACAULAY, and the Fortran polynomial continuation program POLSYS. It is shown that, with four points correspondence, only two (stereo) snapshots are needed to determine the motion uniquely. u003e", "label_annotations": {"Multi-aspect Summary": {"Context": "The problem of recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images is complex. Because the object undergoes uniform expansion and three-dimensional shearing about an unknown point in space, in addition to being subjected to rigid motion. ", "Key idea": "The authors propose that the problem of recovering the three-dimensional motion uniquely can be reduced to the (unique) solution of a set of homogeneous polynomial equations using algebraic geometry, the commutative algebra software package, MACAULAY, and the Fortran polynomial continuation program POLSYS.", "Method": "The authors use this idea in solving the problem that how to determine the motion uniquely with four points correspondence.", "Outcome": "With four points correspondence, only two (stereo) snapshots are needed to determine the motion uniquely.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 11s "}} +{"id": "4fdcceeb-f50f-4c4f-8b92-5985498114f8", "displayed_text": "Title: The recovery of non-rigid motion from stereo images\n\nAbstract: The problem of recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images is discussed. The object undergoes uniform expansion and three-dimensional shearing about an unknown point in space, in addition to being subjected to rigid motion. Feature correspondence over multiple frames is assumed. The problem of recovering the three-dimensional motion uniquely is reduced to the (unique) solution of a set of homogeneous polynomial equations using algebraic geometry, the commutative algebra software package, MACAULAY, and the Fortran polynomial continuation program POLSYS. It is shown that, with four points correspondence, only two (stereo) snapshots are needed to determine the motion uniquely. u003e", "label_annotations": {"Multi-aspect Summary": {"Context": "The problem of recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images is discussed. The object undergoes uniform expansion and three-dimensional shearing about an unknown point in space, in addition to being subjected to rigid motion. Feature correspondence over multiple frames is assumed. ", "Key idea": "Authors reduce the problem of recovering the three-dimensional motion uniquely to the (unique) solution of a set of homogeneous polynomial equations using algebraic geometry, the commutative algebra software package, MACAULAY, and the Fortran polynomial continuation program POLSYS.\r\n", "Method": "N/A", "Outcome": "Authors show that, with four points correspondence, only two (stereo) snapshots are needed to determine the motion uniquely.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 5s "}} +{"id": "532e797a-4b72-488a-80e4-03713d3c8435", "displayed_text": "Title: Structure from Recurrent Motion: From Rigidity to Recurrency\n\nAbstract: This paper proposes a new method for Non-Rigid Structure-from-Motion (NRSfM) from a long monocular video sequence observing a non-rigid object performing recurrent and possibly repetitive dynamic action. Departing from the traditional idea of using linear low-order or low-rank shape model for the task of NRSfM, our method exploits the property of shape recurrency (i.e., many deforming shapes tend to repeat themselves in time). We show that recurrency is in fact a generalized rigidity. Based on this, we reduce NRSfM problems to rigid ones provided that certain recurrency condition is satisfied. Given such a reduction, standard rigid-SfM techniques are directly applicable (without any change) to the reconstruction of non-rigid dynamic shapes. To implement this idea as a practical approach, this paper develops efficient algorithms for automatic recurrency detection, as well as camera view clustering via a rigidity-check. Experiments on both simulated sequences and real data demonstrate the effectiveness of the method. Since this paper offers a novel perspective on rethinking structure-from-motion, we hope it will inspire other new problems in the field.", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous research proposes the traditional idea of using linear low-order or low-rank shape model for the task of Non-Rigid Structure-from-Motion (NRSfM).", "Key idea": "The authors propose a new method for Non-Rigid Structure-from-Motion (NRSfM) from a long monocular video sequence observing a non-rigid object performing recurrent and possibly repetitive dynamic action.", "Method": "The authors evaluate their proposed method on both simulated sequences and real data.", "Outcome": "Experiments on both simulated sequences and real data demonstrate the effectiveness of the method.", "Future Impact": "This paper offers a novel perspective on rethinking structure-from-motion, so it can potentially inspire other new problems in the field."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 14s "}} +{"id": "532e797a-4b72-488a-80e4-03713d3c8435", "displayed_text": "Title: Structure from Recurrent Motion: From Rigidity to Recurrency\n\nAbstract: This paper proposes a new method for Non-Rigid Structure-from-Motion (NRSfM) from a long monocular video sequence observing a non-rigid object performing recurrent and possibly repetitive dynamic action. Departing from the traditional idea of using linear low-order or low-rank shape model for the task of NRSfM, our method exploits the property of shape recurrency (i.e., many deforming shapes tend to repeat themselves in time). We show that recurrency is in fact a generalized rigidity. Based on this, we reduce NRSfM problems to rigid ones provided that certain recurrency condition is satisfied. Given such a reduction, standard rigid-SfM techniques are directly applicable (without any change) to the reconstruction of non-rigid dynamic shapes. To implement this idea as a practical approach, this paper develops efficient algorithms for automatic recurrency detection, as well as camera view clustering via a rigidity-check. Experiments on both simulated sequences and real data demonstrate the effectiveness of the method. Since this paper offers a novel perspective on rethinking structure-from-motion, we hope it will inspire other new problems in the field.", "label_annotations": {"Multi-aspect Summary": {"Context": "Traditional idea to approach the task of Non-Rigid Structure-from-Motion is by using linear low-order or low-rank shape model.", "Key idea": "The author develops algorithms for automatic recurrency detection and camera view clustering via a rigidity-check by applying standard rigid-SfM techniques to the reconstruction of non-rigid dynamic shapes.", "Method": "The author performs experiments on the proposed algorithms with simulated sequences and real data.", "Outcome": "Experiments on both simulated sequences and real data demonstrate the effectiveness of the method.", "Future Impact": "Since this paper offers a novel perspective on rethinking structure-from-motion, the author hopes it will inspire other new problems in the field."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 27s "}} +{"id": "56992082-e04e-4a8b-a985-abfea27fc2e0", "displayed_text": "Title: Dynamic Network Pruning with Interpretable Layerwise Channel Selection\n\nAbstract: Dynamic network pruning achieves runtime acceleration by dynamically determining the inference paths based on different inputs. However, previous methods directly generate continuous decision values for each weight channel, which cannot reflect a clear and interpretable pruning process. In this paper, we propose to explicitly model the discrete weight channel selections, which encourages more diverse weights utilization, and achieves more sparse runtime inference paths. Meanwhile, with the help of interpretable layerwise channel selections in the dynamic network, we can visualize the network decision paths explicitly for model interpretability. We observe that there are clear differences in the layerwise decisions between normal and adversarial examples. Therefore, we propose a novel adversarial example detection algorithm by discriminating the runtime decision features. Experiments show that our dynamic network achieves higher prediction accuracy under the similar computing budgets on CIFAR10 and ImageNet datasets compared to traditional static pruning methods and other dynamic pruning approaches. The proposed adversarial detection algorithm can significantly improve the state-of-the-art detection rate across multiple attacks, which provides an opportunity to build an interpretable and robust model.", "label_annotations": {"Multi-aspect Summary": {"Context": "Dynamic network pruning achieves runtime acceleration by dynamically determining the inference paths based on different inputs. However, previous methods directly generate continuous decision values for each weight channel, which cannot reflect a clear and interpretable pruning process.", "Key idea": "The author proposes to explicitly model the discrete weight channel selections, which encourages more diverse weights utilization, and achieves more sparse runtime inference paths. The author also proposes a novel adversarial example detection algorithm by discriminating the runtime decision features.", "Method": "The author compares the prediction accuracy of the proposed network on the similar computing budgets on CIFAR10 and ImageNet datasets with traditional static pruning methods and other dynamic pruning approaches.", "Outcome": "Experiments show that the proposed dynamic network achieves higher prediction accuracy under the similar computing budgets on CIFAR10 and ImageNet datasets compared to traditional static pruning methods and other dynamic pruning approaches.", "Future Impact": "The proposed adversarial detection algorithm can significantly improve the state-of-the-art detection rate across multiple attacks, which provides an opportunity to build an interpretable and robust model."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 10m 16s "}} +{"id": "56992082-e04e-4a8b-a985-abfea27fc2e0", "displayed_text": "Title: Dynamic Network Pruning with Interpretable Layerwise Channel Selection\n\nAbstract: Dynamic network pruning achieves runtime acceleration by dynamically determining the inference paths based on different inputs. However, previous methods directly generate continuous decision values for each weight channel, which cannot reflect a clear and interpretable pruning process. In this paper, we propose to explicitly model the discrete weight channel selections, which encourages more diverse weights utilization, and achieves more sparse runtime inference paths. Meanwhile, with the help of interpretable layerwise channel selections in the dynamic network, we can visualize the network decision paths explicitly for model interpretability. We observe that there are clear differences in the layerwise decisions between normal and adversarial examples. Therefore, we propose a novel adversarial example detection algorithm by discriminating the runtime decision features. Experiments show that our dynamic network achieves higher prediction accuracy under the similar computing budgets on CIFAR10 and ImageNet datasets compared to traditional static pruning methods and other dynamic pruning approaches. The proposed adversarial detection algorithm can significantly improve the state-of-the-art detection rate across multiple attacks, which provides an opportunity to build an interpretable and robust model.", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous dynamic network pruning methods directly generate continuous decision values for each weight channel, which cannot reflect a clear and interpretable pruning process. \r\nThere are clear differences in the layerwise decisions between normal and adversarial examples. ", "Key idea": "The authors propose to explicitly model the discrete weight channel selections, which encourages more diverse weights utilization, and achieves more sparse runtime inference paths.\r\nThe authors propose a novel adversarial example detection algorithm by discriminating the runtime decision features.", "Method": "The authors compare their novel dynamic network with traditional static pruning methods and other dynamic pruning approaches on CIFAR10 and ImageNet datasets under the similar computing budgets.", "Outcome": "Novel dynamic network achieves higher prediction accuracy under the similar computing budgets on CIFAR10 and ImageNet datasets compared to traditional static pruning methods and other dynamic pruning approaches. \r\nThe proposed adversarial detection algorithm can significantly improve the state-of-the-art detection rate across multiple attacks.", "Future Impact": "The proposed adversarial detection algorithm provides an opportunity to build an interpretable and robust model."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 34s "}} +{"id": "58279154-e623-46d2-a431-cc409e094e2c", "displayed_text": "Title: A model of analogy-driven proof-plan construction\n\nAbstract: This paper addresses a model of analogy-driven theorem proving that is more general and cognitively more adequate than previous approaches. The model works at the level of proof-plans. More precisely, we consider analogy as a control strategy in proof planning that employs a source proof-plan to guide the construction of a proof-plan for the target problem. Our approach includes a reformulation of the source proof-plan. This is in accordance with the well known fact that constructing an analogy in maths often amounts to first finding the appropriate representation which brings out the similarity of two problems, i.e., finding the right concepts and the right level of abstraction. Several well known theorems were processed by our analogy-driven proof-plan construction that could not be proven analogically by previous approaches.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "Authors propose a model that works at the level of proof-plans. More precisely, authors consider analogy as a control strategy in proof planning that employs a source proof-plan to guide the construction of a proof-plan for the target problem.", "Method": "Authors design experiments to test whether the proposed analogy-driven proof-plan construction can prove some well known theorems.", "Outcome": "Several well known theorems were processed by our analogy-driven proof-plan construction that could not be proven analogically by previous approaches.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 23s "}} +{"id": "58279154-e623-46d2-a431-cc409e094e2c", "displayed_text": "Title: A model of analogy-driven proof-plan construction\n\nAbstract: This paper addresses a model of analogy-driven theorem proving that is more general and cognitively more adequate than previous approaches. The model works at the level of proof-plans. More precisely, we consider analogy as a control strategy in proof planning that employs a source proof-plan to guide the construction of a proof-plan for the target problem. Our approach includes a reformulation of the source proof-plan. This is in accordance with the well known fact that constructing an analogy in maths often amounts to first finding the appropriate representation which brings out the similarity of two problems, i.e., finding the right concepts and the right level of abstraction. Several well known theorems were processed by our analogy-driven proof-plan construction that could not be proven analogically by previous approaches.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a model of analogy-driven theorem proving that is more general and cognitively more adequate than previous approaches.", "Method": "The authors use their proposed method to prove several well known theorems analogically.", "Outcome": "Several well known theorems were processed by the authors' analogy-driven proof-plan construction that could not be proven analogically by previous approaches.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 6s "}} +{"id": "5a3da6ef-67b0-41bc-a994-fc5ff455a27b", "displayed_text": "Title: Review spam detection via temporal pattern discovery\n\nAbstract: Online reviews play a crucial role in todayu0027s electronic commerce. It is desirable for a customer to read reviews of products or stores before making the decision of what or from where to buy. Due to the pervasive spam reviews, customers can be misled to buy low-quality products, while decent stores can be defamed by malicious reviews. We observe that, in reality, a great portion (u003e 90% in the data we study) of the reviewers write only one review (singleton review). These reviews are so enormous in number that they can almost determine a storeu0027s rating and impression. However, existing methods did not examine this larger part of the reviews. Are most of these singleton reviews truthful ones? If not, how to detect spam reviews in singleton reviews? We call this problem singleton review spam detection. To address this problem, we observe that the normal reviewersu0027 arrival pattern is stable and uncorrelated to their rating pattern temporally. In contrast, spam attacks are usually bursty and either positively or negatively correlated to the rating. Thus, we propose to detect such attacks via unusually correlated temporal patterns. We identify and construct multidimensional time series based on aggregate statistics, in order to depict and mine such correlations. In this way, the singleton review spam detection problem is mapped to a abnormally correlated pattern detection problem. We propose a hierarchical algorithm to robustly detect the time windows where such attacks are likely to have happened. The algorithm also pinpoints such windows in different time resolutions to facilitate faster human inspection. Experimental results show that the proposed method is effective in detecting singleton review attacks. We discover that singleton review is a significant source of spam reviews and largely affects the ratings of online stores.", "label_annotations": {"Multi-aspect Summary": {"Context": "Due to the pervasive spam reviews, customers can be misled to buy low-quality products, while decent stores can be defamed by malicious reviews.\r\nExisting methods do not adequately address this large subset of reviews.", "Key idea": "The authors propose addressing the issue of spam in singleton reviews via unusually correlated temporal patterns. And the authors propose a hierarchical algorithm to robustly detect the time windows where spam attacks are likely to have happened. ", "Method": "The authors conduct the experiment to evaluate the ability of detecting singleton review attacks using unusually correlated temporal patterns and hierarchical algorithm.", "Outcome": "Experimental results show that the proposed method is effective in detecting singleton review attacks. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "5a3da6ef-67b0-41bc-a994-fc5ff455a27b", "displayed_text": "Title: Review spam detection via temporal pattern discovery\n\nAbstract: Online reviews play a crucial role in todayu0027s electronic commerce. It is desirable for a customer to read reviews of products or stores before making the decision of what or from where to buy. Due to the pervasive spam reviews, customers can be misled to buy low-quality products, while decent stores can be defamed by malicious reviews. We observe that, in reality, a great portion (u003e 90% in the data we study) of the reviewers write only one review (singleton review). These reviews are so enormous in number that they can almost determine a storeu0027s rating and impression. However, existing methods did not examine this larger part of the reviews. Are most of these singleton reviews truthful ones? If not, how to detect spam reviews in singleton reviews? We call this problem singleton review spam detection. To address this problem, we observe that the normal reviewersu0027 arrival pattern is stable and uncorrelated to their rating pattern temporally. In contrast, spam attacks are usually bursty and either positively or negatively correlated to the rating. Thus, we propose to detect such attacks via unusually correlated temporal patterns. We identify and construct multidimensional time series based on aggregate statistics, in order to depict and mine such correlations. In this way, the singleton review spam detection problem is mapped to a abnormally correlated pattern detection problem. We propose a hierarchical algorithm to robustly detect the time windows where such attacks are likely to have happened. The algorithm also pinpoints such windows in different time resolutions to facilitate faster human inspection. Experimental results show that the proposed method is effective in detecting singleton review attacks. We discover that singleton review is a significant source of spam reviews and largely affects the ratings of online stores.", "label_annotations": {"Multi-aspect Summary": {"Context": "In reality, a great portion (u003e 90% in the data we study) of the reviewers write only one review (singleton review). However, existing methods did not examine this larger part of the reviews.\r\n\r\nAuthors observe that the normal reviewersu0027 arrival pattern is stable and uncorrelated to their rating pattern temporally. In contrast, spam attacks are usually bursty and either positively or negatively correlated to the rating", "Key idea": "Authors identify and construct multidimensional time series based on aggregate statistics, in order to depict and mine such correlations. In this way, the singleton review spam detection problem is mapped to a abnormally correlated pattern detection problem. Authors propose a hierarchical algorithm to robustly detect the time windows where such attacks are likely to have happened.", "Method": "Authors setup benchmark to detect singleton review attacks.", "Outcome": " Experimental results show that the proposed method is effective in detecting singleton review attacks. We discover that singleton review is a significant source of spam reviews and largely affects the ratings of online stores.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 56s "}} +{"id": "5b47ca6b-ff9b-429c-adb1-ad9a171eea0e", "displayed_text": "Title: Domain Adaptation of Maximum Entropy Language Models\n\nAbstract: We investigate a recently proposed Bayesian adaptation method for building style-adapted maximum entropy language models for speech recognition, given a large corpus of written language data and a small corpus of speech transcripts. Experiments show that the method consistently outperforms linear interpolation which is typically used in such cases.", "label_annotations": {"Multi-aspect Summary": {"Context": "A Bayesian adaptation method for building style-adapted maximum entropy language models for speech recognition is recently proposed", "Key idea": "The authors investigate a recently proposed Bayesian adaptation method for building style-adapted maximum entropy language models for speech recognition, given a large corpus of written language data and a small corpus of speech transcripts.", "Method": "N/A", "Outcome": "Experiments show that the method consistently outperforms linear interpolation which is typically used in such cases.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 0s "}} +{"id": "5b47ca6b-ff9b-429c-adb1-ad9a171eea0e", "displayed_text": "Title: Domain Adaptation of Maximum Entropy Language Models\n\nAbstract: We investigate a recently proposed Bayesian adaptation method for building style-adapted maximum entropy language models for speech recognition, given a large corpus of written language data and a small corpus of speech transcripts. Experiments show that the method consistently outperforms linear interpolation which is typically used in such cases.", "label_annotations": {"Multi-aspect Summary": {"Context": "A Bayesian adaptation method is recently proposed.", "Key idea": "The authors investigate a Bayesian adaptation method for building style-adapted maximum entropy language models for speech recognition, given a large corpus of written language data and a small corpus of speech transcripts.", "Method": "N/A", "Outcome": "Experiments show that the method consistently outperforms linear interpolation which is typically used in such cases.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 25s "}} +{"id": "5b9f94f9-d93f-455d-a110-007ad67ada6d", "displayed_text": "Title: Coreference Handling in XMG\n\nAbstract: We claim that existing specification languages for tree based grammars fail to adequately support identifier managment. We then show that XMG (eXtensible Meta-Grammar) provides a sophisticated treatment of identifiers which is effective in supporting a linguist-friendly grammar design.", "label_annotations": {"Multi-aspect Summary": {"Context": "Existing specification languages for tree based grammars fail to adequately support identifier management.", "Key idea": "The authors show that XMG (eXtensible Meta-Grammar) provides a sophisticated treatment of identifiers which is effective in supporting a linguist-friendly grammar design.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 57s "}} +{"id": "5b9f94f9-d93f-455d-a110-007ad67ada6d", "displayed_text": "Title: Coreference Handling in XMG\n\nAbstract: We claim that existing specification languages for tree based grammars fail to adequately support identifier managment. We then show that XMG (eXtensible Meta-Grammar) provides a sophisticated treatment of identifiers which is effective in supporting a linguist-friendly grammar design.", "label_annotations": {"Multi-aspect Summary": {"Context": "Existing specification languages for tree based grammars fail to adequately support identifier managment. ", "Key idea": "The authors show that XMG (eXtensible Meta-Grammar) provides a sophisticated treatment of identifiers which is effective in supporting a linguist-friendly grammar design.", "Method": "N/A", "Outcome": "XMG is effective in supporting a linguist-friendly grammar design.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 51s "}} +{"id": "5c268324-e160-489d-9722-6b59d5e3471b", "displayed_text": "Title: Achieving accessibility with self-interested designers: a strategic knowledge-acquisition approach\n\nAbstract: We introduce a new approach towards a more accessible Web by means of more accessible knowledge acquisition mechanisms. Our strategy is to detect the Web designeru0027s needs for knowledge that can be collected from minorities of Web users, and subsequently to design mechanisms that allow the proper elicitation of such knowledge from Web users. We discuss how this scenario places marginal Web users in a privileged position that appeals for their inclusion. Additionally, we illustrate how this approach might help build a more accessible Web, to the benefit of visually-impaired knowledge contributors.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce a new approach towards a more accessible Web by means of more accessible knowledge acquisition mechanisms. ", "Method": "N/A", "Outcome": "This new approach places marginal Web users in a privileged position that appeals for their inclusion and it might help build a more accessible Web, to the benefit of visually-impaired knowledge contributors.", "Future Impact": "The proposed approach might help build a more accessible Web, to the benefit of visually-impaired knowledge contributors."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 36s "}} +{"id": "5c268324-e160-489d-9722-6b59d5e3471b", "displayed_text": "Title: Achieving accessibility with self-interested designers: a strategic knowledge-acquisition approach\n\nAbstract: We introduce a new approach towards a more accessible Web by means of more accessible knowledge acquisition mechanisms. Our strategy is to detect the Web designeru0027s needs for knowledge that can be collected from minorities of Web users, and subsequently to design mechanisms that allow the proper elicitation of such knowledge from Web users. We discuss how this scenario places marginal Web users in a privileged position that appeals for their inclusion. Additionally, we illustrate how this approach might help build a more accessible Web, to the benefit of visually-impaired knowledge contributors.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce a new approach towards a more accessible Web by means of more accessible knowledge acquisition mechanisms.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The proposed approach might help build a more accessible Web, to the benefit of visually-impaired knowledge contributors."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 47s "}} +{"id": "5d87de73-77d4-4efd-b8e1-d7561b13f69f", "displayed_text": "Title: Disambiguation of Preposition Sense Using Linguistically Motivated Features\n\nAbstract: In this paper, we present a supervised classification approach for disambiguation of preposition senses. We use the SemEval 2007 Preposition Sense Disambiguation datasets to evaluate our system and compare its results to those of the systems participating in the workshop. We derived linguistically motivated features from both sides of the preposition. Instead of restricting these to a fixed window size, we utilized the phrase structure. Testing with five different classifiers, we can report an increased accuracy that outperforms the best system in the SemEval task.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author presents a supervised classification approach for disambiguation of preposition senses.", "Method": "The author uses the SemEval 2007 Preposition Sense Disambiguation datasets to evaluate our system and compare its results to those of the systems participating in the workshop. The author also tests the proposed approach's performance on five different classifiers.", "Outcome": "The author derived linguistically motivated features from both sides of the preposition. Testing with five different classifiers, the author reports an increased accuracy that outperforms the best system in the SemEval task.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 49s "}} +{"id": "5d87de73-77d4-4efd-b8e1-d7561b13f69f", "displayed_text": "Title: Disambiguation of Preposition Sense Using Linguistically Motivated Features\n\nAbstract: In this paper, we present a supervised classification approach for disambiguation of preposition senses. We use the SemEval 2007 Preposition Sense Disambiguation datasets to evaluate our system and compare its results to those of the systems participating in the workshop. We derived linguistically motivated features from both sides of the preposition. Instead of restricting these to a fixed window size, we utilized the phrase structure. Testing with five different classifiers, we can report an increased accuracy that outperforms the best system in the SemEval task.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a supervised classification approach for disambiguation of preposition senses. ", "Method": "The authors use the SemEval 2007 Preposition Sense Disambiguation datasets to evaluate supervised classification approach and compare its results to those of the systems participating in the workshop.", "Outcome": "Testing with five different classifiers, the authors conclude that this supervised classification approach obtain an increased accuracy that outperforms the best system in the SemEval task.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 6s "}} +{"id": "5dccca98-2b58-47e3-9b8f-3b1888aa3976", "displayed_text": "Title: The zero-delay data warehouse: mobilizing heterogeneous database\n\nAbstract: Now is the time... for the real-time : In spite of this assertion from Gartner Group the heterogeneity of todayu0027s IT environments and the increasing demands from mobile users are major obstacles for the creation of this vision. Yet its technical foundation is available: software architectures based on innovative middleware components that offer a level of abstraction superior to conventional middleware solutions, including distributed transactions and the seamless integration of mobile devices using open standards, crossing the borders between heterogeneous platforms and systems. Space based computing is a new middleware paradigm meeting these demands. As an example we present the real time build-up of data warehouses.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "N/A", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 15s "}} +{"id": "5dccca98-2b58-47e3-9b8f-3b1888aa3976", "displayed_text": "Title: The zero-delay data warehouse: mobilizing heterogeneous database\n\nAbstract: Now is the time... for the real-time : In spite of this assertion from Gartner Group the heterogeneity of todayu0027s IT environments and the increasing demands from mobile users are major obstacles for the creation of this vision. Yet its technical foundation is available: software architectures based on innovative middleware components that offer a level of abstraction superior to conventional middleware solutions, including distributed transactions and the seamless integration of mobile devices using open standards, crossing the borders between heterogeneous platforms and systems. Space based computing is a new middleware paradigm meeting these demands. As an example we present the real time build-up of data warehouses.", "label_annotations": {"Multi-aspect Summary": {"Context": "The heterogeneity of current IT environments and the increasing demands from mobile users are major obstacles to achieve real-time data integration and processing", "Key idea": "The authors a new middleware paradigm, Space based computing, to meet all kinds of demands. ", "Method": "The authors use new middleware paradigm to create real-time data warehouses. ", "Outcome": " Space based computing is a level of abstraction superior to conventional middleware solutions, including distributed transactions and the seamless integration of mobile devices using open standards, crossing the borders between heterogeneous platforms and systems. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 24s "}} +{"id": "5e1f387c-d883-4d1f-8397-e4a533a3387b", "displayed_text": "Title: A Trainable Spectral-Spatial Sparse Coding Model for Hyperspectral Image Restoration.\n\nAbstract: Hyperspectral imaging offers new perspectives for diverse applications, ranging from the monitoring of the environment using airborne or satellite remote sensing, precision farming, food safety, planetary exploration, or astrophysics. Unfortunately, the spectral diversity of information comes at the expense of various sources of degradation, and the lack of accurate ground-truth \"clean\" hyperspectral signals acquired on the spot makes restoration tasks challenging. In particular, training deep neural networks for restoration is difficult, in contrast to traditional RGB imaging problems where deep models tend to shine. In this paper, we advocate instead for a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data. We show on various denoising benchmarks that our method is computationally efficient and significantly outperforms the state of the art.", "label_annotations": {"Multi-aspect Summary": {"Context": "Hyperspectral imaging offers new perspectives for diverse applications, but the spectral diversity of information comes at the expense of various sources of degradation, and the lack of accurate ground-truth \"clean\" hyperspectral signals acquired on the spot makes restoration tasks challenging.", "Key idea": "The authors propose a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data.", "Method": "The authors evaluate their proposed method on various denoising benchmarks.", "Outcome": "The authors show on various denoising benchmarks that the proposed method is computationally efficient and significantly outperforms the state of the art.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 11m 14s "}} +{"id": "5e1f387c-d883-4d1f-8397-e4a533a3387b", "displayed_text": "Title: A Trainable Spectral-Spatial Sparse Coding Model for Hyperspectral Image Restoration.\n\nAbstract: Hyperspectral imaging offers new perspectives for diverse applications, ranging from the monitoring of the environment using airborne or satellite remote sensing, precision farming, food safety, planetary exploration, or astrophysics. Unfortunately, the spectral diversity of information comes at the expense of various sources of degradation, and the lack of accurate ground-truth \"clean\" hyperspectral signals acquired on the spot makes restoration tasks challenging. In particular, training deep neural networks for restoration is difficult, in contrast to traditional RGB imaging problems where deep models tend to shine. In this paper, we advocate instead for a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data. We show on various denoising benchmarks that our method is computationally efficient and significantly outperforms the state of the art.", "label_annotations": {"Multi-aspect Summary": {"Context": "Hyperspectral imaging offers new perspectives for diverse applications. Unfortunately, the spectral diversity of information comes at the expense of various sources of degradation, and the lack of accurate ground-truth \"clean\" hyperspectral signals acquired on the spot makes restoration tasks challenging. In particular, training deep neural networks for restoration is difficult, in contrast to traditional RGB imaging problems.", "Key idea": "Authors advocate instead for a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data.", "Method": "Authors conduct experiments on various denoising benchmarks.", "Outcome": "Authors show that, on various denoising benchmarks the proposed method is computationally efficient and significantly outperforms the state of the art.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 12s "}} +{"id": "68f8d058-1403-4066-b3d3-a8a2836b35e1", "displayed_text": "Title: Quantifying and Reducing Registration Uncertainty of Spatial Vector Labels on Earth Imagery\n\nAbstract: Given raster imagery features and imperfect vector training labels with registration uncertainty, this paper studies a deep learning framework that can quantify and reduce the registration uncertainty of training labels as well as train neural network parameters simultaneously. The problem is important in broad applications such as streamline classification on Earth imagery or tissue segmentation on medical imagery, whereby annotating precise vector labels is expensive and time-consuming. However, the problem is challenging due to the gap between the vector representation of class labels and the raster representation of image features and the need for training neural networks with uncertain label locations. Existing research on uncertain training labels often focuses on uncertainty in label class semantics or characterizes label registration uncertainty at the pixel level (not contiguous vectors). To fill the gap, this paper proposes a novel learning framework that explicitly quantifies vector labels' registration uncertainty. We propose a registration-uncertainty-aware loss function and design an iterative uncertainty reduction algorithm by re-estimating the posterior of true vector label locations distribution based on a Gaussian process. Evaluations on real-world datasets in National Hydrography Dataset refinement show that the proposed approach significantly outperforms several baselines in the registration uncertainty estimations performance and classification performance.", "label_annotations": {"Multi-aspect Summary": {"Context": "Raster imagery features and imperfect vector training labels have registration uncertainty.\r\nThe problem is challenging due to the gap between the vector representation of class labels and the raster representation of image features and the need for training neural networks with uncertain label locations.", "Key idea": "The authors propose a novel learning framework that explicitly quantifies vector labels' registration uncertainty and a registration-uncertainty-aware loss function and design an iterative uncertainty reduction algorithm by re-estimating the posterior of true vector label locations distribution based on a Gaussian process.", "Method": "The authors evaluate novel learning framework on real-world datasets in National Hydrography Dataset refinement ", "Outcome": "Evaluations on real-world datasets in National Hydrography Dataset refinement show that the novel learning framework significantly outperforms several baselines in the registration uncertainty estimations performance and classification performance.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 31s "}} +{"id": "68f8d058-1403-4066-b3d3-a8a2836b35e1", "displayed_text": "Title: Quantifying and Reducing Registration Uncertainty of Spatial Vector Labels on Earth Imagery\n\nAbstract: Given raster imagery features and imperfect vector training labels with registration uncertainty, this paper studies a deep learning framework that can quantify and reduce the registration uncertainty of training labels as well as train neural network parameters simultaneously. The problem is important in broad applications such as streamline classification on Earth imagery or tissue segmentation on medical imagery, whereby annotating precise vector labels is expensive and time-consuming. However, the problem is challenging due to the gap between the vector representation of class labels and the raster representation of image features and the need for training neural networks with uncertain label locations. Existing research on uncertain training labels often focuses on uncertainty in label class semantics or characterizes label registration uncertainty at the pixel level (not contiguous vectors). To fill the gap, this paper proposes a novel learning framework that explicitly quantifies vector labels' registration uncertainty. We propose a registration-uncertainty-aware loss function and design an iterative uncertainty reduction algorithm by re-estimating the posterior of true vector label locations distribution based on a Gaussian process. Evaluations on real-world datasets in National Hydrography Dataset refinement show that the proposed approach significantly outperforms several baselines in the registration uncertainty estimations performance and classification performance.", "label_annotations": {"Multi-aspect Summary": {"Context": "The problem of quantifying and reducing the registration uncertainty of training labels is important but challenging. Existing research on uncertain training labels often focuses on uncertainty in label class semantics or characterizes label registration uncertainty at the pixel level (not contiguous vectors).", "Key idea": "The authors propose a deep learning framework that can quantify and reduce the registration uncertainty of training labels as well as train neural network parameters simultaneously, specifically, a registration-uncertainty-aware loss function and an iterative uncertainty reduction algorithm by re-estimating the posterior of true vector label locations distribution based on a Gaussian process.", "Method": "The authors evaluate the proposed method on real-world datasets in National Hydrography Dataset refinement and compare it with several baselines.", "Outcome": "Evaluations on real-world datasets in National Hydrography Dataset refinement show that the proposed approach significantly outperforms several baselines in the registration uncertainty estimations performance and classification performance.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 43s "}} +{"id": "69aacc53-6730-4db0-b420-9a45b96a642e", "displayed_text": "Title: Planning from first principles for geometric constraint satisfaction\n\nAbstract: An important problem in geometric reasoning is to find the configuration of a collection of geometric bodies so as to satisfy a set of given constraints. Recently, it has been suggested that this problem can be solved efficiently by symbolically reasoning about geometry using a degrees of freedom analysis. The approach employs a set of specialized routines called plan fragments that specify how to change the configuration of a set of bodies to satisfy a new constraint while preserving existing constraints. In this paper we show how these plan fragments can be automatically synthesized using first principles about geometric bodies, actions, and topology.", "label_annotations": {"Multi-aspect Summary": {"Context": "An important problem in geometric reasoning is to find the configuration of a collection of geometric bodies so as to satisfy a set of given constraints. ", "Key idea": "The authors introduce an approach solving geometric reasoning problem efficiently by symbolically reasoning about geometry using a degrees of freedom analysis. ", "Method": "The authors use new method to show how these plan fragments can be automatically synthesized using first principles about geometric bodies, actions, and topology.", "Outcome": "New method can efficiently solve geometric configuration problems, satisfying a new constraint while preserving existing constraints.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "69aacc53-6730-4db0-b420-9a45b96a642e", "displayed_text": "Title: Planning from first principles for geometric constraint satisfaction\n\nAbstract: An important problem in geometric reasoning is to find the configuration of a collection of geometric bodies so as to satisfy a set of given constraints. Recently, it has been suggested that this problem can be solved efficiently by symbolically reasoning about geometry using a degrees of freedom analysis. The approach employs a set of specialized routines called plan fragments that specify how to change the configuration of a set of bodies to satisfy a new constraint while preserving existing constraints. In this paper we show how these plan fragments can be automatically synthesized using first principles about geometric bodies, actions, and topology.", "label_annotations": {"Multi-aspect Summary": {"Context": "An important problem in geometric reasoning is to find the configuration of a collection of geometric bodies so as to satisfy a set of given constraints. Recently, it has been suggested that this problem can be solved efficiently by symbolically reasoning about geometry using a degrees of freedom analysis.", "Key idea": "Authors use some ways to show that how plan fragments, i.e. a set of specialized routines, can be automatically synthesized using first principles about geometric bodies, actions, and topology.", "Method": "N/A", "Outcome": "Authors show that how plan fragments can be automatically synthesized using first principles about geometric bodies, actions, and topology.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "6dc39f88-d613-4ec0-b70d-d5daa6f3643c", "displayed_text": "Title: Demonstration of BitGourmet: Data Analysis via Deterministic Approximation\n\nAbstract: We demonstrate BitGourmet, a novel data analysis system that supports deterministic approximate query processing (DAQ). The system executes aggregation queries and produces deterministic bounds that are guaranteed to contain the true value. The system allows users to set a precision constraint on query results. Given a user-defined target precision, we operate on a carefully selected data subset to satisfy the precision constraint. More precisely, we divide each column vertically, bit-by-bit. Our specialized query processing engine evaluates queries on subsets of these bit vectors. This involves a scenario-specific query optimizer which relies on quality and cost models to decide the optimal bit selection and execution plan. In our demonstration, we show that DAQ realizes an interesting trade-off between result quality and execution time, making data analysis more interactive. We also offer manual control over the query plan, i.e., the bit selection and the execution plan, so that users can gain more insights into our system and DAQ in general.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a novel data analysis system that supports deterministic approximate query processing (DAQ) named BitGourmet.", "Method": "The authors give a demonstration to show BitGourmet performance on deterministic approximate query processing task.", "Outcome": "BitGourmet realizes an interesting trade-off between result quality and execution time, making data analysis more interactive and it also allows users manual control over the query plan for deeper system insights.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 12s "}} +{"id": "6dc39f88-d613-4ec0-b70d-d5daa6f3643c", "displayed_text": "Title: Demonstration of BitGourmet: Data Analysis via Deterministic Approximation\n\nAbstract: We demonstrate BitGourmet, a novel data analysis system that supports deterministic approximate query processing (DAQ). The system executes aggregation queries and produces deterministic bounds that are guaranteed to contain the true value. The system allows users to set a precision constraint on query results. Given a user-defined target precision, we operate on a carefully selected data subset to satisfy the precision constraint. More precisely, we divide each column vertically, bit-by-bit. Our specialized query processing engine evaluates queries on subsets of these bit vectors. This involves a scenario-specific query optimizer which relies on quality and cost models to decide the optimal bit selection and execution plan. In our demonstration, we show that DAQ realizes an interesting trade-off between result quality and execution time, making data analysis more interactive. We also offer manual control over the query plan, i.e., the bit selection and the execution plan, so that users can gain more insights into our system and DAQ in general.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose BitGourmet, a novel data analysis system that supports deterministic approximate query processing (DAQ).", "Method": "N/A", "Outcome": "The authors show that DAQ realizes an interesting trade-off between result quality and execution time, making data analysis more interactive in their demonstration.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 34s "}} +{"id": "6de74297-fb80-448f-b7ae-41f8d9701044", "displayed_text": "Title: Can We Consider Central Catadioptric Cameras and Fisheye Cameras within a Unified Imaging Model\n\nAbstract: There are two kinds of omnidirectional cameras often used in computer vision: central catadioptric cameras and fisheye cameras. Previous literatures use different imaging models to describe them separately. A unified imaging model is however presented in this paper. The unified model in this paper can be considered as an extension of the unified imaging model for central catadioptric cameras proposed by Geyer and Daniilidis. We show that our unified model can cover some existing models for fisheye cameras and fit well for many actual fisheye cameras used in previous literatures. Under our unified model, central catadioptric cameras and fisheye cameras can be classified by the model\u2019s characteristic parameter, and a fisheye image can be transformed into a central catadioptric one, vice versa. An important merit of our new unified model is that existing calibration methods for central catadioptric cameras can be directly applied to fisheye cameras. Furthermore, the metric calibration from single fisheye image only using projections of lines becomes possible via our unified model but the existing methods for fisheye cameras in the literatures till now are all non-metric under the same conditions. Experimental results of calibration from some central catadioptric and fisheye images confirm the validity and usefulness of our new unified model.", "label_annotations": {"Multi-aspect Summary": {"Context": "Since there are two kinds of omnidirectional cameras often used in computer vision: central catadioptric cameras and fisheye cameras, previous literatures use different imaging models to describe them separately. ", "Key idea": "The authors propose a unified imaging model, which can cover some existing models for fisheye cameras and fit well for many actual fisheye cameras used in previous literatures.", "Method": "The authors compare unified imaging model with other. existing methods for fisheye cameras in the literatures on the metric calibration from single fisheye image and evaluate unified imaging model on alibration from some central catadioptric and fisheye images.", "Outcome": "While the existing methods for fisheye cameras in the literatures are all non-metric, unified imaging model only use projections of lines on the metric calibration from single fisheye image and experimental results of calibration from some central catadioptric and fisheye images confirm the validity and usefulness of our new unified model.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 22m 39s "}} +{"id": "6de74297-fb80-448f-b7ae-41f8d9701044", "displayed_text": "Title: Can We Consider Central Catadioptric Cameras and Fisheye Cameras within a Unified Imaging Model\n\nAbstract: There are two kinds of omnidirectional cameras often used in computer vision: central catadioptric cameras and fisheye cameras. Previous literatures use different imaging models to describe them separately. A unified imaging model is however presented in this paper. The unified model in this paper can be considered as an extension of the unified imaging model for central catadioptric cameras proposed by Geyer and Daniilidis. We show that our unified model can cover some existing models for fisheye cameras and fit well for many actual fisheye cameras used in previous literatures. Under our unified model, central catadioptric cameras and fisheye cameras can be classified by the model\u2019s characteristic parameter, and a fisheye image can be transformed into a central catadioptric one, vice versa. An important merit of our new unified model is that existing calibration methods for central catadioptric cameras can be directly applied to fisheye cameras. Furthermore, the metric calibration from single fisheye image only using projections of lines becomes possible via our unified model but the existing methods for fisheye cameras in the literatures till now are all non-metric under the same conditions. Experimental results of calibration from some central catadioptric and fisheye images confirm the validity and usefulness of our new unified model.", "label_annotations": {"Multi-aspect Summary": {"Context": "There are two kinds of omnidirectional cameras often used in computer vision: central catadioptric cameras and fisheye cameras. Previous literatures use different imaging models to describe them separately.", "Key idea": "The authors present a unified imaging model, which can be considered as an extension of the unified imaging model for central catadioptric cameras proposed by Geyer and Daniilidis.", "Method": "The authors evaluate their unified model by calibrating some central catadioptric and fisheye images.", "Outcome": "Experimental results of calibration from some central catadioptric and fisheye images confirm the validity and usefulness of the new unified model.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 36s "}} +{"id": "6ecf725b-661e-4897-8169-22d71826d0e8", "displayed_text": "Title: OIE@OIA: an Adaptable and Efficient Open Information Extraction Framework\n\nAbstract: Different Open Information Extraction (OIE) tasks require different types of information, so the OIE field requires strong adaptability of OIE algorithms to meet different task requirements. This paper discusses the adaptability problem in existing OIE systems and designs a new adaptable and efficient OIE system - OIE@OIA as a solution. OIE@OIA follows the methodology of Open Information eXpression (OIX): parsing a sentence to an Open Information Annotation (OIA) Graph and then adapting the OIA graph to different OIE tasks with simple rules. As the core of our OIE@OIA system, we implement an end-to-end OIA generator by annotating a dataset (we make it open available) and designing an efficient learning algorithm for the complex OIA graph. We easily adapt the OIE@OIA system to accomplish three popular OIE tasks. The experimental show that our OIE@OIA achieves new SOTA performances on these tasks, showing the great adaptability of our OIE@OIA system. Furthermore, compared to other end-to-end OIE baselines that need millions of samples for training, our OIE@OIA needs much fewer training samples (12K), showing a significant advantage in terms of efficiency.", "label_annotations": {"Multi-aspect Summary": {"Context": "Different Open Information Extraction (OIE) tasks require different types of information, so the OIE field requires strong adaptability of OIE algorithms to meet different task requirements.", "Key idea": "The authors designs a new adaptable and efficient OIE system - OIE@OIA as a solution to the adaptability problem in existing OIE systems.", "Method": "The authors conduct experiments to evaluate OIE@OIA system on three popular OIE tasks and compare it with other end-to-end OIE baselines.", "Outcome": "On three popular OIE tasks, OIE@OIA achieves new SOTA performances, showing the great adaptability. And compared to other end-to-end OIE baselines that need millions of samples for training, OIE@OIA needs much fewer training samples (12K), showing a significant advantage in terms of efficiency.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 7m 13s "}} +{"id": "6ecf725b-661e-4897-8169-22d71826d0e8", "displayed_text": "Title: OIE@OIA: an Adaptable and Efficient Open Information Extraction Framework\n\nAbstract: Different Open Information Extraction (OIE) tasks require different types of information, so the OIE field requires strong adaptability of OIE algorithms to meet different task requirements. This paper discusses the adaptability problem in existing OIE systems and designs a new adaptable and efficient OIE system - OIE@OIA as a solution. OIE@OIA follows the methodology of Open Information eXpression (OIX): parsing a sentence to an Open Information Annotation (OIA) Graph and then adapting the OIA graph to different OIE tasks with simple rules. As the core of our OIE@OIA system, we implement an end-to-end OIA generator by annotating a dataset (we make it open available) and designing an efficient learning algorithm for the complex OIA graph. We easily adapt the OIE@OIA system to accomplish three popular OIE tasks. The experimental show that our OIE@OIA achieves new SOTA performances on these tasks, showing the great adaptability of our OIE@OIA system. Furthermore, compared to other end-to-end OIE baselines that need millions of samples for training, our OIE@OIA needs much fewer training samples (12K), showing a significant advantage in terms of efficiency.", "label_annotations": {"Multi-aspect Summary": {"Context": "Different Open Information Extraction (OIE) tasks require different types of information, so the OIE field requires strong adaptability of OIE algorithms to meet different task requirements.", "Key idea": "The authors discuss the adaptability problem in existing OIE systems and design a new adaptable and efficient OIE system - OIE@OIA as a solution.", "Method": "The authors use the proposed OIE@OIA system to accomplish three popular OIE tasks, and compare it to other end-to-end OIE baselines.", "Outcome": "The experimental show that the proposed OIE@OIA achieves new SOTA performances on three popular OIE tasks, showing the great adaptability of OIE@OIA system. Furthermore, compared to other end-to-end OIE baselines that need millions of samples for training, OIE@OIA needs much fewer training samples (12K), showing a significant advantage in terms of efficiency.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 52s "}} +{"id": "73e353a8-e0d6-466f-af93-6fccf38fcb18", "displayed_text": "Title: Learning Reasoning Paths over Semantic Graphs for Video-grounded Dialogues\n\nAbstract: Compared to traditional visual question answering, video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting. Previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modelling the inherent information flows at the turn level. We propose to discover information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer. We then introduce a new approach that learns to predict reasoning paths over this semantic graph. Our path prediction model predicts a path from the current turn through past dialogue turns that contain additional visual cues to answer the current question. Our reasoning model sequentially processes both visual and textual information through this reasoning path and the propagated features are used to generate the answer. Our experimental results demonstrate the effectiveness of our method and provide additional insights on how models use semantic dependencies in a dialogue context to retrieve visual cues.", "label_annotations": {"Multi-aspect Summary": {"Context": "Compared to traditional visual question answering, video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting. Previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modelling the inherent information flows at the turn level.", "Key idea": "The authors propose to introduce a new approach that learns to predict reasoning paths over this semantic graph for video-grounded dialogues.", "Method": "N/A", "Outcome": "The experimental results demonstrate the effectiveness of t he proposed method.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 38s "}} +{"id": "73e353a8-e0d6-466f-af93-6fccf38fcb18", "displayed_text": "Title: Learning Reasoning Paths over Semantic Graphs for Video-grounded Dialogues\n\nAbstract: Compared to traditional visual question answering, video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting. Previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modelling the inherent information flows at the turn level. We propose to discover information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer. We then introduce a new approach that learns to predict reasoning paths over this semantic graph. Our path prediction model predicts a path from the current turn through past dialogue turns that contain additional visual cues to answer the current question. Our reasoning model sequentially processes both visual and textual information through this reasoning path and the propagated features are used to generate the answer. Our experimental results demonstrate the effectiveness of our method and provide additional insights on how models use semantic dependencies in a dialogue context to retrieve visual cues.", "label_annotations": {"Multi-aspect Summary": {"Context": "Video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting, but previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modeling the inherent information flows at the turn level.", "Key idea": "The author introduces an approach that a path prediction model predicts a path from the current turn through past dialogue turns that contain additional visual cues to answer the current question, and than a reasoning model processes both visual and textual information through this reasoning path and the propagated features are used to generate the answer.", "Method": "N/A", "Outcome": "The proposed method is demonstrated as effective, and insights on how models use semantic dependencies in a dialogue context to retrieve visual cues is are provided.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 30s "}} +{"id": "754f792b-fcf3-42f4-bb2c-5c1dcce21d7a", "displayed_text": "Title: Anaphora and logical form: on formal meaning representations for natural language\n\nAbstract: We argue, on general grounds, in favor of formal meaning representations for natural language. We then adopt, as a Forcing function\" for the adequacy of such as representation the problem of identifying the possible antecedents of anachoric expressions. This suggests certain structural properties of a representation which facilitate the identification of a possible antecedents. Given an appropriate representation. language with such properties, it is then possible to deal with a surprisingly rich class of anachora.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors adopt, as a Forcing function\" for the adequacy of such as representation the problem of identifying the possible antecedents of anachoric expressions.", "Method": "N/A", "Outcome": "Certain structural properties of a representation which facilitate the identification of a possible antecedents and language with such properties, it is then possible to deal with a surprisingly rich class of anachora.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 25m 59s "}} +{"id": "754f792b-fcf3-42f4-bb2c-5c1dcce21d7a", "displayed_text": "Title: Anaphora and logical form: on formal meaning representations for natural language\n\nAbstract: We argue, on general grounds, in favor of formal meaning representations for natural language. We then adopt, as a Forcing function\" for the adequacy of such as representation the problem of identifying the possible antecedents of anachoric expressions. This suggests certain structural properties of a representation which facilitate the identification of a possible antecedents. Given an appropriate representation. language with such properties, it is then possible to deal with a surprisingly rich class of anachora.", "label_annotations": {"Multi-aspect Summary": {"Context": "People argue, on general grounds, in favor of formal meaning representations for natural language, and then adopt, as a \"Forcing function\" for the adequacy of such as representation the problem of identifying the possible antecedents of anachoric expressions.", "Key idea": "The authors investigate certain structural properties of a representation which facilitate the identification of a possible antecedents.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 41s "}} +{"id": "7b0671d7-c8f6-4e81-828b-c73958a6a63a", "displayed_text": "Title: Education, entertainment and authenticity: lessons learned from designing an interactive exhibit about medieval music\n\nAbstract: In this paper we describe the design experience gathered from creating an interactive exhibit about medieval music. This system was designed as an educational exhibit that relies on audio as its only feedback channel. We focused our work on three major goals: educational value, entertainment aspects, and historic authenticity. We present insight into the challenges in designing a system with these goals, and how they could be solved.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors describe the design experience gathered from creating an interactive exhibit about medieval music.", "Method": "There are three primary goals, educational value, entertainment aspects, and historic authenticity, that design experience focus on.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "7b0671d7-c8f6-4e81-828b-c73958a6a63a", "displayed_text": "Title: Education, entertainment and authenticity: lessons learned from designing an interactive exhibit about medieval music\n\nAbstract: In this paper we describe the design experience gathered from creating an interactive exhibit about medieval music. This system was designed as an educational exhibit that relies on audio as its only feedback channel. We focused our work on three major goals: educational value, entertainment aspects, and historic authenticity. We present insight into the challenges in designing a system with these goals, and how they could be solved.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors describe the design experience of an interactive exhibit about medieval music, which relies on audio as its only feedback channel and serves an educational purpose.", "Method": "N/A", "Outcome": "The authors reveal the challenges in designing a system with the goals of educational value, entertainment aspects, and historic authenticity, and how they could be solved.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 37s "}} +{"id": "7b21425c-a2b7-4d19-b030-a8350b2a7a80", "displayed_text": "Title: Conditional set generation using Seq2seq models\n\nAbstract: Conditional set generation learns a mapping from an input sequence of tokens to a set. Several NLP tasks, such as entity typing and dialogue emotion tagging, are instances of set generation. Seq2Seq models, a popular choice for set generation, treat a set as a sequence and do not fully leverage its key properties, namely order-invariance and cardinality. We propose a novel algorithm for effectively sampling informative orders over the combinatorial space of label orders. We jointly model the set cardinality and output by prepending the set size and taking advantage of the autoregressive factorization used by Seq2Seq models. Our method is a model-independent data augmentation approach that endows any Seq2Seq model with the signals of order-invariance and cardinality. Training a Seq2Seq model on this augmented data (without any additional annotations) gets an average relative improvement of 20% on four benchmark datasets across various models: BART, T5, and GPT-3. Code to use SETAUG available at: https://setgen.structgen.com. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Conditional set generation learns a mapping from an input sequence of tokens to a set. Several NLP tasks, such as entity typing and dialogue emotion tagging, are instances of set generation. Seq2Seq models, a popular choice for set generation, treat a set as a sequence and do not fully leverage its key properties, namely order-invariance and cardinality.", "Key idea": "The authors propose a novel algorithm for effectively sampling informative orders over the combinatorial space of label orders, which is a model-independent data augmentation approach that endows any Seq2Seq model with the signals of order-invariance and cardinality.", "Method": "The authors train several Seq2Seq models based their method and evaluate these models on four benchmark datasets.", "Outcome": "Training a Seq2Seq model on this augmented data (without any additional annotations) gets an average relative improvement of 20% on four benchmark datasets across various models: BART, T5, and GPT-3.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 54s "}} +{"id": "7b21425c-a2b7-4d19-b030-a8350b2a7a80", "displayed_text": "Title: Conditional set generation using Seq2seq models\n\nAbstract: Conditional set generation learns a mapping from an input sequence of tokens to a set. Several NLP tasks, such as entity typing and dialogue emotion tagging, are instances of set generation. Seq2Seq models, a popular choice for set generation, treat a set as a sequence and do not fully leverage its key properties, namely order-invariance and cardinality. We propose a novel algorithm for effectively sampling informative orders over the combinatorial space of label orders. We jointly model the set cardinality and output by prepending the set size and taking advantage of the autoregressive factorization used by Seq2Seq models. Our method is a model-independent data augmentation approach that endows any Seq2Seq model with the signals of order-invariance and cardinality. Training a Seq2Seq model on this augmented data (without any additional annotations) gets an average relative improvement of 20% on four benchmark datasets across various models: BART, T5, and GPT-3. Code to use SETAUG available at: https://setgen.structgen.com. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Seq2Seq models, a popular choice for set generation, treat a set as a sequence and do not fully leverage its key properties, namely order-invariance and cardinality.", "Key idea": "The authors propose a novel algorithm for effectively sampling informative orders over the combinatorial space of label orders. This method is a model-independent data augmentation approach that endows any Seq2Seq model with the signals of order-invariance and cardinality. ", "Method": "The authors train a Seq2Seq model on new augmented data (without any additional annotations) to evaluate performance across various models: BART, T5, and GPT-3.", "Outcome": "Training a Seq2Seq model on augmented data (without any additional annotations) gets an average relative improvement of 20% on four benchmark datasets across various models: BART, T5, and GPT-3. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 13s "}} +{"id": "7c065e41-7c2e-430e-a918-6a776037bf50", "displayed_text": "Title: Report on the DB/IR panel at SIGMOD 2005\n\nAbstract: This paper summarizes the salient aspects of the SIGMOD 2005 panel on Databases and Information Retrieval: Rethinking the Great . The goal of the panel was to discuss whether we should rethink data management systems architectures to truly merge Database (DB) and Information Retrieval (IR) technologies. The panel had very high attendance and generated lively discussions.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "N/A", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 18s "}} +{"id": "7c065e41-7c2e-430e-a918-6a776037bf50", "displayed_text": "Title: Report on the DB/IR panel at SIGMOD 2005\n\nAbstract: This paper summarizes the salient aspects of the SIGMOD 2005 panel on Databases and Information Retrieval: Rethinking the Great . The goal of the panel was to discuss whether we should rethink data management systems architectures to truly merge Database (DB) and Information Retrieval (IR) technologies. The panel had very high attendance and generated lively discussions.", "label_annotations": {"Multi-aspect Summary": {"Context": "The SIGMOD 2005 panel on Databases and Information Retrieval: Rethinking the Great aims ad discussing whether we should rethink data management systems architectures to truly merge Database (DB) and Information Retrieval (IR) technologies. The panel had very high attendance and generated lively discussions.", "Key idea": "The authors summarize the salient aspects of the SIGMOD 2005 panel on Databases and Information Retrieval: Rethinking the Great.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 28s "}} +{"id": "7d850699-212d-466b-976f-0afed0653fef", "displayed_text": "Title: Informing determiner and preposition error correction with word clusters\n\nAbstract: We extend our n-gram-based data-driven prediction approach from the Helping Our Own (HOO) 2011 Shared Task (Boyd and Meurers, 2011) to identify determiner and preposition errors in non-native English essays from the Cambridge Learner Corpus FCE Dataset (Yannakoudakis et al., 2011) as part of the HOO 2012 Shared Task. Our system focuses on three error categories: missing determiner, incorrect determiner, and incorrect preposition. Approximately two-thirds of the errors annotated in HOO 2012 training and test data fall into these three categories. To improve our approach, we developed a missing determiner detector and incorporated word clustering (Brown et al., 1992) into the n-gram prediction approach.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors extend their n-gram-based data-driven prediction approach from the Helping Our Own (HOO) 2011 Shared Task (Boyd and Meurers, 2011) to identify determiner and preposition errors in non-native English essays from the Cambridge Learner Corpus FCE Dataset (Yannakoudakis et al., 2011) as part of the HOO 2012 Shared Task.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 29s "}} +{"id": "7d850699-212d-466b-976f-0afed0653fef", "displayed_text": "Title: Informing determiner and preposition error correction with word clusters\n\nAbstract: We extend our n-gram-based data-driven prediction approach from the Helping Our Own (HOO) 2011 Shared Task (Boyd and Meurers, 2011) to identify determiner and preposition errors in non-native English essays from the Cambridge Learner Corpus FCE Dataset (Yannakoudakis et al., 2011) as part of the HOO 2012 Shared Task. Our system focuses on three error categories: missing determiner, incorrect determiner, and incorrect preposition. Approximately two-thirds of the errors annotated in HOO 2012 training and test data fall into these three categories. To improve our approach, we developed a missing determiner detector and incorporated word clustering (Brown et al., 1992) into the n-gram prediction approach.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors extend their n-gram-based data-driven prediction approach. They develop a missing determiner detector and incorporating word clustering into the n-gram prediction approach.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 12m 25s "}} +{"id": "802a5b78-a022-4d38-bfb3-f28eee4ef89a", "displayed_text": "Title: Transition-based Dependency Parsing Using Two Heterogeneous Gated Recursive Neural Networks\n\nAbstract: Recently, neural network based dependency parsing has attracted much interest, which can effectively alleviate the problems of data sparsity and feature engineering by using the dense features. However, it is still a challenge problem to sufficiently model the complicated syntactic and semantic compositions of the dense features in neural network based methods. In this paper, we propose two heterogeneous gated recursive neural networks: tree structured gated recursive neural network (Tree-GRNN) and directed acyclic graph structured gated recursive neural network (DAG-GRNN). Then we integrate them to automatically learn the compositions of the dense features for transition-based dependency parsing. Specifically, Tree-GRNN models the feature combinations for the trees in stack, which already have partial dependency structures. DAG-GRNN models the feature combinations of the nodes whose dependency relations have not been built yet. Experiment results on two prevalent benchmark datasets (PTB3 and CTB5) show the effectiveness of our proposed model.", "label_annotations": {"Multi-aspect Summary": {"Context": "It is still a challenge problem to sufficiently model the complicated syntactic and semantic compositions of the dense features in neural network based methods.", "Key idea": "The authors propose two heterogeneous gated recursive neural networks: tree structured gated recursive neural network (Tree-GRNN) and directed acyclic graph structured gated recursive neural network (DAG-GRNN) and integrate them to automatically learn the compositions of the dense features for transition-based dependency parsing.", "Method": "The authors applied tree structured gated recursive neural network (Tree-GRNN) and directed acyclic graph structured gated recursive neural network (DAG-GRNN) on two prevalent benchmark datasets (PTB3 and CTB5) to check their effectiveness.", "Outcome": "Experiments conducted on two benchmark datasets, PTB3 and CTB5, demonstrate the effectiveness of the proposed models, tree structured gated recursive neural network (Tree-GRNN) and directed acyclic graph structured gated recursive neural network (DAG-GRNN).", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 14s "}} +{"id": "802a5b78-a022-4d38-bfb3-f28eee4ef89a", "displayed_text": "Title: Transition-based Dependency Parsing Using Two Heterogeneous Gated Recursive Neural Networks\n\nAbstract: Recently, neural network based dependency parsing has attracted much interest, which can effectively alleviate the problems of data sparsity and feature engineering by using the dense features. However, it is still a challenge problem to sufficiently model the complicated syntactic and semantic compositions of the dense features in neural network based methods. In this paper, we propose two heterogeneous gated recursive neural networks: tree structured gated recursive neural network (Tree-GRNN) and directed acyclic graph structured gated recursive neural network (DAG-GRNN). Then we integrate them to automatically learn the compositions of the dense features for transition-based dependency parsing. Specifically, Tree-GRNN models the feature combinations for the trees in stack, which already have partial dependency structures. DAG-GRNN models the feature combinations of the nodes whose dependency relations have not been built yet. Experiment results on two prevalent benchmark datasets (PTB3 and CTB5) show the effectiveness of our proposed model.", "label_annotations": {"Multi-aspect Summary": {"Context": "Recently, neural network based dependency parsing has attracted much interest, which can effectively alleviate the problems of data sparsity and feature engineering by using the dense features. However, it is still a challenge problem to sufficiently model the complicated syntactic and semantic compositions of the dense features in neural network based methods.", "Key idea": "The authors propose two heterogeneous gated recursive neural networks: tree structured gated recursive neural network (Tree-GRNN) and directed acyclic graph structured gated recursive neural network (DAG-GRNN), and the integration of both networks.", "Method": "The authors evaluated their proposed networks on two prevalent benchmark datasets (PTB3 and CTB5).", "Outcome": "Experiment results on two prevalent benchmark datasets (PTB3 and CTB5) show the effectiveness of our proposed model.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 59s "}} +{"id": "80698baf-89cb-4a50-9f5c-0c74151b798b", "displayed_text": "Title: Grading the Graders: Motivating Peer Graders in a MOOC\n\nAbstract: In this paper, we detail our efforts at creating and running a controlled study designed to examine how students in a MOOC might be motivated to do a better job during peer grading. This study involves more than one thousand students of a popular MOOC. We ask two specific questions: (1) When a student knows that his or her own peer grading efforts are being examined by peers, does this knowledge alone tend to motivate the student to do a better job when grading assignments? And (2) when a student not only knows that his or her own peer grading efforts are being examined by peers, but he or she is also given a number of other peer grading efforts to evaluate (so the peer graders see how other peer graders evaluate assignments), do both of these together tend to motivate the student to do a better job when grading assignments? We find strong statistical evidence that ``grading the graders'' does in fact tend to increase the quality of peer grading.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author details their efforts at creating and running a controlled study designed to examine how students in a MOOC might be motivated to do a better job during peer grading on two scenarios: (1) When a student knows that his or her own peer grading efforts are being examined by peers, and (2) when a student not only knows that his or her own peer grading efforts are being examined by peers, but is also given a number of other peer grading efforts to evaluate.", "Method": "The author collects samples from more than one thousand students of a popular MOOC and perform statistical analysis on whether \"grading the graders\" increases the quality of peer grading.", "Outcome": "The author finds strong statistical evidence that \"grading the graders\" does in fact tend to increase the quality of peer grading.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 5h 52m 32s "}} +{"id": "80698baf-89cb-4a50-9f5c-0c74151b798b", "displayed_text": "Title: Grading the Graders: Motivating Peer Graders in a MOOC\n\nAbstract: In this paper, we detail our efforts at creating and running a controlled study designed to examine how students in a MOOC might be motivated to do a better job during peer grading. This study involves more than one thousand students of a popular MOOC. We ask two specific questions: (1) When a student knows that his or her own peer grading efforts are being examined by peers, does this knowledge alone tend to motivate the student to do a better job when grading assignments? And (2) when a student not only knows that his or her own peer grading efforts are being examined by peers, but he or she is also given a number of other peer grading efforts to evaluate (so the peer graders see how other peer graders evaluate assignments), do both of these together tend to motivate the student to do a better job when grading assignments? We find strong statistical evidence that ``grading the graders'' does in fact tend to increase the quality of peer grading.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors create and run a controlled study designed to examine how students in a MOOC might be motivated to do a better job during peer grading.", "Method": "The authors gather more than one thousand students of a popular MOOC, and ask them two specific questions regarding their graders.", "Outcome": "The authors find strong statistical evidence that \"grading the graders\" does in fact tend to increase the quality of peer grading.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 42s "}} +{"id": "816707fd-9214-4435-ac40-b2655e55c9d0", "displayed_text": "Title: CoDA21: Evaluating Language Understanding Capabilities of NLP Models With Context-Definition Alignment\n\nAbstract: Pretrained language models (PLMs) have achieved superhuman performance on many benchmarks, creating a need for harder tasks. We introduce CoDA21 (Context Definition Alignment), a challenging benchmark that measures natural language understanding (NLU) capabilities of PLMs: Given a definition and a context each for k words, but not the words themselves, the task is to align the k definitions with the k contexts. CoDA21 requires a deep understanding of contexts and definitions, including complex inference and world knowledge. We find that there is a large gap between human and PLM performance, suggesting that CoDA21 measures an aspect of NLU that is not sufficiently covered in existing benchmarks.(1)", "label_annotations": {"Multi-aspect Summary": {"Context": "Pretrained language models (PLMs) have achieved superhuman performance on many benchmarks, creating a need for harder tasks. ", "Key idea": "The authors introduce CoDA21 (Context Definition Alignment), a challenging benchmark that measures natural language understanding (NLU) capabilities of PLMs.", "Method": "The authors use CoDA21 as benchmark to compare human and PLM's natural language understanding (NLU) capabilities.", "Outcome": "Human has much better performance on natural language understanding(NLU) capabilities than Pretrained language models (PLM).\r\nCoDA21 measures an aspect of NLU that is not sufficiently covered in existing benchmarks.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 11s "}} +{"id": "816707fd-9214-4435-ac40-b2655e55c9d0", "displayed_text": "Title: CoDA21: Evaluating Language Understanding Capabilities of NLP Models With Context-Definition Alignment\n\nAbstract: Pretrained language models (PLMs) have achieved superhuman performance on many benchmarks, creating a need for harder tasks. We introduce CoDA21 (Context Definition Alignment), a challenging benchmark that measures natural language understanding (NLU) capabilities of PLMs: Given a definition and a context each for k words, but not the words themselves, the task is to align the k definitions with the k contexts. CoDA21 requires a deep understanding of contexts and definitions, including complex inference and world knowledge. We find that there is a large gap between human and PLM performance, suggesting that CoDA21 measures an aspect of NLU that is not sufficiently covered in existing benchmarks.(1)", "label_annotations": {"Multi-aspect Summary": {"Context": "Pretrained language models (PLMs) have achieved superhuman performance on many benchmarks, creating a need for harder tasks.", "Key idea": "We introduce CoDA21 (Context Definition Alignment), a challenging benchmark that measures natural language understanding (NLU) capabilities of PLMs: Given a definition and a context each for k words, but not the words themselves, the task is to align the k definitions with the k contexts.", "Method": "The author compares the performance of human and pretrained language models on the proposed benchmark.", "Outcome": "There is a large gap between human and PLM performance, suggesting that CoDA21 measures an aspect of NLU that is not sufficiently covered in existing benchmarks.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 24s "}} +{"id": "8737b031-f77c-4f32-8a68-4be9b0c9ecf8", "displayed_text": "Title: Effective Use of Synthetic Data for Urban Scene Semantic Segmentation\n\nAbstract: Training a deep network to perform semantic segmentation requires large amounts of labeled data. To alleviate the manual effort of annotating real images, researchers have investigated the use of synthetic data, which can be labeled automatically. Unfortunately, a network trained on synthetic data performs relatively poorly on real images. While this can be addressed by domain adaptation, existing methods all require having access to real images during training. In this paper, we introduce a drastically different way to handle synthetic images that does not require seeing any real images at training time. Our approach builds on the observation that foreground and background classes are not affected in the same manner by the domain shift, and thus should be treated differently. In particular, the former should be handled in a detection-based manner to better account for the fact that, while their texture in synthetic images is not photo-realistic, their shape looks natural. Our experiments evidence the effectiveness of our approach on Cityscapes and CamVid with models trained on synthetic data only.", "label_annotations": {"Multi-aspect Summary": {"Context": "A network trained on synthetic data performs relatively poorly on real images. Although this can be addressed by domain adaptation, existing methods all require having access to real images during training.", "Key idea": "The authors introduce a drastically different way to handle synthetic images that does not require seeing any real images at training time.", "Method": "The authors evaluate new method on Cityscapes and CamVid with models trained on synthetic data only.", "Outcome": "Experiments results prove the effectiveness of new method on Cityscapes and CamVid with models trained on synthetic data only.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 23s "}} +{"id": "8737b031-f77c-4f32-8a68-4be9b0c9ecf8", "displayed_text": "Title: Effective Use of Synthetic Data for Urban Scene Semantic Segmentation\n\nAbstract: Training a deep network to perform semantic segmentation requires large amounts of labeled data. To alleviate the manual effort of annotating real images, researchers have investigated the use of synthetic data, which can be labeled automatically. Unfortunately, a network trained on synthetic data performs relatively poorly on real images. While this can be addressed by domain adaptation, existing methods all require having access to real images during training. In this paper, we introduce a drastically different way to handle synthetic images that does not require seeing any real images at training time. Our approach builds on the observation that foreground and background classes are not affected in the same manner by the domain shift, and thus should be treated differently. In particular, the former should be handled in a detection-based manner to better account for the fact that, while their texture in synthetic images is not photo-realistic, their shape looks natural. Our experiments evidence the effectiveness of our approach on Cityscapes and CamVid with models trained on synthetic data only.", "label_annotations": {"Multi-aspect Summary": {"Context": "Training a deep network to perform semantic segmentation requires large amounts of labeled data. The use of synthetic data to alleviate the manual effort of annotating will result in poorer performance, and existing methods to resolve this all require having access to real images during training.", "Key idea": "The authors introduce a way to handle synthetic images that does not require seeing any real images at training time.", "Method": "The authors train their models using the proposed approach on Cityscapes and CamVid, using synthetic data only.", "Outcome": "The experiments evidence the effectiveness of the proposed approach on Cityscapes and CamVid with models trained on synthetic data only.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 9m 23s "}} +{"id": "8ae36735-e4ac-48da-bd2b-5538a6a00a73", "displayed_text": "Title: Overcoming Relational Learning Biases to Accurately Predict Preferences in Large Scale Networks\n\nAbstract: Many individuals on social networking sites provide traits about themselves, such as interests or demographics. Social networking sites can use this information to provide better content to match their usersu0027 interests, such as recommending scheduled events or various relevant products. These tasks require accurate probability estimates to determine the correct answer to return. Relational machine learning (RML) is an excellent framework for these problems as it jointly models the user labels given their attributes and the relational structure. Further, semi-supervised learning methods could enable RML methods to exploit the large amount of unlabeled data in networks. However, existing RML approaches have limitations that prevent their application in large scale domains. First, semi-supervised methods for RML do not fully utilize all the unlabeled instances in the network. Second, the collective inference procedures necessary to jointly infer the missing labels are generally viewed as too expensive to apply in large scale domains. In this work, we address each of these limitations. We analyze the effect of full semi-supervised RML and find that collective inference methods can introduce considerable bias into predictions. We correct this by implementing a maximum entropy constraint on the inference step, forcing the predictions to have the same distribution as the observed labels. Next, we outline a massively scalable variational inference algorithm for large scale relational network domains. We extend this inference algorithm to incorporate the maximum entropy constraint, proving that it only requires a constant amount of overhead while remaining massively parallel. We demonstrate our methodu0027s improvement over a variety of baselines on seven real world datasets, including large scale networks with over five million edges.", "label_annotations": {"Multi-aspect Summary": {"Context": "Relational machine learning (RML) is an excellent framework for problems that require accurate probability estimates to determine the correct answer as it jointly models the user labels given their attributes and the relational structure. However, existing RML approaches have limitations that prevent their application in large scale domains.", "Key idea": "The authors propose a method, implementing a maximum entropy constraint on the inference step, can solve the limitations of Relational machine learning (RML).", "Method": "The authors conduct experiments over a variety of baselines on seven real world datasets to evaluate new method.", "Outcome": "Experiments prove new method's improvement over a variety of baselines on seven real world datasets, including large scale networks with over five million edges.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 28s "}} +{"id": "8ae36735-e4ac-48da-bd2b-5538a6a00a73", "displayed_text": "Title: Overcoming Relational Learning Biases to Accurately Predict Preferences in Large Scale Networks\n\nAbstract: Many individuals on social networking sites provide traits about themselves, such as interests or demographics. Social networking sites can use this information to provide better content to match their usersu0027 interests, such as recommending scheduled events or various relevant products. These tasks require accurate probability estimates to determine the correct answer to return. Relational machine learning (RML) is an excellent framework for these problems as it jointly models the user labels given their attributes and the relational structure. Further, semi-supervised learning methods could enable RML methods to exploit the large amount of unlabeled data in networks. However, existing RML approaches have limitations that prevent their application in large scale domains. First, semi-supervised methods for RML do not fully utilize all the unlabeled instances in the network. Second, the collective inference procedures necessary to jointly infer the missing labels are generally viewed as too expensive to apply in large scale domains. In this work, we address each of these limitations. We analyze the effect of full semi-supervised RML and find that collective inference methods can introduce considerable bias into predictions. We correct this by implementing a maximum entropy constraint on the inference step, forcing the predictions to have the same distribution as the observed labels. Next, we outline a massively scalable variational inference algorithm for large scale relational network domains. We extend this inference algorithm to incorporate the maximum entropy constraint, proving that it only requires a constant amount of overhead while remaining massively parallel. We demonstrate our methodu0027s improvement over a variety of baselines on seven real world datasets, including large scale networks with over five million edges.", "label_annotations": {"Multi-aspect Summary": {"Context": "Relational machine learning (RML), combined with semi-supervised learning methods, is an excellent framework to provide better contents to uses based on the traits they posted on social networking sites. However, existing RML approaches have several limitations that prevent their application in large scale domains.", "Key idea": "The authors find the limitation of the collective inference methods of full semi-supervised RML and implement a maximum entropy constraint on the inference step to correct this.", "Method": "The authors evaluate their proposed method on seven real world datasets, including large scale networks with over five million edges, and compare it to a variety of baselines.", "Outcome": "The authors demonstrate their method's improvement over a variety of baselines on seven real world datasets, including large scale networks with over five million edges.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 9m 53s "}} +{"id": "8aedb046-2f51-4229-bc19-ea6db98355cb", "displayed_text": "Title: AggregateRank: bringing order to web sites\n\nAbstract: Since the website is one of the most important organizational structures of the Web, how to effectively rank websites has been essential to many Web applications, such as Web search and crawling. In order to get the ranks of websites, researchers used to describe the inter-connectivity among websites with a so-called HostGraph in which the nodes denote websites and the edges denote linkages between websites (if and only if there are hyperlinks from the pages in one website to the pages in the other, there will be an edge between these two websites), and then adopted the random walk model in the HostGraph. However, as pointed in this paper, the random walk over such a HostGraph is not reasonable because it is not in accordance with the browsing behavior of web surfers. Therefore, the derivate rank cannot represent the true probability of visiting the corresponding website.In this work, we mathematically proved that the probability of visiting a website by the random web surfer should be equal to the sum of the PageRank values of the pages inside that website. Nevertheless, since the number of web pages is much larger than that of websites, it is not feasible to base the calculation of the ranks of websites on the calculation of PageRank. To tackle this problem, we proposed a novel method named AggregateRank rooted in the theory of stochastic complement, which cannot only approximate the sum of PageRank accurately, but also have a lower computational complexity than PageRank. Both theoretical analysis and experimental evaluation show that AggregateRank is a better method for ranking websites than previous methods.", "label_annotations": {"Multi-aspect Summary": {"Context": " In order to get the ranks of websites, researchers used to describe the inter-connectivity among websites with a so-called HostGraph in which the nodes denote websites and the edges denote linkages between websites, and then adopted the random walk model in the HostGraph. However, the random walk over such a HostGraph is not reasonable because it is not in accordance with the browsing behavior of web surfers. Therefore, the derivate rank cannot represent the true probability of visiting the corresponding website.", "Key idea": "Authors mathematically prove that the probability of visiting a website by the random web surfer should be equal to the sum of the PageRank values of the pages inside that website. Authors also propose a novel method named AggregateRank, which can not only approximate the sum of PageRank accurately, but also have a lower computational complexity than PageRank. ", "Method": "Authors discuss the propose methods in theory and experiments.", "Outcome": "Both theoretical analysis and experimental evaluation show that AggregateRank is a better method for ranking websites than previous methods.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 22s "}} +{"id": "8aedb046-2f51-4229-bc19-ea6db98355cb", "displayed_text": "Title: AggregateRank: bringing order to web sites\n\nAbstract: Since the website is one of the most important organizational structures of the Web, how to effectively rank websites has been essential to many Web applications, such as Web search and crawling. In order to get the ranks of websites, researchers used to describe the inter-connectivity among websites with a so-called HostGraph in which the nodes denote websites and the edges denote linkages between websites (if and only if there are hyperlinks from the pages in one website to the pages in the other, there will be an edge between these two websites), and then adopted the random walk model in the HostGraph. However, as pointed in this paper, the random walk over such a HostGraph is not reasonable because it is not in accordance with the browsing behavior of web surfers. Therefore, the derivate rank cannot represent the true probability of visiting the corresponding website.In this work, we mathematically proved that the probability of visiting a website by the random web surfer should be equal to the sum of the PageRank values of the pages inside that website. Nevertheless, since the number of web pages is much larger than that of websites, it is not feasible to base the calculation of the ranks of websites on the calculation of PageRank. To tackle this problem, we proposed a novel method named AggregateRank rooted in the theory of stochastic complement, which cannot only approximate the sum of PageRank accurately, but also have a lower computational complexity than PageRank. Both theoretical analysis and experimental evaluation show that AggregateRank is a better method for ranking websites than previous methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "How to effectively rank websites is essential to many Web applications, but the previous research, the random walk over HostGraph is not reasonable because it is not in accordance with the browsing behavior of web surfers. Therefore, the derivate rank cannot represent the true probability of visiting the corresponding website.", "Key idea": "The authors mathematically prove that the probability of visiting a website by the random web surfer should be equal to the sum of the PageRank values of the pages inside that website. They further propose a novel method named AggregateRank rooted in the theory of stochastic complement, which cannot only approximate the sum of PageRank accurately, but also have a lower computational complexity than PageRank.", "Method": "N/A", "Outcome": "Both theoretical analysis and experimental evaluation show that the proposed AggregateRank is a better method for ranking websites than previous methods.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "927df1bd-273a-4088-8c56-2e79cac37072", "displayed_text": "Title: Scalable k-means++\n\nAbstract: Over half a century old and showing no signs of aging, k-means remains one of the most popular data processing algorithms. As is well-known, a proper initialization of k-means is crucial for obtaining a good final solution. The recently proposed k-means++ initialization algorithm achieves this, obtaining an initial set of centers that is provably close to the optimum solution. A major downside of the k-means++ is its inherent sequential nature, which limits its applicability to massive data: one must make k passes over the data to find a good initial set of centers. In this work we show how to drastically reduce the number of passes needed to obtain, in parallel, a good initialization. This is unlike prevailing efforts on parallelizing k-means that have mostly focused on the post-initialization phases of k-means. We prove that our proposed initialization algorithm k-means|| obtains a nearly optimal solution after a logarithmic number of passes, and then show that in practice a constant number of passes suffices. Experimental evaluation on real-world large-scale data demonstrates that k-means|| outperforms k-means++ in both sequential and parallel settings.", "label_annotations": {"Multi-aspect Summary": {"Context": "k-means remains one of the most popular data processing algorithms. The recently proposed k-means++ initialization algorithm achieves a proper initialization, obtaining an initial set of centers that is provably close to the optimum solution. A major downside of the k-means++ is its inherent sequential nature, which limits its applicability to massive data: one must make k passes over the data to find a good initial set of centers.", "Key idea": "Authors show how to drastically reduce the number of passes needed to obtain, in parallel, a good initialization. \r\nAuthors prove that our proposed initialization algorithm k-means|| obtains a nearly optimal solution after a logarithmic number of passes, and then show that in practice a constant number of passes suffices.", "Method": "Authors evaluate k-means|| and k-means++ (as a baseline) on real-world large-scale data.", "Outcome": "Experimental evaluation on real-world large-scale data demonstrates that k-means|| outperforms k-means++ in both sequential and parallel settings.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "927df1bd-273a-4088-8c56-2e79cac37072", "displayed_text": "Title: Scalable k-means++\n\nAbstract: Over half a century old and showing no signs of aging, k-means remains one of the most popular data processing algorithms. As is well-known, a proper initialization of k-means is crucial for obtaining a good final solution. The recently proposed k-means++ initialization algorithm achieves this, obtaining an initial set of centers that is provably close to the optimum solution. A major downside of the k-means++ is its inherent sequential nature, which limits its applicability to massive data: one must make k passes over the data to find a good initial set of centers. In this work we show how to drastically reduce the number of passes needed to obtain, in parallel, a good initialization. This is unlike prevailing efforts on parallelizing k-means that have mostly focused on the post-initialization phases of k-means. We prove that our proposed initialization algorithm k-means|| obtains a nearly optimal solution after a logarithmic number of passes, and then show that in practice a constant number of passes suffices. Experimental evaluation on real-world large-scale data demonstrates that k-means|| outperforms k-means++ in both sequential and parallel settings.", "label_annotations": {"Multi-aspect Summary": {"Context": "K-means remains one of the most popular data processing algorithms. K-means++ initialization algorithm achieves this, obtaining an initial set of centers that is provably close to the optimum solution. A major downside of the k-means++ is its inherent sequential nature, which limits its applicability to massive data.", "Key idea": "This paper introduces an initialization algorithm k-means||, designed to reduce the number of sequential passes required by k-means++ while maintaining nearly optimal initialization quality. ", "Method": "The authors try to use experiment to prove that k-means|| algorithm obtains a nearly optimal solution after a logarithmic number of passes, and a constant number of passes suffices in practice.", "Outcome": "Experimental evaluation on real-world large-scale data demonstrates that k-means|| outperforms k-means++ in both sequential and parallel settings.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "9292bc08-a70f-4adb-a6d7-920728893a39", "displayed_text": "Title: Propositional argumentation and causal reasoning\n\nAbstract: The paper introduces a number of propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics. An assumption-based argumentation framework [Bondarenko et al., 1997] will constitute a special case of this construction. In addition, a stronger argumentation system in a full classical language will be shown to be equivalent to a system of causal reasoning [Giunchiglia et al., 2004]. The implications of this correspondence for the respective nonmonotonic theories of argumentation and causal reasoning are discussed.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "Authors introduces a number of propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics. An assumption-based argumentation framework [Bondarenko et al., 1997] will constitute a special case of authors' construction. In addition, a stronger argumentation system in a full classical language will be shown to be equivalent to a system of causal reasoning", "Method": "Author make theoretical framework to introduce propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics.", "Outcome": "The implications of this correspondence for the respective nonmonotonic theories of argumentation and causal reasoning are discussed.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 9m 43s "}} +{"id": "9292bc08-a70f-4adb-a6d7-920728893a39", "displayed_text": "Title: Propositional argumentation and causal reasoning\n\nAbstract: The paper introduces a number of propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics. An assumption-based argumentation framework [Bondarenko et al., 1997] will constitute a special case of this construction. In addition, a stronger argumentation system in a full classical language will be shown to be equivalent to a system of causal reasoning [Giunchiglia et al., 2004]. The implications of this correspondence for the respective nonmonotonic theories of argumentation and causal reasoning are discussed.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce a number of propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics.", "Method": "The authors show the equivalence of a stronger argumentation system in a full classical language and a system of causal reasoning, and discuss the implications of this correspondence.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 51s "}} +{"id": "93bf12a1-2174-43e3-9cd4-c2b8aeed2f93", "displayed_text": "Title: One-sided unsupervised domain mapping\n\nAbstract: In unsupervised domain mapping, the learner is given two unmatched datasets A and B. The goal is to learn a mapping GAB that translates a sample in A to the analog sample in B. Recent approaches have shown that when learning simultaneously both GAB and the inverse mapping GBA, convincing mappings are obtained. In this work, we present a method of learning GAB without learning GBA. This is done by learning a mapping that maintains the distance between a pair of samples. Moreover, good mappings are obtained, even by maintaining the distance between different parts of the same sample before and after mapping. We present experimental results that the new method not only allows for one sided mapping learning, but also leads to preferable numerical results over the existing circularity-based constraint. Our entire code is made publicly available at https://github.com/sagiebenaim/DistanceGAN.", "label_annotations": {"Multi-aspect Summary": {"Context": "Recent approaches in unsupervised domain mapping has shown that convincing mappings are obtained when learning simultaneously both GAB and the inverse mapping GBA, where learning a mapping GAB means translating a sample in dataset A to the analog sample in dataset B, vise versa.", "Key idea": "The author presents a method of learning GAB without learning GBA by learning a mapping that maintains the distance between a pair of samples.", "Method": "N/A", "Outcome": "Good mappings are obtained, even by maintaining the distance between different parts of the same sample before and after mapping. The proposed method allows for one sided mapping learning and leads to preferable numerical results over the existing circularity-based constraint.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 37s "}} +{"id": "93bf12a1-2174-43e3-9cd4-c2b8aeed2f93", "displayed_text": "Title: One-sided unsupervised domain mapping\n\nAbstract: In unsupervised domain mapping, the learner is given two unmatched datasets A and B. The goal is to learn a mapping GAB that translates a sample in A to the analog sample in B. Recent approaches have shown that when learning simultaneously both GAB and the inverse mapping GBA, convincing mappings are obtained. In this work, we present a method of learning GAB without learning GBA. This is done by learning a mapping that maintains the distance between a pair of samples. Moreover, good mappings are obtained, even by maintaining the distance between different parts of the same sample before and after mapping. We present experimental results that the new method not only allows for one sided mapping learning, but also leads to preferable numerical results over the existing circularity-based constraint. Our entire code is made publicly available at https://github.com/sagiebenaim/DistanceGAN.", "label_annotations": {"Multi-aspect Summary": {"Context": "The goal of unsupervised domain mapping is to learn a mapping GAB that translates a sample in A to the analog sample in B, where A and B are two unmatched datasets. Recent approaches have shown that when learning simultaneously both GAB and the inverse mapping GBA, convincing mappings are obtained.", "Key idea": "The authors present a method of learning the mapping GAB without learning the inverse mapping GBA in unsupervised domain mapping.", "Method": "N/A", "Outcome": "The authors present experimental results that the new method not only allows for one sided mapping learning, but also leads to preferable numerical results over the existing circularity-based constraint.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 38m 12s "}} +{"id": "9748b061-590a-4602-9015-a68e879ffced", "displayed_text": "Title: xPrint: A Modularized Liquid Printer for Smart Materials Deposition\n\nAbstract: To meet the increasing requirements of HCI researchers who are looking into using liquid-based materials (e.g., hydrogels) to create novel interfaces, we present a design strategy for HCI researchers to build and customize a liquid-based smart material printing platform with off-the-shelf or easy-to-machine parts. For the hardware, we suggest a magnetic assembly-based modular design. These modularized parts can be easily and precisely reconfigured with off-the-shelf or easy-to-machine parts that can meet different processing requirements such as mechanical mixing, chemical reaction, light activation, and solution vaporization. In addition, xPrint supports an open-source, highly customizable software design and simulation platform, which is applicable for simulating and facilitating smart material constructions. Furthermore, compared to inkjet or pneumatic syringe-based printing systems, xPrint has a large range of printable materials from synthesized polymers to natural micro-organism-living cells with a printing resolution from 10\u03bcm up to 5mm (droplet size). In this paper, we will introduce the system design in detail and three use cases to demonstrate the material variability and the customizability for users with different demands (e.g., designers, scientific researchers, or artists).", "label_annotations": {"Multi-aspect Summary": {"Context": "HCI researchers who are looking into using liquid-based materials (e.g., hydrogels) to create novel interfaces have the increasing requirements.", "Key idea": "The paper introduces a design strategy named xPrint, that allows HCI researchers to build and customize a smart material printing platform using off-the-shelf or easily machinable parts. \r\n", "Method": "These modularized parts of xPrint can be easily and precisely reconfigured with off-the-shelf or easy-to-machine parts that can meet different processing requirements such as mechanical mixing, chemical reaction, light activation, and solution vaporization. ", "Outcome": "xPrint supports an open-source, highly customizable software design and simulation platform, which is applicable for simulating and facilitating smart material constructions. And xPrint has a large range of printable materials from synthesized polymers to natural micro-organism-living cells with a printing resolution from 10\u03bcm up to 5mm (droplet size).", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "9748b061-590a-4602-9015-a68e879ffced", "displayed_text": "Title: xPrint: A Modularized Liquid Printer for Smart Materials Deposition\n\nAbstract: To meet the increasing requirements of HCI researchers who are looking into using liquid-based materials (e.g., hydrogels) to create novel interfaces, we present a design strategy for HCI researchers to build and customize a liquid-based smart material printing platform with off-the-shelf or easy-to-machine parts. For the hardware, we suggest a magnetic assembly-based modular design. These modularized parts can be easily and precisely reconfigured with off-the-shelf or easy-to-machine parts that can meet different processing requirements such as mechanical mixing, chemical reaction, light activation, and solution vaporization. In addition, xPrint supports an open-source, highly customizable software design and simulation platform, which is applicable for simulating and facilitating smart material constructions. Furthermore, compared to inkjet or pneumatic syringe-based printing systems, xPrint has a large range of printable materials from synthesized polymers to natural micro-organism-living cells with a printing resolution from 10\u03bcm up to 5mm (droplet size). In this paper, we will introduce the system design in detail and three use cases to demonstrate the material variability and the customizability for users with different demands (e.g., designers, scientific researchers, or artists).", "label_annotations": {"Multi-aspect Summary": {"Context": "There have been increasing requirements of HCI researchers who are looking into using liquid-based materials (e.g., hydrogels) to create novel interfaces.", "Key idea": "The author presents a design for HCI researchers to build and customize a liquid-based smart material printing platform. The design consists a hardware of a magnetic assembly-based modular design, modularized to be easily and precisely reconfigured with off-the-shelf or easy-to-machine parts to meet different processing requirements, and it also supports open-source, highly customizable software design and simulation platform.", "Method": "The author compares the design with inkjet or pneumatic syringe-based printing systems on its range of printable material, and applies the design on three use cases to demonstrate the material variability and the customizability for users with different demands.", "Outcome": "Compared to inkjet or pneumatic syringe-based printing systems, the proposed design has a large range of printable materials from synthesized polymers to natural micro-organism-living cells with a printing resolution from 10\u03bcm up to 5mm (droplet size). The design is also demonstrated for material variability and the customizability for users with different demands.", "Future Impact": "The design strategy is beneficial for HCI researchers to build and customize a liquid-based smart material printing platform with off-the-shelf or easy-to-machine parts."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 59s "}} +{"id": "977698a6-56ef-4e57-94c5-b3a68a451a80", "displayed_text": "Title: A Question Answering Approach for Emotion Cause Extraction\n\nAbstract: Emotion cause extraction aims to identify the reasons behind a certain emotion expressed in text. It is a much more difficult task compared to emotion classification. Inspired by recent advances in using deep memory networks for question answering (QA), we propose a new approach which considers emotion cause identification as a reading comprehension task in QA. Inspired by convolutional neural networks, we propose a new mechanism to store relevant context in different memory slots to model context information. Our proposed approach can extract both word level sequence features and lexical features. Performance evaluation shows that our method achieves the state-of-the-art performance on a recently released emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.", "label_annotations": {"Multi-aspect Summary": {"Context": "Emotion cause extraction aims to identify the reasons behind a certain emotion expressed in text, which is a much more difficult task compared to emotion classification. Recent research has advances in using deep memory networks for question answering (QA).", "Key idea": "The authors propose a new approach which considers emotion cause identification as a reading comprehension task in QA. Inspired by convolutional neural networks, they also propose a new mechanism to store relevant context in different memory slots to model context information.", "Method": "The authors evaluate their proposed approach on a recently released emotion cause dataset.", "Outcome": "The proposed approach can extract both word level sequence features and lexical features. Performance evaluation shows that the proposed method achieves the state-of-the-art performance on a recently released emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 11s "}} +{"id": "977698a6-56ef-4e57-94c5-b3a68a451a80", "displayed_text": "Title: A Question Answering Approach for Emotion Cause Extraction\n\nAbstract: Emotion cause extraction aims to identify the reasons behind a certain emotion expressed in text. It is a much more difficult task compared to emotion classification. Inspired by recent advances in using deep memory networks for question answering (QA), we propose a new approach which considers emotion cause identification as a reading comprehension task in QA. Inspired by convolutional neural networks, we propose a new mechanism to store relevant context in different memory slots to model context information. Our proposed approach can extract both word level sequence features and lexical features. Performance evaluation shows that our method achieves the state-of-the-art performance on a recently released emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.", "label_annotations": {"Multi-aspect Summary": {"Context": "Emotion cause extraction is more challenging than emotion classification and aims to identify the reasons behind a certain emotion expressed in text. The recent approach uses deep memory networks for question answering.", "Key idea": "The author proposes a mechanism to emotion cause extraction to store relevant context in different memory slots to model context information. The proposed approach can extract both word level sequence features and lexical features.", "Method": "The author assesses the performance of the proposed approach against a recently released emotion cause database.", "Outcome": "Performance evaluation shows that the author\u2019s method achieves the state-of-the-art performance on a recently released emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 26s "}} +{"id": "9821d5f7-72b0-4841-a54f-d2af4a04ea3a", "displayed_text": "Title: Inverse rendering in SUV space with a linear texture model\n\nAbstract: In this paper, we consider the problem of inverse rendering in the case where surface texture can be approximated by a linear basis. Assuming a dichromatic reflectance model, we show that spherical harmonic illumination coefficients and texture parameters can be estimated in a specular invariant colour subspace by solving a system of bilinear equations. We focus on the case of faces, where both shape and texture can be efficiently described by a linear statistical model. In this context, we are able to fit a 3D morphable model to a single colour image, accounting for both non-Lambertian specular reflectance and complex illumination of the same light source colour. We are able to recover statistical texture model parameters with an accuracy comparable to more computationally expensive analysis-by-synthesis approaches. Moreover, our approach requires only the solution of convex optimisation problems.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors a method, assuming a dichromatic reflectance model and solving spherical harmonic illumination coefficients and texture parameters can be estimated in a specular invariant colour subspace by solving a system of bilinear equations. ", "Method": "The authors apply new method on problems focusing on the case of faces.\r\nThe authors compare new method with more computationally expensive analysis-by-synthesis approaches on statistical texture model parameters recovering problem. \r\n", "Outcome": "When focusing on the case of faces, new method is able to fit a 3D morphable model to a single colour image, accounting for both non-Lambertian specular reflectance and complex illumination of the same light source colour.\r\nNew method is able to recover statistical texture model parameters with an accuracy comparable to more computationally expensive analysis-by-synthesis approaches, while only requires the solution of convex optimisation problems.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 57s "}} +{"id": "9821d5f7-72b0-4841-a54f-d2af4a04ea3a", "displayed_text": "Title: Inverse rendering in SUV space with a linear texture model\n\nAbstract: In this paper, we consider the problem of inverse rendering in the case where surface texture can be approximated by a linear basis. Assuming a dichromatic reflectance model, we show that spherical harmonic illumination coefficients and texture parameters can be estimated in a specular invariant colour subspace by solving a system of bilinear equations. We focus on the case of faces, where both shape and texture can be efficiently described by a linear statistical model. In this context, we are able to fit a 3D morphable model to a single colour image, accounting for both non-Lambertian specular reflectance and complex illumination of the same light source colour. We are able to recover statistical texture model parameters with an accuracy comparable to more computationally expensive analysis-by-synthesis approaches. Moreover, our approach requires only the solution of convex optimisation problems.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors show that for a dichromatic reflectance model, spherical harmonic illumination coefficients and texture parameters can be estimated in a specular invariant colour subspace by solving a system of bilinear equations.", "Method": "The authors recover statistical texture model parameters using their proposed method and compare them with the results of more computationally expensive analysis-by-synthesis approaches.", "Outcome": "The authors are able to recover statistical texture model parameters with an accuracy comparable to more computationally expensive analysis-by-synthesis approaches.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 34m 59s "}} +{"id": "9999e1df-f439-4f62-bd03-337e494e9da3", "displayed_text": "Title: A data type encapsulation scheme utilizing base language operators\n\nAbstract: A data type encapsulation scheme in which the \u201cspace\u201d operations are expressed naturally in terms of the base language operators is described. The scheme results from a conceptual separation of operators and procedure calls in the base language and produces a language of considerable expressive power. The scheme has been implemented and several examples are given.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "Authors propose a data type encapsulation scheme in which the \u201cspace\u201d operations are expressed naturally in terms of the base language operators. The scheme results from a conceptual separation of operators and procedure calls in the base language and produces a language of considerable expressive power. ", "Method": "N/A", "Outcome": "The data type encapsulation scheme has been implemented and several examples are given.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 6s "}} +{"id": "9999e1df-f439-4f62-bd03-337e494e9da3", "displayed_text": "Title: A data type encapsulation scheme utilizing base language operators\n\nAbstract: A data type encapsulation scheme in which the \u201cspace\u201d operations are expressed naturally in terms of the base language operators is described. The scheme results from a conceptual separation of operators and procedure calls in the base language and produces a language of considerable expressive power. The scheme has been implemented and several examples are given.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a data type encapsulation scheme in which the \u201cspace\u201d operations are expressed naturally in terms of the base language operators.", "Method": "N/A", "Outcome": "The scheme has been implemented and several examples are given.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 46s "}} +{"id": "99bef0ca-923c-480d-a89e-9a2f27e02157", "displayed_text": "Title: Abstraction for solving large incomplete-information games\n\nAbstract: Most real-world games and many recreational games are games of incomplete information. Over the last dozen years, abstraction has emerged as a key enabler for solving large incomplete-information games. First, the game is abstracted to generate a smaller, abstract game that is strategically similar to the original game. Second, an approximate equilibrium is computed in the abstract game. Third, the strategy from the abstract game is mapped back to the original game.\r\n\r\nIn this paper, I will review key developments in the field. I present reasons for abstracting games, and point out the issue of abstraction pathology. I then review the practical algorithms for information abstraction and action abstraction. I then cover recent theoretical breakthroughs that beget bounds on the quality of the strategy from the abstract game, when measured in the original game. I then discuss how to reverse map the opponentu0027s action into the abstraction if the opponent makes a move that is not in the abstraction. Finally, I discuss other topics of current and future research.", "label_annotations": {"Multi-aspect Summary": {"Context": "Abstraction has emerged as a key enabler for solving large incomplete-information games. ", "Key idea": "The authors review key developments of abstraction, the practical algorithms for information abstraction and action abstraction and recent theoretical breakthroughs.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "99bef0ca-923c-480d-a89e-9a2f27e02157", "displayed_text": "Title: Abstraction for solving large incomplete-information games\n\nAbstract: Most real-world games and many recreational games are games of incomplete information. Over the last dozen years, abstraction has emerged as a key enabler for solving large incomplete-information games. First, the game is abstracted to generate a smaller, abstract game that is strategically similar to the original game. Second, an approximate equilibrium is computed in the abstract game. Third, the strategy from the abstract game is mapped back to the original game.\r\n\r\nIn this paper, I will review key developments in the field. I present reasons for abstracting games, and point out the issue of abstraction pathology. I then review the practical algorithms for information abstraction and action abstraction. I then cover recent theoretical breakthroughs that beget bounds on the quality of the strategy from the abstract game, when measured in the original game. I then discuss how to reverse map the opponentu0027s action into the abstraction if the opponent makes a move that is not in the abstraction. Finally, I discuss other topics of current and future research.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most real-world games and many recreational games are games of incomplete information.\r\n\r\nAbstraction has emerged as a key enabler for solving large incomplete-information games. First, the game is abstracted to generate a smaller, abstract game that is strategically similar to the original game. Second, an approximate equilibrium is computed in the abstract game. Third, the strategy from the abstract game is mapped back to the original game.", "Key idea": "The author review key developments and present reasons for abstracting games, and point out the issue of abstraction pathology. \r\n\r\nThe author then review the practical algorithms for information abstraction and action abstraction. She then cover recent theoretical breakthroughs that beget bounds on the quality of the strategy from the abstract game, when measured in the original game. \r\n\r\nShe then discuss how to reverse map the opponentu0027s action into the abstraction if the opponent makes a move that is not in the abstraction. Finally, she discuss other topics of current and future research.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 3s "}} +{"id": "9b95b592-1562-4ef9-b0ed-e0655fadc73b", "displayed_text": "Title: Surface Matching with Large Deformations and Arbitrary Topology: A Geodesic Distance Evolution Scheme on a 3-Manifold\n\nAbstract: A general formulation for geodesic distance propagation of surfaces is presented. Starting from a surface lying on a 3-manifold in IR4, we set up a partial differential equation governing the propagation of surfaces at equal geodesic distance (on the 3-manifold) from the given original surface. This propagation scheme generalizes a result of Kimmel et al. [11] and provides a way to compute distance maps on manifolds. Moreover, the propagation equation is generalized to any number of dimensions. Using an eulerian formulation with level-sets, it gives stable numerical algorithms for computing distance maps. This theory is used to present a new method for surface matching which generalizes a curve matching method [5]. Matching paths are obtained as the orbits of the vector field defined as the sum of two distance mapsu0027 gradient values. This surface matching technique applies to the case of large deformation and topological changes.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author proposes a new method for surface matching which generalizes a curve matching method. It is based on a partial differential equation governing the propagation of surfaces at equal geodesic distance from the given original surface, which uses an eulerian formulation with level-sets to give stable numerical algorithms for computing distance maps.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 48s "}} +{"id": "9b95b592-1562-4ef9-b0ed-e0655fadc73b", "displayed_text": "Title: Surface Matching with Large Deformations and Arbitrary Topology: A Geodesic Distance Evolution Scheme on a 3-Manifold\n\nAbstract: A general formulation for geodesic distance propagation of surfaces is presented. Starting from a surface lying on a 3-manifold in IR4, we set up a partial differential equation governing the propagation of surfaces at equal geodesic distance (on the 3-manifold) from the given original surface. This propagation scheme generalizes a result of Kimmel et al. [11] and provides a way to compute distance maps on manifolds. Moreover, the propagation equation is generalized to any number of dimensions. Using an eulerian formulation with level-sets, it gives stable numerical algorithms for computing distance maps. This theory is used to present a new method for surface matching which generalizes a curve matching method [5]. Matching paths are obtained as the orbits of the vector field defined as the sum of two distance mapsu0027 gradient values. This surface matching technique applies to the case of large deformation and topological changes.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a general formulation for geodesic distance propagation of surfaces.\r\n\r\n", "Method": "The authors evaluate propagation scheme on computing distance maps on manifolds", "Outcome": "This propagation scheme provide a way to compute distance maps on manifolds.\r\nThe propagation equation can be generalized to any number of dimensions and gives stable numerical algorithms for computing distance maps.\r\nThis theory presents a new method for surface matching which generalizes a curve matching method and surface matching technique can be applies to the case of large deformation and topological changes.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 15m 30s "}} +{"id": "9cd7e7e1-8893-4db6-8327-48f098187699", "displayed_text": "Title: Identifying Relevant Messages in a Twitter-based Citizen Channel for Natural Disaster Situations\n\nAbstract: During recent years the online social networks (in particular Twitter) have become an important alternative information channel to traditional media during natural disasters, but the amount and diversity of messages poses the challenge of information overload to end users. The goal of our research is to develop an automatic classifier of tweets to feed a mobile application that reduces the difficulties that citizens face to get relevant information during natural disasters. In this paper, we present in detail the process to build a classifier that filters tweets relevant and non-relevant to an earthquake. By using a dataset from the Chilean earthquake of 2010, we first build and validate a ground truth, and then we contribute by presenting in detail the effect of class imbalance and dimensionality reduction over 5 classifiers. We show how the performance of these models is affected by these variables, providing important considerations at the moment of building these systems.", "label_annotations": {"Multi-aspect Summary": {"Context": "During recent years the online social networks (in particular Twitter) have become an important alternative information channel to traditional media during natural disasters, but the amount and diversity of messages poses the challenge of information overload to end users.", "Key idea": "Authors develop an automatic classifier of tweets to feed a mobile application that reduces the difficulties that citizens face to get relevant information during natural disasters. Authors present in detail the process to build a classifier that filters tweets relevant and non-relevant to an earthquake.", "Method": "Using a dataset from the Chilean earthquake of 2010, authors build and validate a ground truth, and then they contribute by presenting in detail the effect of class imbalance and dimensionality reduction over 5 classifiers.", "Outcome": "Authors show that the performance of classifiers is affected by these variables, providing important considerations at the moment of building these systems.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 13s "}} +{"id": "9cd7e7e1-8893-4db6-8327-48f098187699", "displayed_text": "Title: Identifying Relevant Messages in a Twitter-based Citizen Channel for Natural Disaster Situations\n\nAbstract: During recent years the online social networks (in particular Twitter) have become an important alternative information channel to traditional media during natural disasters, but the amount and diversity of messages poses the challenge of information overload to end users. The goal of our research is to develop an automatic classifier of tweets to feed a mobile application that reduces the difficulties that citizens face to get relevant information during natural disasters. In this paper, we present in detail the process to build a classifier that filters tweets relevant and non-relevant to an earthquake. By using a dataset from the Chilean earthquake of 2010, we first build and validate a ground truth, and then we contribute by presenting in detail the effect of class imbalance and dimensionality reduction over 5 classifiers. We show how the performance of these models is affected by these variables, providing important considerations at the moment of building these systems.", "label_annotations": {"Multi-aspect Summary": {"Context": "During recent years the online social networks (in particular Twitter) have become an important alternative information channel to traditional media during natural disasters, but the amount and diversity of messages poses the challenge of information overload to end users.", "Key idea": "The authors present in detail the process to build a classifier that filters tweets relevant and non-relevant to an earthquake.", "Method": "The authors build their models and validate a ground truth on a dataset from the Chilean earthquake of 2010.", "Outcome": "The authors show how the performance of classifiers that filters tweets relevant and non-relevant to an earthquake is affected by class imbalance and dimensionality reduction.", "Future Impact": "The research can provide important considerations at the moment of building classifiers that filters information about natural disasters."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "9f641e23-3886-4ac6-b65e-28db86ee48be", "displayed_text": "Title: GSP: The Cinderella of Mechanism Design\n\nAbstract: Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction. By all theoretical accounts including their own [Varian 14], this was the wrong auction --- the Vickrey-Clarke-Groves (VCG) auction would have been the proper choice --- yet GSP has succeeded spectacularly.\r\n\r\nWe give a deep justification for GSPu0027s success: advertisersu0027 preferences map to a model we call value maximization; they do not maximize profit as the standard theory would believe. For value maximizers, GSP is the truthful auction [Aggarwal 09]. Moreover, this implies an axiomatization of GSP --- it is an auction whose prices are truthful for value maximizers --- that can be applied much more broadly than the simple model for which GSP was originally designed. In particular, applying it to arbitrary single-parameter domains recovers the folklore definition of GSP. Through the lens of value maximization, GSP metamorphosizes into a powerful auction, sound in its principles and elegant in its simplicity.", "label_annotations": {"Multi-aspect Summary": {"Context": " Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction. By all theoretical accounts including their own [Varian 14], this was the wrong auction --- the Vickrey-Clarke-Groves (VCG) auction would have been the proper choice --- yet GSP has succeeded spectacularly.", "Key idea": "Authors justify for GSPu0027s success: advertisersu0027 preferences map to a model we call value maximization; they do not maximize profit as the standard theory would believe. For value maximizers, authors argue that GSP is the truthful auction. Moreover, this implies an axiomatization of GSP that can be applied much more broadly than the simple model for which GSP was originally designed.", "Method": "N/A", "Outcome": "Authors justify GSPu0027s success: advertisers preferences map to a model we call value maximization. For value maximizers, Authors argue that GSP is the truthful auction.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 18s "}} +{"id": "9f641e23-3886-4ac6-b65e-28db86ee48be", "displayed_text": "Title: GSP: The Cinderella of Mechanism Design\n\nAbstract: Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction. By all theoretical accounts including their own [Varian 14], this was the wrong auction --- the Vickrey-Clarke-Groves (VCG) auction would have been the proper choice --- yet GSP has succeeded spectacularly.\r\n\r\nWe give a deep justification for GSPu0027s success: advertisersu0027 preferences map to a model we call value maximization; they do not maximize profit as the standard theory would believe. For value maximizers, GSP is the truthful auction [Aggarwal 09]. Moreover, this implies an axiomatization of GSP --- it is an auction whose prices are truthful for value maximizers --- that can be applied much more broadly than the simple model for which GSP was originally designed. In particular, applying it to arbitrary single-parameter domains recovers the folklore definition of GSP. Through the lens of value maximization, GSP metamorphosizes into a powerful auction, sound in its principles and elegant in its simplicity.", "label_annotations": {"Multi-aspect Summary": {"Context": "Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction, which is theoretically wrong. However, GSP succeeded spectacularly despite the fact that the Vickrey-Clarke-Groves (VCG) auction would have been the proper choice.", "Key idea": "The authors justify for GSP's success.", "Method": "The authors make justification based on advertisers' preferences map to a model called value maximization.", "Outcome": "The authors show that through the lens of value maximization, GSP becomes a powerful auction, sound in its principles and elegant in its simplicity.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 16m 46s "}} +{"id": "a10dafe9-6093-47f0-8429-7b62c46566ea", "displayed_text": "Title: Enabling enterprise mashups over unstructured text feeds with InfoSphere MashupHub and SystemT\n\nAbstract: Enterprise mashup scenarios often involve feeds derived from data created primarily for eye consumption, such as email, news, calendars, blogs, and web feeds. These data sources can test the capabilities of current data mashup products, as the attributes needed to perform join, aggregation, and other operations are often buried within unstructured feed text. Information extraction technology is a key enabler in such scenarios, using annotators to convert unstructured text into structured information that can facilitate mashup operations. Our demo presents the integration of SystemT, an information extraction system from IBM Research, with IBMu0027s InfoSphere MashupHub. We show how to build domain-specific annotators with SystemTu0027s declarative rule language, AQL, and how to use these annotators to combine structured and unstructured information in an enterprise mashup.", "label_annotations": {"Multi-aspect Summary": {"Context": "Enterprise mashup scenarios often involve feeds derived from data created primarily for eye consumption, such as email, news, calendars, blogs, and web feeds, which can test the capabilities of current data mashup products. Information extraction technology is a key enabler in such scenarios, using annotators to convert unstructured text into structured information that can facilitate mashup operations.", "Key idea": "The authors present the integration of SystemT, an information extraction system from IBM Research, with IBM's InfoSphere MashupHub.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 1h 45m 26s "}} +{"id": "a10dafe9-6093-47f0-8429-7b62c46566ea", "displayed_text": "Title: Enabling enterprise mashups over unstructured text feeds with InfoSphere MashupHub and SystemT\n\nAbstract: Enterprise mashup scenarios often involve feeds derived from data created primarily for eye consumption, such as email, news, calendars, blogs, and web feeds. These data sources can test the capabilities of current data mashup products, as the attributes needed to perform join, aggregation, and other operations are often buried within unstructured feed text. Information extraction technology is a key enabler in such scenarios, using annotators to convert unstructured text into structured information that can facilitate mashup operations. Our demo presents the integration of SystemT, an information extraction system from IBM Research, with IBMu0027s InfoSphere MashupHub. We show how to build domain-specific annotators with SystemTu0027s declarative rule language, AQL, and how to use these annotators to combine structured and unstructured information in an enterprise mashup.", "label_annotations": {"Multi-aspect Summary": {"Context": "Enterprise mashup scenarios often involve feeds derived from data created primarily for eye consumption, such as email, news, calendars, blogs, and web feeds. These data sources can test the capabilities of current data mashup products, as the attributes needed to perform join, aggregation, and other operations are often buried within unstructured feed text. Information extraction technology is a key enabler in such scenarios, using annotators to convert unstructured text into structured information that can facilitate mashup operations.", "Key idea": "Authors present the integration of SystemT, an information extraction system from IBM Research, with IBMu0027s InfoSphere MashupHub.", "Method": "Authors show demos on how to build domain-specific annotators with SystemTu0027s declarative rule language, AQL, and how to use these annotators to combine structured and unstructured information in an enterprise mashup.", "Outcome": "Authors show that the proposed SystemT can be used to build domain-specific annotators with declarative rule language, AQL, and these annotators can combine structured and unstructured information in an enterprise mashup.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 12s "}} +{"id": "a1739057-ef00-4b01-9c26-4ab2b5d5708e", "displayed_text": "Title: A framework to support multiple query optimization for complex mining tasks\n\nAbstract: With an increasing use of data mining tools and techniques, we envision that a Knowledge Discovery and Data Mining System (KDDMS) will have to support and optimize for the following scenarios: 1) Sequence of Queries: A user may analyze one or more datasets by issuing a sequence of related complex mining queries, and 2) Multiple Simultaneous Queries: Several users may be analyzing a set of datasets concurrently, and may issue related complex queries.This paper presents a systematic mechanism to optimize for the above cases, targetting the class of mining queries involving frequent pattern mining on one or multiple datasets. We present a system architecture and propose new algorithms for this purpose. We show the design of a knowledgeable cache which can store the past query results from queries on multiple datasets. We present algorithms which enable the use of the results stored in such a cache to further optimize multiple queries.We have implemented and evaluated our system with both real and synthetic datasets. Our experimental results show that our techniques can achieve a speedup of up to a factor of 9, compared with the systems which do not support caching or optimize for multiple queries.", "label_annotations": {"Multi-aspect Summary": {"Context": "The use of data mining tools and techniques is increasing.", "Key idea": "The authors envision that a Knowledge Discovery and Data Mining System (KDDMS) will have to support and optimize for two scenarios. They further propose a systematic mechanism to optimize for these two cases, and a system architecture along with new algorithms for this purpose.", "Method": "The authors implement and evaluate their system with both real and synthetic datasets.", "Outcome": "The experimental results show that our techniques can achieve a speedup of up to a factor of 9, compared with the systems which do not support caching or optimize for multiple queries.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 5s "}} +{"id": "a1739057-ef00-4b01-9c26-4ab2b5d5708e", "displayed_text": "Title: A framework to support multiple query optimization for complex mining tasks\n\nAbstract: With an increasing use of data mining tools and techniques, we envision that a Knowledge Discovery and Data Mining System (KDDMS) will have to support and optimize for the following scenarios: 1) Sequence of Queries: A user may analyze one or more datasets by issuing a sequence of related complex mining queries, and 2) Multiple Simultaneous Queries: Several users may be analyzing a set of datasets concurrently, and may issue related complex queries.This paper presents a systematic mechanism to optimize for the above cases, targetting the class of mining queries involving frequent pattern mining on one or multiple datasets. We present a system architecture and propose new algorithms for this purpose. We show the design of a knowledgeable cache which can store the past query results from queries on multiple datasets. We present algorithms which enable the use of the results stored in such a cache to further optimize multiple queries.We have implemented and evaluated our system with both real and synthetic datasets. Our experimental results show that our techniques can achieve a speedup of up to a factor of 9, compared with the systems which do not support caching or optimize for multiple queries.", "label_annotations": {"Multi-aspect Summary": {"Context": "With an increasing use of data mining tools and techniques, authors envision that a Knowledge Discovery and Data Mining System (KDDMS) will have to support and optimize for the following scenarios: 1) Sequence of Queries: A user may analyze one or more datasets by issuing a sequence of related complex mining queries, and 2) Multiple Simultaneous Queries: Several users may be analyzing a set of datasets concurrently, and may issue related complex queries.", "Key idea": "Authors presents a systematic mechanism to optimize for the above cases, targetting the class of mining queries involving frequent pattern mining on one or multiple datasets. Authors present a system architecture and propose new algorithms for this purpose. Authors show the design of a knowledgeable cache which can store the past query results from queries on multiple datasets. Authors present algorithms which enable the use of the results stored in such a cache to further optimize multiple queries.", "Method": "Authors evaluate the system with both real and synthetic datasets", "Outcome": "Experimental results show that our techniques can achieve a speedup of up to a factor of 9, compared with the systems which do not support caching or optimize for multiple queries.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 48s "}} +{"id": "a5bc5ab8-70fb-4dff-824b-7606228e44a9", "displayed_text": "Title: Adding math to Web pages with EzMath\n\nAbstract: EzMath provides an easy to learn notation for embedding mathematical expressions in Web pages.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce EzMath, which provides an easy to learn notation for embedding mathematical expressions in Web pages.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 33s "}} +{"id": "a5bc5ab8-70fb-4dff-824b-7606228e44a9", "displayed_text": "Title: Adding math to Web pages with EzMath\n\nAbstract: EzMath provides an easy to learn notation for embedding mathematical expressions in Web pages.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose EzMath, which provides an easy to learn notation for embedding mathematical expressions in Web pages.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 32s "}} +{"id": "a722b600-3725-4738-a47b-435aebd63e13", "displayed_text": "Title: A Design Philosophy for Agents in the Smart Home\n\nAbstract: The home is often the most private space in people's lives, and not one in which they expect to be surveilled. However, today's market for smart home devices has quickly evolved to include products that monitor, automate, and present themselves as human. After documenting some of the more unusual emergent problems with contemporary devices, this body of work seeks to develop a design philosophy for intelligent agents in the smart home that can act as an alternative to the ways that these devices are currently built. This is then applied to the design of privacy empowering technologies, representing the first steps from the devices of the present towards a more respectful future.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "The home is often the most private space in people's lives, and not one in which they expect to be surveilled. However, today's market for smart home devices has quickly evolved to include products that monitor, automate, and present themselves as human.", "Key idea": "The authors develop a design philosophy for intelligent agents in the smart home that can act as an alternative to the ways that these devices are currently built.", "Method": "N/A", "Outcome": "The proposed philosophy is applied to the design of privacy empowering technologies.", "Future Impact": "The authors' work represent the first steps from the devices of the present towards a more respectful future"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 27s "}} +{"id": "a722b600-3725-4738-a47b-435aebd63e13", "displayed_text": "Title: A Design Philosophy for Agents in the Smart Home\n\nAbstract: The home is often the most private space in people's lives, and not one in which they expect to be surveilled. However, today's market for smart home devices has quickly evolved to include products that monitor, automate, and present themselves as human. After documenting some of the more unusual emergent problems with contemporary devices, this body of work seeks to develop a design philosophy for intelligent agents in the smart home that can act as an alternative to the ways that these devices are currently built. This is then applied to the design of privacy empowering technologies, representing the first steps from the devices of the present towards a more respectful future.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Today's market for smart home devices has quickly evolved to include products that monitor, automate, and present themselves as human. ", "Key idea": "Authors develop a design philosophy for intelligent agents in the smart home that can act as an alternative to the ways that these devices are currently built.", "Method": "Authors apply the design philosophy to the design of privacy empowering technologies.", "Outcome": "Authors argue that this work marks the first steps from the devices of the present towards a more respectful future.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 15s "}} +{"id": "ad69e31c-2c7f-4db0-916c-3deccaab37fd", "displayed_text": "Title: A Scalable Machine Learning Approach to Go\n\nAbstract: Go is an ancient board game that poses unique opportunities and challenges for AI and machine learning. Here we develop a machine learning approach to Go, and related board games, focusing primarily on the problem of learning a good evaluation function in a scalable way. Scalability is essential at multiple levels, from the library of local tactical patterns, to the integration of patterns across the board, to the size of the board itself. The system we propose is capable of automatically learning the propensity of local patterns from a library of games. Propensity and other local tactical information are fed into a recursive neural network, derived from a Bayesian network architecture. The network integrates local information across the board and produces local outputs that represent local territory ownership probabilities. The aggregation of these probabilities provides an effective strategic evaluation function that is an estimate of the expected area at the end (or at other stages) of the game. Local area targets for training can be derived from datasets of human games. A system trained using only 9 \u00d7 9 amateur game data performs surprisingly well on a test set derived from 19 \u00d7 19 professional game data. Possible directions for further improvements are briefly discussed.", "label_annotations": {"Multi-aspect Summary": {"Context": "Go is an ancient board game that poses unique opportunities and challenges for AI and machine learning.", "Key idea": "The author proposes a system that is capable of automatically learning the propensity of local patterns from a library of games through a recursive neural network. The network integrates local information across the board and produces local outputs that represent local territory ownership probabilities, which then provides an effective strategic evaluation function on the expected area at the end of the game.", "Method": "The author tests the system by training on 9 \u00d7 9 amateur game data and testing on 19 \u00d7 19 professional game data.", "Outcome": "A system trained using only 9 \u00d7 9 amateur game data performs surprisingly well on a test set derived from 19 \u00d7 19 professional game data.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 24s "}} +{"id": "ad69e31c-2c7f-4db0-916c-3deccaab37fd", "displayed_text": "Title: A Scalable Machine Learning Approach to Go\n\nAbstract: Go is an ancient board game that poses unique opportunities and challenges for AI and machine learning. Here we develop a machine learning approach to Go, and related board games, focusing primarily on the problem of learning a good evaluation function in a scalable way. Scalability is essential at multiple levels, from the library of local tactical patterns, to the integration of patterns across the board, to the size of the board itself. The system we propose is capable of automatically learning the propensity of local patterns from a library of games. Propensity and other local tactical information are fed into a recursive neural network, derived from a Bayesian network architecture. The network integrates local information across the board and produces local outputs that represent local territory ownership probabilities. The aggregation of these probabilities provides an effective strategic evaluation function that is an estimate of the expected area at the end (or at other stages) of the game. Local area targets for training can be derived from datasets of human games. A system trained using only 9 \u00d7 9 amateur game data performs surprisingly well on a test set derived from 19 \u00d7 19 professional game data. Possible directions for further improvements are briefly discussed.", "label_annotations": {"Multi-aspect Summary": {"Context": "Go is an ancient board game that poses unique opportunities and challenges for AI and machine learning.", "Key idea": "The authors develop a machine learning approach to Go, and related board games, focusing primarily on the problem of learning a good evaluation function in a scalable way. They propose a system capable of automatically learning the propensity of local patterns from a library of games.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The authors discuss possible directions for further improvements"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 16m 28s "}} +{"id": "adfd8058-64b3-4062-953c-034b732e2fa0", "displayed_text": "Title: Video Compression Scheme Using DEMD Based Texture Synthesis\n\nAbstract: In this paper we present a video coding scheme based on texture synthesis through Directional Empirical Mode Decomposition (DEMD). In this scheme P and B-frames of the video sequence are decomposed and parametrically coded with the help of DEMD algorithm, while I-frames are coded with the help of H.264. All P and B frames are decomposed into Intrinsic Mode Function (IMF) image and its residue. Only the first level IMF image for P and B frames are coded. At decoder stage subsequent IMF images are synthesized with the help of correlation search. Wavelet decomposition is performed over residual image and energy level at the HH band is used as a decision criterion for number of decomposition to be performed for optimum synthesis. The experimental result demonstrates the effectiveness of the algorithm in multi-resolution parametric modeling of image data which can be efficiently coded to achieve significant compression with acceptable quality. This scheme also enables to perform scalable coding of IMF parameters to achieve higher compression with perceptual quality.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author presents a video coding scheme based on texture synthesis through Directional Empirical Mode Decomposition (DEMD). At encoding stage, scheme P and B-frames of the video sequence are decomposed into Intrinsic Mode Function (IMF) image and its residue and parametrically coded on only their first level with the help of DEMD algorithm and I-frames are coded with the help of H.264, while at encoding stage subsequent IMF images are synthesized with the help of correlation search.\r\n\r\nIn this paper we present a video coding scheme based on texture synthesis through Directional Empirical Mode Decomposition (DEMD). In this scheme P and B-frames of the video sequence are decomposed and parametrically coded with the help of DEMD algorithm, while I-frames are coded with the help of H.264. All P and B frames are decomposed into Intrinsic Mode Function (IMF) image and its residue. Only the first level IMF image for P and B frames are coded. At decoder stage subsequent IMF images are synthesized with the help of correlation search. Wavelet decomposition is performed over residual image and energy level at the HH band is used as a decision criterion for number of decomposition to be performed for optimum synthesis.", "Method": "The author tests the effectiveness of the algorithm in multi-resolution parametric modeling of image data. The author also assesses the scalability of coding of IMF parameters.", "Outcome": "The experimental result demonstrates the effectiveness of the algorithm in multi-resolution parametric modeling of image data which can be efficiently coded to achieve significant compression with acceptable quality. This scheme also enables to perform scalable coding of IMF parameters to achieve higher compression with perceptual quality.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 12m 54s "}} +{"id": "adfd8058-64b3-4062-953c-034b732e2fa0", "displayed_text": "Title: Video Compression Scheme Using DEMD Based Texture Synthesis\n\nAbstract: In this paper we present a video coding scheme based on texture synthesis through Directional Empirical Mode Decomposition (DEMD). In this scheme P and B-frames of the video sequence are decomposed and parametrically coded with the help of DEMD algorithm, while I-frames are coded with the help of H.264. All P and B frames are decomposed into Intrinsic Mode Function (IMF) image and its residue. Only the first level IMF image for P and B frames are coded. At decoder stage subsequent IMF images are synthesized with the help of correlation search. Wavelet decomposition is performed over residual image and energy level at the HH band is used as a decision criterion for number of decomposition to be performed for optimum synthesis. The experimental result demonstrates the effectiveness of the algorithm in multi-resolution parametric modeling of image data which can be efficiently coded to achieve significant compression with acceptable quality. This scheme also enables to perform scalable coding of IMF parameters to achieve higher compression with perceptual quality.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a video coding scheme based on texture synthesis through Directional Empirical Mode Decomposition (DEMD).", "Method": "N/A", "Outcome": "The experimental result demonstrates the effectiveness of the algorithm in multi-resolution parametric modeling of image data which can be efficiently coded to achieve significant compression with acceptable quality. This scheme also enables to perform scalable coding of IMF parameters to achieve higher compression with perceptual quality.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 7s "}} +{"id": "b0ad60d7-3a3a-42c2-acbc-fba55e708ba0", "displayed_text": "Title: A powerful and general approach to context exploitation in natural language processing\n\nAbstract: In natural language, the meaning of a lexeme often varies due to the specific surrounding context. Computational approaches to natural language processing can benefit from a reliable, long-range-context-dependent representation of the meaning of each lexeme that appears in a given sentence. We have developed a general new technique that produces a context-dependent u0027meaningu0027 representation for a lexeme in a specific surrounding context. The u0027meaningu0027 of a lexeme in a specific context is represented by a list of semantically replaceable elements the members of which are other lexemes from our experimental lexicon. We have performed experiments with a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases. The resulting lists can be used to compare the u0027meaningu0027 of conceptual units (individual words or frequently-occurring phrases) in different contexts and also can serve as features for machine learning approaches to classify semantic roles and relationships.", "label_annotations": {"Multi-aspect Summary": {"Context": " In natural language, the meaning of a lexeme often varies due to the specific surrounding context. Computational approaches to natural language processing can benefit from a reliable, long-range-context-dependent representation of the meaning of each lexeme that appears in a given sentence. ", "Key idea": "Authors develope a general new technique that produces a context-dependent u0027meaningu0027 representation for a lexeme in a specific surrounding context. \r\n The u0027meaningu0027 of a lexeme in a specific context is represented by a list of semantically replaceable elements the members of which are other lexemes from our experimental lexicon.", "Method": "Authors perform experiments with a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases.", "Outcome": "The lists generated by the propose technique can be used to compare the u0027meaningu0027 of conceptual units (individual words or frequently-occurring phrases) in different contexts and also can serve as features for machine learning approaches to classify semantic roles and relationships.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "b0ad60d7-3a3a-42c2-acbc-fba55e708ba0", "displayed_text": "Title: A powerful and general approach to context exploitation in natural language processing\n\nAbstract: In natural language, the meaning of a lexeme often varies due to the specific surrounding context. Computational approaches to natural language processing can benefit from a reliable, long-range-context-dependent representation of the meaning of each lexeme that appears in a given sentence. We have developed a general new technique that produces a context-dependent u0027meaningu0027 representation for a lexeme in a specific surrounding context. The u0027meaningu0027 of a lexeme in a specific context is represented by a list of semantically replaceable elements the members of which are other lexemes from our experimental lexicon. We have performed experiments with a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases. The resulting lists can be used to compare the u0027meaningu0027 of conceptual units (individual words or frequently-occurring phrases) in different contexts and also can serve as features for machine learning approaches to classify semantic roles and relationships.", "label_annotations": {"Multi-aspect Summary": {"Context": "Computational approaches to natural language processing can benefit from a reliable, long-range-context-dependent representation of the meaning of each lexeme that appears in a given sentence. ", "Key idea": "The authors developed a general new technique that produces a context-dependent 'meaning' representation for a lexeme in a specific surrounding context. ", "Method": "The authors have performed experiments with a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases to evaluate the effectiveness new technique, that produces a context-dependent 'meaning' representation for a lexeme in a specific surrounding context. ", "Outcome": "The resulting lists generated by new technique can be used to compare the 'meaning' of conceptual units (individual words or frequently-occurring phrases) in different contexts and also can serve as features for machine learning approaches to classify semantic roles and relationships.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 49s "}} +{"id": "b14f2bc1-607f-4d8b-a731-b4afdf30a633", "displayed_text": "Title: MetaLogic: Logical Reasoning Explanations with Fine-Grained Structure\n\nAbstract: In this paper, we propose a comprehensive benchmark to investigate models' logical reasoning capabilities in complex real-life scenarios. Current explanation datasets often employ synthetic data with simple reasoning structures. Therefore, it cannot express more complex reasoning processes, such as the rebuttal to a reasoning step and the degree of certainty of the evidence. To this end, we propose a comprehensive logical reasoning explanation form. Based on the multi-hop chain of reasoning, the explanation form includes three main components: (1) The condition of rebuttal that the reasoning node can be challenged; (2) Logical formulae that uncover the internal texture of reasoning nodes; (3) Reasoning strength indicated by degrees of certainty. The fine-grained structure conforms to the real logical reasoning scenario, better fitting the human cognitive process but, simultaneously, is more challenging for the current models. We evaluate the current best models' performance on this new explanation form. The experimental results show that generating reasoning graphs remains a challenging task for current models, even with the help of giant pre-trained language models. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Current explanation datasets often employ synthetic data with simple reasoning structures. Therefore, it cannot express more complex reasoning processes, such as the rebuttal to a reasoning step and the degree of certainty of the evidence.", "Key idea": "The authors propose a comprehensive benchmark to investigate models' logical reasoning capabilities in complex real-life scenarios, as well as a logical reasoning explanation form.", "Method": "The authors evaluate the current best models' performance on their proposed new explanation form.", "Outcome": "The experimental results show that generating reasoning graphs remains a challenging task for current models, even with the help of giant pre-trained language models.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 12h 59m 8s "}} +{"id": "b14f2bc1-607f-4d8b-a731-b4afdf30a633", "displayed_text": "Title: MetaLogic: Logical Reasoning Explanations with Fine-Grained Structure\n\nAbstract: In this paper, we propose a comprehensive benchmark to investigate models' logical reasoning capabilities in complex real-life scenarios. Current explanation datasets often employ synthetic data with simple reasoning structures. Therefore, it cannot express more complex reasoning processes, such as the rebuttal to a reasoning step and the degree of certainty of the evidence. To this end, we propose a comprehensive logical reasoning explanation form. Based on the multi-hop chain of reasoning, the explanation form includes three main components: (1) The condition of rebuttal that the reasoning node can be challenged; (2) Logical formulae that uncover the internal texture of reasoning nodes; (3) Reasoning strength indicated by degrees of certainty. The fine-grained structure conforms to the real logical reasoning scenario, better fitting the human cognitive process but, simultaneously, is more challenging for the current models. We evaluate the current best models' performance on this new explanation form. The experimental results show that generating reasoning graphs remains a challenging task for current models, even with the help of giant pre-trained language models. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Current explanation datasets often employ synthetic data with simple reasoning structures. Therefore, it cannot express more complex reasoning processes, such as the rebuttal to a reasoning step and the degree of certainty of the evidence.", "Key idea": "The authors propose a comprehensive logical reasoning explanation form including rebuttal conditions, logical formulae, and reasoning strength to investigate models' logical reasoning capabilities in complex real-life scenarios. ", "Method": "The authors evaluate the current best models' performance on this new comprehensive logical reasoning explanation form.", "Outcome": "The experimental results show that generating reasoning graphs remains a challenging task for current models, even with the help of giant pre-trained language models.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "b391a193-83e3-4f11-801f-1842647d626e", "displayed_text": "Title: Uncertain Graph Neural Networks For Facial Action Unit Detection\n\nAbstract: Capturing the dependencies among different facial action units (AU) is extremely important for the AU detection task. Many studies have employed graph-based deep learning methods to exploit the dependencies among AUs. However, the dependencies among AUs in real world data are often noisy and the uncertainty is essential to be taken into consideration. Rather than employing a deterministic mode, we propose an uncertain graph neural network (UGN) to learn the probabilistic mask that simultaneously captures both the individual dependencies among AUs and the uncertainties. Further, we propose an adaptive weighted loss function based on the epistemic uncertainties to adaptively vary the weights of the training samples during the training process to account for unbalanced data distributions among AUs. We also provide an insightful analysis on how the uncertainties are related to the performance of AU detection. Extensive experiments, conducted on two benchmark datasets, i.e., BP4D and DISFA, demonstrate our method achieves the state-of-the-art performance.", "label_annotations": {"Multi-aspect Summary": {"Context": "Capturing the dependencies among different facial action units (AU) is extremely important for the AU detection task, but the dependencies among AUs in real world data are often noisy and uncertain. Many studies have employed graph-based deep learning methods to exploit the dependencies among AUs.", "Key idea": "The authors propose an uncertain graph neural network (UGN) to learn the probabilistic mask that simultaneously captures both the individual dependencies among AUs and the uncertainties. They further propose an adaptive weighted loss function based on the epistemic uncertainties to adaptively vary the weights of the training samples during the training process to account for unbalanced data distributions among AUs.", "Method": "They analyze how the uncertainties are related to the performance of AU detection and conduct extensive experiments on two benchmark datasets, i.e., BP4D and DISFA.", "Outcome": "Extensive experiments, conducted on two benchmark datasets, i.e., BP4D and DISFA, demonstrate our method achieves the state-of-the-art performance.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 12s "}} +{"id": "b391a193-83e3-4f11-801f-1842647d626e", "displayed_text": "Title: Uncertain Graph Neural Networks For Facial Action Unit Detection\n\nAbstract: Capturing the dependencies among different facial action units (AU) is extremely important for the AU detection task. Many studies have employed graph-based deep learning methods to exploit the dependencies among AUs. However, the dependencies among AUs in real world data are often noisy and the uncertainty is essential to be taken into consideration. Rather than employing a deterministic mode, we propose an uncertain graph neural network (UGN) to learn the probabilistic mask that simultaneously captures both the individual dependencies among AUs and the uncertainties. Further, we propose an adaptive weighted loss function based on the epistemic uncertainties to adaptively vary the weights of the training samples during the training process to account for unbalanced data distributions among AUs. We also provide an insightful analysis on how the uncertainties are related to the performance of AU detection. Extensive experiments, conducted on two benchmark datasets, i.e., BP4D and DISFA, demonstrate our method achieves the state-of-the-art performance.", "label_annotations": {"Multi-aspect Summary": {"Context": "Capturing the dependencies among different facial action units (AU) is extremely important for the AU detection task.\r\nMany studies have employed graph-based deep learning methods to exploit the dependencies among AUs. However, the dependencies among AUs in real world data are often noisy and the uncertainty is essential to be taken into consideration.", "Key idea": "Authors propose an uncertain graph neural network (UGN) to learn the probabilistic mask that simultaneously captures both the individual dependencies among AUs and the uncertainties. Further, authors propose an adaptive weighted loss function based on the epistemic uncertainties to adaptively vary the weights of the training samples during the training process to account for unbalanced data distributions among AUs. ", "Method": "Authors conduct experiments on two benchmark datasets, i.e., BP4D and DISFA.", "Outcome": "Extensive experiments, conducted on two benchmark datasets, i.e., BP4D and DISFA, demonstrate the proposed method achieves the state-of-the-art performance.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 39s "}} +{"id": "b6b29c8a-7c8c-444c-b434-2ff9e166d9aa", "displayed_text": "Title: Pessimistic Bootstrapping for Uncertainty-Driven Offline Reinforcement Learning\n\nAbstract: Offline Reinforcement Learning (RL) aims to learn policies from previously collected datasets without exploring the environment. Directly applying off-policy algorithms to offline RL usually fails due to the extrapolation error caused by the out-of-distribution (OOD) actions. Previous methods tackle such problem by penalizing the Q-values of OOD actions or constraining the trained policy to be close to the behavior policy. Nevertheless, such methods typically prevent the generalization of value functions beyond the offline data and also lack precise characterization of OOD data. In this paper, we propose Pessimistic Bootstrapping for offline RL (PBRL), a purely uncertainty-driven offline algorithm without explicit policy constraints. Specifically, PBRL conducts uncertainty quantification via the disagreement of bootstrapped Q-functions, and performs pessimistic updates by penalizing the value function based on the estimated uncertainty. To tackle the extrapolating error, we further propose a novel OOD sampling method. We show that such OOD sampling and pessimistic bootstrapping yields provable uncertainty quantifier in linear MDPs, thus providing the theoretical underpinning for PBRL. Extensive experiments on D4RL benchmark show that PBRL has better performance compared to the state-of-the-art algorithms. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Directly applying off-policy algorithms to offline RL usually fails due to the extrapolation error caused by the out-of-distribution (OOD) actions. Previous methods tackle such problem by penalizing the Q-values of OOD actions or constraining the trained policy to be close to the behavior policy, which typically prevent the generalization of value functions beyond the offline data and also lack precise characterization of OOD data.", "Key idea": "The authors propose Pessimistic Bootstrapping for offline RL (PBRL), a purely uncertainty-driven offline algorithm without explicit policy constraints, and further propose a novel OOD sampling method to tackle the extrapolating error.", "Method": "The authors conduct extensive experiments to evaluate their proposed method on D4RL benchmark.", "Outcome": "The authors show that their proposed OOD sampling and pessimistic bootstrapping yields provable uncertainty quantifier in linear MDPs. Extensive experiments on D4RL benchmark further show that the proposed PBRL has better performance compared to the state-of-the-art algorithms.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 25s "}} +{"id": "b6b29c8a-7c8c-444c-b434-2ff9e166d9aa", "displayed_text": "Title: Pessimistic Bootstrapping for Uncertainty-Driven Offline Reinforcement Learning\n\nAbstract: Offline Reinforcement Learning (RL) aims to learn policies from previously collected datasets without exploring the environment. Directly applying off-policy algorithms to offline RL usually fails due to the extrapolation error caused by the out-of-distribution (OOD) actions. Previous methods tackle such problem by penalizing the Q-values of OOD actions or constraining the trained policy to be close to the behavior policy. Nevertheless, such methods typically prevent the generalization of value functions beyond the offline data and also lack precise characterization of OOD data. In this paper, we propose Pessimistic Bootstrapping for offline RL (PBRL), a purely uncertainty-driven offline algorithm without explicit policy constraints. Specifically, PBRL conducts uncertainty quantification via the disagreement of bootstrapped Q-functions, and performs pessimistic updates by penalizing the value function based on the estimated uncertainty. To tackle the extrapolating error, we further propose a novel OOD sampling method. We show that such OOD sampling and pessimistic bootstrapping yields provable uncertainty quantifier in linear MDPs, thus providing the theoretical underpinning for PBRL. Extensive experiments on D4RL benchmark show that PBRL has better performance compared to the state-of-the-art algorithms. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous methods tackle problem that off-policy algorithms failures due to the extrapolation error caused by the out-of-distribution (OOD) actions by penalizing the Q-values of OOD actions or constraining the trained policy to be close to the behavior policy. And these methods typically prevent the generalization of value functions beyond the offline data and also lack precise characterization of OOD data. ", "Key idea": "The authors propose Pessimistic Bootstrapping for offline RL (PBRL), a purely uncertainty-driven offline algorithm without explicit policy constraints and further propose a novel OOD sampling method to tackle the extrapolating error.", "Method": "The authors evaluate Pessimistic Bootstrapping for offline RL (PBRL) performance on D4RL benchmark.", "Outcome": "The authors prove OOD sampling and pessimistic bootstrapping yields provable uncertainty quantifier in linear MDPs, thus providing the theoretical underpinning for PBRL. \r\nExtensive experiments on D4RL benchmark show that PBRL has better performance compared to the state-of-the-art algorithms.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 40m 53s "}} +{"id": "bd84a1fd-ee85-4ac9-a6ae-e534b6013506", "displayed_text": "Title: Harmonic Decompositions of Convolutional Networks\n\nAbstract: We consider convolutional networks from a reproducing kernel Hilbert space viewpoint. We establish harmonic decompositions of convolutional networks, that is expansions into sums of elementary functions of increasing order. The elementary functions are related to the spherical harmonics, a fundamental class of special functions on spheres. The harmonic decompositions allow us to characterize the integral operators associated with convolutional networks, and obtain as a result statistical bounds for convolutional networks. ", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors consider convolutional networks from a reproducing kernel Hilbert space viewpoint and establish harmonic decompositions of convolutional networks, that is expansions into sums of elementary functions of increasing order. ", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 49s "}} +{"id": "bd84a1fd-ee85-4ac9-a6ae-e534b6013506", "displayed_text": "Title: Harmonic Decompositions of Convolutional Networks\n\nAbstract: We consider convolutional networks from a reproducing kernel Hilbert space viewpoint. We establish harmonic decompositions of convolutional networks, that is expansions into sums of elementary functions of increasing order. The elementary functions are related to the spherical harmonics, a fundamental class of special functions on spheres. The harmonic decompositions allow us to characterize the integral operators associated with convolutional networks, and obtain as a result statistical bounds for convolutional networks. ", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors establish harmonic decompositions of convolutional networks, which is expansions into sums of elementary functions of increasing order.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The harmonic decompositions allow characterizing the integral operators associated with convolutional networks, and obtaining as a result statistical bounds for convolutional networks."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 23s "}} +{"id": "c268a190-6974-4190-8f48-db5dcbda8bc8", "displayed_text": "Title: Worst Case Optimal Joins on Relational and XML data\n\nAbstract: In recent data management ecosystem, one of the greatest challenges is the data variety. Data varies in multiple formats such as relational and (semi-)structured data. Traditional database handles a single type of data format and thus its ability to deal with different types of data formats is limited. To overcome such limitation, we propose a multi-model processing framework for relational and semi-structured data (i.e. XML), and design a worst-case optimal join algorithm. The salient feature of our algorithm is that it can guarantee that the intermediate results are no larger than the worst-case join results. Preliminary results show that our multi-model algorithm significantly outperforms the baseline join methods in terms of running time and intermediate result size.", "label_annotations": {"Multi-aspect Summary": {"Context": "Data varies in multiple formats such as relational and (semi-)structured data. Traditional database handles a single type of data format and thus its ability to deal with different types of data formats is limited.", "Key idea": "The author proposes a multi-model processing framework for relational and semi-structured data, and designs a worst-case optimal join algorithm. The salient feature of the algorithm is that it can guarantee that the intermediate results are no larger than the worst-case join results.", "Method": "The author assesses the proposed algorithm against the baseline join model on unning time and intermediate result size.", "Outcome": "The multi-model algorithm significantly outperforms the baseline join methods in terms of running time and intermediate result size.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 27s "}} +{"id": "c268a190-6974-4190-8f48-db5dcbda8bc8", "displayed_text": "Title: Worst Case Optimal Joins on Relational and XML data\n\nAbstract: In recent data management ecosystem, one of the greatest challenges is the data variety. Data varies in multiple formats such as relational and (semi-)structured data. Traditional database handles a single type of data format and thus its ability to deal with different types of data formats is limited. To overcome such limitation, we propose a multi-model processing framework for relational and semi-structured data (i.e. XML), and design a worst-case optimal join algorithm. The salient feature of our algorithm is that it can guarantee that the intermediate results are no larger than the worst-case join results. Preliminary results show that our multi-model algorithm significantly outperforms the baseline join methods in terms of running time and intermediate result size.", "label_annotations": {"Multi-aspect Summary": {"Context": "In recent data management ecosystem, one of the greatest challenges is the data variety. Data varies in multiple formats such as relational and (semi-)structured data. Traditional database handles a single type of data format and thus its ability to deal with different types of data formats is limited.", "Key idea": "The authors propose a multi-model processing framework for relational and semi-structured data (i.e. XML), and design a worst-case optimal join algorithm.", "Method": "The authors compare their algorithm with baseline join methods", "Outcome": "Preliminary results show that the proposed multi-model algorithm significantly outperforms the baseline join methods in terms of running time and intermediate result size.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 40s "}} +{"id": "c37e0f9f-1654-4f1f-b812-bf6f67c0c840", "displayed_text": "Title: Attack Agnostic Statistical Method for Adversarial Detection\n\nAbstract: Deep Learning based AI systems have shown great promise in various domains such as vision, audio, autonomous systems (vehicles, drones), etc. Recent research on neural networks has shown the susceptibility of deep networks to adversarial attacks - a technique of adding small perturbations to the inputs which can fool a deep network into misclassifying them. Developing defenses against such adversarial attacks is an active research area, with some approaches proposing robust models that are immune to such adversaries, while other techniques attempt to detect such adversarial inputs. In this paper, we present a novel statistical approach for adversarial detection in image classification. Our approach is based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class. For this purpose, we make use of various statistical distances such as ED (Energy Distance), MMD (Maximum Mean Discrepancy) for adversarial detection, and analyze the performance of each metric. We experimentally show that our approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size and the degree of adversarial perturbation.", "label_annotations": {"Multi-aspect Summary": {"Context": "Recent research on neural networks has shown the susceptibility of deep networks to adversarial attacks by adding small perturbations to the inputs which can fool a deep network into misclassifying them. Current developed defenses against such adversarial attacks are using robust models that are immune to such adversaries and detecting such adversarial inputs.", "Key idea": "The author present a statistical approach for adversarial detection in image classification, which is based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class. The author uses various statistical distances such as ED and MMD for adversarial detection and analyzes the performance of each metric.", "Method": "The author uses MNIST and CIFAR-10 datasets, and various attack methods, sample sizes and degrees of adversarial perturbation to test the effectiveness of the proposal.", "Outcome": "The proposed method achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size and the degree of adversarial perturbation.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 49s "}} +{"id": "c37e0f9f-1654-4f1f-b812-bf6f67c0c840", "displayed_text": "Title: Attack Agnostic Statistical Method for Adversarial Detection\n\nAbstract: Deep Learning based AI systems have shown great promise in various domains such as vision, audio, autonomous systems (vehicles, drones), etc. Recent research on neural networks has shown the susceptibility of deep networks to adversarial attacks - a technique of adding small perturbations to the inputs which can fool a deep network into misclassifying them. Developing defenses against such adversarial attacks is an active research area, with some approaches proposing robust models that are immune to such adversaries, while other techniques attempt to detect such adversarial inputs. In this paper, we present a novel statistical approach for adversarial detection in image classification. Our approach is based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class. For this purpose, we make use of various statistical distances such as ED (Energy Distance), MMD (Maximum Mean Discrepancy) for adversarial detection, and analyze the performance of each metric. We experimentally show that our approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size and the degree of adversarial perturbation.", "label_annotations": {"Multi-aspect Summary": {"Context": "Susceptibility of deep networks to adversarial attacks - a technique of adding small perturbations to the inputs which can fool a deep network into misclassifying them.", "Key idea": "The authors present a novel statistical approach, based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class, for adversarial detection in image classification. ", "Method": "The authors make use of various statistical distances such as ED (Energy Distance), MMD (Maximum Mean Discrepancy) for adversarial detection, and analyze the performance of each metric to evaluate new statistical approach's adversarial detection performance on MNIST and CIFAR-10 datasets.", "Outcome": "New statistical approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size and the degree of adversarial perturbation.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "c3aa9543-0695-42b4-99af-d033e3912801", "displayed_text": "Title: Data Science for the Real Estate Industry\n\nAbstract: World's major industries, such as Financial Services, Telecom, Advertising, Healthcare, Education, etc, have attracted the attention of the KDD community for decades. Hundreds of KDD papers have been published on topics related to these industries and dozens of workshops organized---some of which have become an integral part of the conference agenda (e.g. the Health Day). Somewhat unexpectedly, the KDD conference has barely addressed the real estate industry, despite its enormous size and prominence. The reason for that apparent mismatch is two-fold: (a) until recently, the real estate industry did not appreciate the value data science methods could add (with some exceptions, such as econometrics methods for creating real-estate price indices); (b) the Data Science community has not been aware of challenging real estate problems that are perfectly suited to its methods. This tutorial provides a step towards resolving this issue. We provide an introduction to real estate for data scientists, and outline a spectrum of data science problems, many of which are being tackled by new \"prop-tech\" companies, while some are yet to be approached. We present concrete examples from three of these companies (where the authors work): Airbnb -- the most popular short-term rental marketplace, Cherre -- a real estate data integration platform, and Compass -- the largest independent real estate brokerage in the U.S.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Hundreds of KDD papers have been published on topics related to these industries and dozens of workshops organized. However, KDD conference has barely addressed the real estate industry, despite its enormous size and prominence. ", "Key idea": "Authors discuss the reasons why KDD has never discuss real estate industry, (a) until recently, the real estate industry did not appreciate the value data science methods could add (with some exceptions, such as econometrics methods for creating real-estate price indices); (b) the Data Science community has not been aware of challenging real estate problems that are perfectly suited to its methods.\r\n\r\nThis tutorial provides a step towards resolving this issue.", "Method": "Authors provide an introduction to real estate for data scientists, and outline a spectrum of data science problems, many of which are being tackled by new \"prop-tech\" companies, while some are yet to be approached. Authors present concrete examples from three of these companies (where the authors work): Airbnb -- the most popular short-term rental marketplace, Cherre -- a real estate data integration platform, and Compass -- the largest independent real estate brokerage in the U.S.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 17s "}} +{"id": "c3aa9543-0695-42b4-99af-d033e3912801", "displayed_text": "Title: Data Science for the Real Estate Industry\n\nAbstract: World's major industries, such as Financial Services, Telecom, Advertising, Healthcare, Education, etc, have attracted the attention of the KDD community for decades. Hundreds of KDD papers have been published on topics related to these industries and dozens of workshops organized---some of which have become an integral part of the conference agenda (e.g. the Health Day). Somewhat unexpectedly, the KDD conference has barely addressed the real estate industry, despite its enormous size and prominence. The reason for that apparent mismatch is two-fold: (a) until recently, the real estate industry did not appreciate the value data science methods could add (with some exceptions, such as econometrics methods for creating real-estate price indices); (b) the Data Science community has not been aware of challenging real estate problems that are perfectly suited to its methods. This tutorial provides a step towards resolving this issue. We provide an introduction to real estate for data scientists, and outline a spectrum of data science problems, many of which are being tackled by new \"prop-tech\" companies, while some are yet to be approached. We present concrete examples from three of these companies (where the authors work): Airbnb -- the most popular short-term rental marketplace, Cherre -- a real estate data integration platform, and Compass -- the largest independent real estate brokerage in the U.S.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "KDD conference has barely addressed the real estate industry, despite its enormous size and prominence.", "Key idea": "This tutorial provides a step towards resolving the problems that the real estate industry did not appreciate the value data science methods could add and the Data Science community has not been aware of challenging real estate problems that are perfectly suited to its methods. ", "Method": "This tutorial provides an introduction to real estate for data scientists, and outline a spectrum of data science problems. It presents concrete examples from three of these companies (where the authors work): Airbnb -- the most popular short-term rental marketplace, Cherre -- a real estate data integration platform, and Compass -- the largest independent real estate brokerage in the U.S.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 4s "}} +{"id": "c5c22c6f-e9e9-4b0a-ac57-d3baa100033b", "displayed_text": "Title: Conditional gradient methods for stochastically constrained convex minimization\n\nAbstract: We propose two novel conditional gradient-based methods for solving structured stochastic convex optimization problems with a large number of linear constraints. Instances of this template naturally arise from SDP-relaxations of combinatorial problems, which involve a number of constraints that is polynomial in the problem dimension. The most important feature of our framework is that only a subset of the constraints is processed at each iteration, thus gaining a computational advantage over prior works that require full passes. Our algorithms rely on variance reduction and smoothing used in conjunction with conditional gradient steps, and are accompanied by rigorous convergence guarantees. Preliminary numerical experiments are provided for illustrating the practical performance of the methods. ", "label_annotations": {"Multi-aspect Summary": {"Context": " Instances of structured stochastic convex optimization problems with a large number of linear constraints naturally arise from SDP-relaxations of combinatorial problems, which involve a number of constraints that is polynomial in the problem dimension.", "Key idea": "Authors propose two novel conditional gradient-based methods for solving structured stochastic convex optimization problems with a large number of linear constraints. The most important feature of proposed framework is that only a subset of the constraints is processed at each iteration, thus gaining a computational advantage over prior works that require full passes. ", "Method": "Authors conduct rigorous convergence analysis theoretically. They also provide preliminary numerical experiments.", "Outcome": "Proposed algorithms rely on variance reduction and smoothing used in conjunction with conditional gradient steps, and are accompanied by rigorous convergence guarantees. Preliminary numerical experiments are provided for illustrating the practical performance of the methods.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 42s "}} +{"id": "c5c22c6f-e9e9-4b0a-ac57-d3baa100033b", "displayed_text": "Title: Conditional gradient methods for stochastically constrained convex minimization\n\nAbstract: We propose two novel conditional gradient-based methods for solving structured stochastic convex optimization problems with a large number of linear constraints. Instances of this template naturally arise from SDP-relaxations of combinatorial problems, which involve a number of constraints that is polynomial in the problem dimension. The most important feature of our framework is that only a subset of the constraints is processed at each iteration, thus gaining a computational advantage over prior works that require full passes. Our algorithms rely on variance reduction and smoothing used in conjunction with conditional gradient steps, and are accompanied by rigorous convergence guarantees. Preliminary numerical experiments are provided for illustrating the practical performance of the methods. ", "label_annotations": {"Multi-aspect Summary": {"Context": "Previous methods for solving structured stochastic convex optimization problems require full passes of the constraints.", "Key idea": "The authors propose two novel conditional gradient-based methods for solving structured stochastic convex optimization problems with a large number of linear constraints, which has a computational advantage over prior works.", "Method": "The authors run preliminary numerical experiments for their proposed methods.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 19s "}} +{"id": "c676aecf-7468-4258-bb41-22bc1811bc3a", "displayed_text": "Title: DMCS : Density Modularity based Community Search\n\nAbstract: Community Search, or finding a connected subgraph (known as a community) containing the given query nodes in a social network, is a fundamental problem. Most of the existing community search models only focus on the internal cohesiveness of a community. However, a high-quality community often has high modularity, which means dense connections inside communities and sparse connections to the nodes outside the community. In this paper, we conduct a pioneer study on searching a community with high modularity. We point out that while modularity has been popularly used in community detection (without query nodes), it has not been adopted for community search, surprisingly, and its application in community search (related to query nodes) brings in new challenges. We address these challenges by designing a new graph modularity function named Density Modularity. To the best of our knowledge, this is the first work on the community search problem using graph modularity. The community search based on the density modularity, termed as DMCS, is to find a community in a social network that contains all the query nodes and has high density-modularity. We prove that the DMCS problem is NP-hard. To efficiently address DMCS, we present new algorithms that run in log-linear time to the graph size. We conduct extensive experimental studies in real-world and synthetic networks, which offer insights into the efficiency and effectiveness of our algorithms. In particular, our algorithm achieves up to 8.5 times higher accuracy in terms of NMI than baseline algorithms.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most of the existing community search models only focus on the internal cohesiveness of a community, and a high-quality community often has dense connections inside communities and sparse connections to the nodes outside the community. Modularity in community search is not practiced and challenging.", "Key idea": "The author designs a first-ever graph modularity on community search and tries to efficiently address density modularity community search by using algorithms that run in log-linear time to the graph size.", "Method": "The author tests the proposed algorithms by conducting studies in real-world and synthetic networks.", "Outcome": "The author\u2019s algorithm achieves up to 8.5 times higher accuracy in terms of NMI than baseline algorithms.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 45s "}} +{"id": "c676aecf-7468-4258-bb41-22bc1811bc3a", "displayed_text": "Title: DMCS : Density Modularity based Community Search\n\nAbstract: Community Search, or finding a connected subgraph (known as a community) containing the given query nodes in a social network, is a fundamental problem. Most of the existing community search models only focus on the internal cohesiveness of a community. However, a high-quality community often has high modularity, which means dense connections inside communities and sparse connections to the nodes outside the community. In this paper, we conduct a pioneer study on searching a community with high modularity. We point out that while modularity has been popularly used in community detection (without query nodes), it has not been adopted for community search, surprisingly, and its application in community search (related to query nodes) brings in new challenges. We address these challenges by designing a new graph modularity function named Density Modularity. To the best of our knowledge, this is the first work on the community search problem using graph modularity. The community search based on the density modularity, termed as DMCS, is to find a community in a social network that contains all the query nodes and has high density-modularity. We prove that the DMCS problem is NP-hard. To efficiently address DMCS, we present new algorithms that run in log-linear time to the graph size. We conduct extensive experimental studies in real-world and synthetic networks, which offer insights into the efficiency and effectiveness of our algorithms. In particular, our algorithm achieves up to 8.5 times higher accuracy in terms of NMI than baseline algorithms.", "label_annotations": {"Multi-aspect Summary": {"Context": "Most of the existing community search models only focus on the internal cohesiveness of a community. \r\nWhile modularity has been popularly used in community detection (without query nodes), it has not been adopted for community search and its application in community search (related to query nodes) brings in new challenges. ", "Key idea": "The authors design a new graph modularity function named Density Modularity, and the community search based on the density modularity, termed as DMCS, is to find a community in a social network that contains all the query nodes and has high density-modularity. And the authors present new algorithms that run in log-linear time to the graph size to address DMCS.", "Method": "The authors conduct extensive experimental studies using new algorithm to address the community search based on the density modularity(DMCS) problem in real-world and synthetic networks in terms of NMI.", "Outcome": " New algorithm solving the community search based on the density modularity(DMCS) problem achieves up to 8.5 times higher accuracy in terms of NMI than baseline algorithms.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "ca53b2c4-2912-4515-aae6-938c3f268a60", "displayed_text": "Title: Detecting epidemic tendency by mining search logs\n\nAbstract: We consider the problem of detecting epidemic tendency by mining search logs. We propose an algorithm based on click-through information to select epidemic related queries/terms. We adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search logs. The results show our algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. We also find the proposed method performs better when combining different ERTs than using single ERT.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose an algorithm based on click-through information to select epidemic related queries/terms, and adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search log.", "Method": "N/A", "Outcome": "The results show the proposed algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. The proposed method performs better when combining different ERTs than using single ERT.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "ca53b2c4-2912-4515-aae6-938c3f268a60", "displayed_text": "Title: Detecting epidemic tendency by mining search logs\n\nAbstract: We consider the problem of detecting epidemic tendency by mining search logs. We propose an algorithm based on click-through information to select epidemic related queries/terms. We adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search logs. The results show our algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. We also find the proposed method performs better when combining different ERTs than using single ERT.", "label_annotations": {"Multi-aspect Summary": {"Context": "The problem of detecting epidemic tendency is important.", "Key idea": "Authors propose an algorithm based on click-through information to select epidemic related queries/terms. Authors adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search logs.", "Method": "Authors design experiments to find epidemic related terms (ERTs).", "Outcome": "Experimental results show the proposed algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. Authors also find the proposed method performs better when combining different ERTs than using single ERT.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 51s "}} +{"id": "ccae9338-7379-4af1-8fea-8945ba429c5c", "displayed_text": "Title: Triplet Loss in Siamese Network for Object Tracking\n\nAbstract: Object tracking is still a critical and challenging problem with many applications in computer vision. For this challenge, more and more researchers pay attention to applying deep learning to get powerful feature for better tracking accuracy. In this paper, a novel triplet loss is proposed to extract expressive deep feature for object tracking by adding it into Siamese network framework instead of pairwise loss for training. Without adding any inputs, our approach is able to utilize more elements for training to achieve more powerful feature via the combination of original samples. Furthermore, we propose a theoretical analysis by combining comparison of gradients and back-propagation, to prove the effectiveness of our method. In experiments, we apply the proposed triplet loss for three real-time trackers based on Siamese network. And the results on several popular tracking benchmarks show our variants operate at almost the same frame-rate with baseline trackers and achieve superior tracking performance than them, as well as the comparable accuracy with recent state-of-the-art real-time trackers.", "label_annotations": {"Multi-aspect Summary": {"Context": "Object tracking is still a critical and challenging problem with many applications in computer vision. For this challenge, more and more researchers pay attention to applying deep learning to get powerful feature for better tracking accuracy.", "Key idea": "The authors propose a novel triplet loss to extract expressive deep feature for object tracking by adding it into Siamese network framework instead of pairwise loss for training.", "Method": "The authors propose a theoretical analysis by combining comparison of gradients and back-propagation, to prove the effectiveness of their proposed method. They apply the proposed triplet loss for three real-time trackers based on Siamese network on several popular tracking benchmarks.", "Outcome": "The experimental results on several popular tracking benchmarks show that the proposed methods operate at almost the same frame-rate with baseline trackers and achieve superior tracking performance than them, as well as the comparable accuracy with recent state-of-the-art real-time trackers.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 41s "}} +{"id": "ccae9338-7379-4af1-8fea-8945ba429c5c", "displayed_text": "Title: Triplet Loss in Siamese Network for Object Tracking\n\nAbstract: Object tracking is still a critical and challenging problem with many applications in computer vision. For this challenge, more and more researchers pay attention to applying deep learning to get powerful feature for better tracking accuracy. In this paper, a novel triplet loss is proposed to extract expressive deep feature for object tracking by adding it into Siamese network framework instead of pairwise loss for training. Without adding any inputs, our approach is able to utilize more elements for training to achieve more powerful feature via the combination of original samples. Furthermore, we propose a theoretical analysis by combining comparison of gradients and back-propagation, to prove the effectiveness of our method. In experiments, we apply the proposed triplet loss for three real-time trackers based on Siamese network. And the results on several popular tracking benchmarks show our variants operate at almost the same frame-rate with baseline trackers and achieve superior tracking performance than them, as well as the comparable accuracy with recent state-of-the-art real-time trackers.", "label_annotations": {"Multi-aspect Summary": {"Context": "Object tracking is still a critical and challenging problem with many applications in computer vision. For this challenge, more and more researchers pay attention to applying deep learning to get powerful feature for better tracking accuracy", "Key idea": "Authors propose a novel triplet loss is proposed to extract expressive deep feature for object tracking by adding it into Siamese network framework instead of pairwise loss for training. Without adding any inputs, the proposed approach is able to utilize more elements for training to achieve more powerful feature via the combination of original samples. Furthermore, authors propose a theoretical analysis by combining comparison of gradients and back-propagation, to prove the effectiveness of our method.", "Method": "Authors conduct experiments on several popular tracking benchmarks.", "Outcome": "Results on several popular tracking benchmarks show the proposed variants operate at almost the same frame-rate with baseline trackers and achieve superior tracking performance than them, as well as the comparable accuracy with recent state-of-the-art real-time trackers.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 13s "}} +{"id": "ce74316d-c5dc-47f1-b0c4-0591bc3fb4b6", "displayed_text": "Title: On the integration of structure indexes and inverted lists\n\nAbstract: Several methods have been proposed to evaluate queries over a native XML DBMS, where the queries specify both path and keyword constraints. These broadly consist of graph traversal approaches, optimized with auxiliary structures known as structure indexes; and approaches based on information-retrieval style inverted lists. We propose a strategy that combines the two forms of auxiliary indexes, and a query evaluation algorithm for branching path expressions based on this strategy. Our technique is general and applicable for a wide range of choices of structure indexes and inverted list join algorithms. Our experiments over the Niagara XML DBMS show the benefit of integrating the two forms of indexes. We also consider algorithmic issues in evaluating path expression queries when the notion of relevance ranking is incorporated. By integrating the above techniques with the Threshold Algorithm proposed by Fagin et al., we obtain instance optimal algorithms to push down top k computation.", "label_annotations": {"Multi-aspect Summary": {"Context": "Several methods have been proposed to evaluate queries over a native XML DBMS and these methods broadly consist of graph traversal approaches, optimized with auxiliary structures known as structure indexes; and approaches based on information-retrieval style inverted lists. ", "Key idea": "The authors propose a strategy that combines the two forms of auxiliary indexes, and a query evaluation algorithm for branching path expressions based on this strategy.", "Method": "The authors perform experiments over the Niagara XML DBMS to evaluate benefits of new strategy.", "Outcome": "Experiments over the Niagara XML DBMS show the benefit of integrating the two forms of indexes. Integrating the techniques that combines the two forms of auxiliary indexes with the Threshold Algorithm proposed by Fagin et al., instance optimal algorithms to push down top k computation has been obtained.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 10s "}} +{"id": "ce74316d-c5dc-47f1-b0c4-0591bc3fb4b6", "displayed_text": "Title: On the integration of structure indexes and inverted lists\n\nAbstract: Several methods have been proposed to evaluate queries over a native XML DBMS, where the queries specify both path and keyword constraints. These broadly consist of graph traversal approaches, optimized with auxiliary structures known as structure indexes; and approaches based on information-retrieval style inverted lists. We propose a strategy that combines the two forms of auxiliary indexes, and a query evaluation algorithm for branching path expressions based on this strategy. Our technique is general and applicable for a wide range of choices of structure indexes and inverted list join algorithms. Our experiments over the Niagara XML DBMS show the benefit of integrating the two forms of indexes. We also consider algorithmic issues in evaluating path expression queries when the notion of relevance ranking is incorporated. By integrating the above techniques with the Threshold Algorithm proposed by Fagin et al., we obtain instance optimal algorithms to push down top k computation.", "label_annotations": {"Multi-aspect Summary": {"Context": "Several methods have been proposed to evaluate queries over a native XML DBMS, where the queries specify both path and keyword constraints. These broadly consist of graph traversal approaches, optimized with auxiliary structures known as structure indexes; and approaches based on information-retrieval style inverted lists.", "Key idea": "The authors propose a strategy that combines the two forms of auxiliary indexes, and a query evaluation algorithm for branching path expressions based on this strategy. They further integrate these techniques with the Threshold Algorithm proposed by Fagin et al., and obtain instance optimal algorithms to push down top k computation.", "Method": "The authors conduct experiments over the Niagara XML DBMS", "Outcome": "The experiments over the Niagara XML DBMS show the benefit of adapting their proposed stratgey.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 36s "}} +{"id": "db900c02-9a35-4a83-aa19-15b763259100", "displayed_text": "Title: Quantum Algorithms for Deep Convolutional Neural Networks\n\nAbstract: Quantum computing is a powerful computational paradigm with applications in several fields, including machine learning. In the last decade, deep learning, and in particular Convolutional Neural Networks (CNN), have become essential for applications in signal processing and image recognition. Quantum deep learning, however, remains a challenging problem, as it is difficult to implement non linearities with quantum unitaries. In this paper we propose a quantum algorithm for evaluating and training deep convolutional neural networks with potential speedups over classical CNNs for both the forward and backward passes. The quantum CNN (QCNN) reproduces completely the outputs of the classical CNN and allows for non linearities and pooling operations. The QCNN is in particular interesting for deep networks and could allow new frontiers in the image recognition domain, by allowing for many more convolution kernels, larger kernels, high dimensional inputs and high depth input channels. We also present numerical simulations for the classification of the MNIST dataset to provide practical evidence for the efficiency of the QCNN.", "label_annotations": {"Multi-aspect Summary": {"Context": "Quantum computing is a powerful computational paradigm with applications in several fields and deep learning is essential for applications in signal processing and image recognition. Quantum deep learning, however, remains a challenging problem, as it is difficult to implement non linearities with quantum unitaries.", "Key idea": "The authors propose a quantum algorithm for evaluating and training deep convolutional neural networks with potential speedups over classical CNNs for both the forward and backward passes.", "Method": "The authors evaluate their proposed method using the classification of the MNIST dataset.", "Outcome": "The authors present numerical simulations for the classification of the MNIST dataset to provide practical evidence for the efficiency of the QCNN.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 3s "}} +{"id": "db900c02-9a35-4a83-aa19-15b763259100", "displayed_text": "Title: Quantum Algorithms for Deep Convolutional Neural Networks\n\nAbstract: Quantum computing is a powerful computational paradigm with applications in several fields, including machine learning. In the last decade, deep learning, and in particular Convolutional Neural Networks (CNN), have become essential for applications in signal processing and image recognition. Quantum deep learning, however, remains a challenging problem, as it is difficult to implement non linearities with quantum unitaries. In this paper we propose a quantum algorithm for evaluating and training deep convolutional neural networks with potential speedups over classical CNNs for both the forward and backward passes. The quantum CNN (QCNN) reproduces completely the outputs of the classical CNN and allows for non linearities and pooling operations. The QCNN is in particular interesting for deep networks and could allow new frontiers in the image recognition domain, by allowing for many more convolution kernels, larger kernels, high dimensional inputs and high depth input channels. We also present numerical simulations for the classification of the MNIST dataset to provide practical evidence for the efficiency of the QCNN.", "label_annotations": {"Multi-aspect Summary": {"Context": "Quantum deep learning remains a challenging problem, as it is difficult to implement non linearities with quantum unitaries.", "Key idea": "The author proposes a quantum algorithm for evaluating and training deep convolutional neural networks with potential speedups over classical CNNs for both the forward and backward passes. The quantum CNN (QCNN) reproduces completely the outputs of the classical CNN and allows for non linearities and pooling operations and it is in particular interesting for deep networks and could allow new frontiers in the image recognition domain.", "Method": "The author presents numerical simulations for the classification of the MNIST dataset.", "Outcome": "The numerical simulations for the classification of the MNIST dataset provide practical evidence for the efficiency of the QCNN.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 25s "}} +{"id": "dca09f6f-b63a-42a1-9eb5-fbfa45bc6389", "displayed_text": "Title: Modeling Heart Rate and Activity Data for Personalized Fitness Recommendation\n\nAbstract: Activity logs collected from wearable devices (e.g. Apple Watch, Fitbit, etc.) are a promising source of data to facilitate a wide range of applications such as personalized exercise scheduling, workout recommendation, and heart rate anomaly detection. However, such data are heterogeneous, noisy, diverse in scale and resolution, and have complex interdependencies, making them challenging to model. In this paper, we develop context-aware sequential models to capture the personalized and temporal patterns of fitness data. Specifically, we propose FitRec - an LSTM-based model that captures two levels of context information: context within a specific activity, and context across a user's activity history. We are specifically interested in (a) estimating a user's heart rate profile for a candidate activity; and (b) predicting and recommending suitable activities on this basis. We evaluate our model on a novel dataset containing over 250 thousand workout records coupled with hundreds of millions of parallel sensor measurements (e.g. heart rate, GPS) and metadata. We demonstrate that the model is able to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise. We evaluate the proposed model against baselines on several personalized recommendation tasks, showing the promise of using wearable data for activity modeling and recommendation.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Activity logs collected from wearable devices (e.g. Apple Watch, Fitbit, etc.) are a promising source of data to facilitate a wide range of applications such as personalized exercise scheduling, workout recommendation, and heart rate anomaly detection. However, such data are heterogeneous, noisy, diverse in scale and resolution, and have complex interdependencies, making them challenging to model.", "Key idea": "Authors develop context-aware sequential models to capture the personalized and temporal patterns of fitness data. Specifically, they propose FitRec - an LSTM-based model that captures two levels of context information: context within a specific activity, and context across a user's activity history.", "Method": "Authors evaluate our model on a novel dataset containing over 250 thousand workout records coupled with hundreds of millions of parallel sensor measurements (e.g. heart rate, GPS) and metadata. ", "Outcome": "Authors demonstrate that the model is able to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise. \r\nAuthors also evaluate the proposed model against baselines on several personalized recommendation tasks, showing the promise of using wearable data for activity modeling and recommendation.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 31s "}} +{"id": "dca09f6f-b63a-42a1-9eb5-fbfa45bc6389", "displayed_text": "Title: Modeling Heart Rate and Activity Data for Personalized Fitness Recommendation\n\nAbstract: Activity logs collected from wearable devices (e.g. Apple Watch, Fitbit, etc.) are a promising source of data to facilitate a wide range of applications such as personalized exercise scheduling, workout recommendation, and heart rate anomaly detection. However, such data are heterogeneous, noisy, diverse in scale and resolution, and have complex interdependencies, making them challenging to model. In this paper, we develop context-aware sequential models to capture the personalized and temporal patterns of fitness data. Specifically, we propose FitRec - an LSTM-based model that captures two levels of context information: context within a specific activity, and context across a user's activity history. We are specifically interested in (a) estimating a user's heart rate profile for a candidate activity; and (b) predicting and recommending suitable activities on this basis. We evaluate our model on a novel dataset containing over 250 thousand workout records coupled with hundreds of millions of parallel sensor measurements (e.g. heart rate, GPS) and metadata. We demonstrate that the model is able to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise. We evaluate the proposed model against baselines on several personalized recommendation tasks, showing the promise of using wearable data for activity modeling and recommendation.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Since data collected from wearable devices are heterogeneous, noisy, diverse in scale and resolution, and have complex interdependencies, it's challenging to model these data.", "Key idea": "The authors propose FitRec, an LSTM-based model, which is a context-aware sequential models to capture the personalized and temporal patterns of fitness data.", "Method": "The authors evaluate FitRec model on a novel dataset containing over 250 thousand workout records coupled with hundreds of millions of parallel sensor measurements (e.g. heart rate, GPS) and metadata. And the authors compare FitRec model against baselines on several personalized recommendation tasks.", "Outcome": "FitRec model is able to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise. FitRec model have better performance on several personalized recommendation task than baselines.", "Future Impact": "FitRec model provides better promise of using wearable data for activity modeling and recommendation."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "dd282632-ee41-45da-add8-d68d89c57e2d", "displayed_text": "Title: Bagging Regional Classification Activation Maps for Weakly Supervised Object Localization.\n\nAbstract: Classification activation map (CAM), utilizing the classification structure to generate pixel-wise localization maps, is a crucial mechanism for weakly supervised object localization (WSOL). However, CAM directly uses the classifier trained on image-level features to locate objects, making it prefers to discern global discriminative factors rather than regional object cues. Thus only the discriminative locations are activated when feeding pixel-level features into this classifier. To solve this issue, this paper elaborates a plug-and-play mechanism called BagCAMs to better project a well-trained classifier for the localization task without refining or re-training the baseline structure. Our BagCAMs adopts a proposed regional localizer generation (RLG) strategy to define a set of regional localizers and then derive them from a well-trained classifier. These regional localizers can be viewed as the base learner that only discerns region-wise object factors for localization tasks, and their results can be effectively weighted by our BagCAMs to form the final localization map. Experiments indicate that adopting our proposed BagCAMs can improve the performance of baseline WSOL methods to a great extent and obtains state-of-the-art performance on three WSOL benchmarks. Code are released at https://github.com/zh460045050/BagCAMs.", "label_annotations": {"Multi-aspect Summary": {"Context": "Classification activation map (CAM) is a crucial mechanism for weakly supervised object localization (WSOL). However, CAM directly uses the classifier trained on image-level features to locate objects, making it prefers to discern global discriminative factors rather than regional object cues, so only the discriminative locations are activated when feeding pixel-level features into this classifier.", "Key idea": "The authors elaborate a plug-and-play mechanism called BagCAMs to better project a well-trained classifier for the localization task without refining or re-training the baseline structure.", "Method": "The authors evaluate BagCAMs on three WSOL benchmarks and compare it to baseline WSOL methods.", "Outcome": "Experiments indicate that adopting the proposed BagCAMs can improve the performance of baseline WSOL methods to a great extent and obtains state-of-the-art performance on three WSOL benchmarks.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 10m 36s "}} +{"id": "dd282632-ee41-45da-add8-d68d89c57e2d", "displayed_text": "Title: Bagging Regional Classification Activation Maps for Weakly Supervised Object Localization.\n\nAbstract: Classification activation map (CAM), utilizing the classification structure to generate pixel-wise localization maps, is a crucial mechanism for weakly supervised object localization (WSOL). However, CAM directly uses the classifier trained on image-level features to locate objects, making it prefers to discern global discriminative factors rather than regional object cues. Thus only the discriminative locations are activated when feeding pixel-level features into this classifier. To solve this issue, this paper elaborates a plug-and-play mechanism called BagCAMs to better project a well-trained classifier for the localization task without refining or re-training the baseline structure. Our BagCAMs adopts a proposed regional localizer generation (RLG) strategy to define a set of regional localizers and then derive them from a well-trained classifier. These regional localizers can be viewed as the base learner that only discerns region-wise object factors for localization tasks, and their results can be effectively weighted by our BagCAMs to form the final localization map. Experiments indicate that adopting our proposed BagCAMs can improve the performance of baseline WSOL methods to a great extent and obtains state-of-the-art performance on three WSOL benchmarks. Code are released at https://github.com/zh460045050/BagCAMs.", "label_annotations": {"Multi-aspect Summary": {"Context": "Classification activation map (CAM), utilizing the classification structure to generate pixel-wise localization maps, is a crucial mechanism for weakly supervised object localization (WSOL). However, CAM directly uses the classifier trained on image-level features to locate objects, making it prefers to discern global discriminative factors rather than regional object cues. Thus only the discriminative locations are activated when feeding pixel-level features into this classifier. ", "Key idea": "Authors propose a plug-and-play mechanism called BagCAMs to better project a well-trained classifier for the localization task without refining or re-training the baseline structure. BagCAMs adopts a proposed regional localizer generation (RLG) strategy to define a set of regional localizers and then derive them from a well-trained classifier. These regional localizers can be viewed as the base learner that only discerns region-wise object factors for localization tasks, and their results can be effectively weighted by our BagCAMs to form the final localization map. ", "Method": "Authors experiment on three eakly supervised object localization (WSOL) benchmarks.", "Outcome": "Authors show that adopting our proposed BagCAMs can improve the performance of baseline WSOL methods to a great extent and obtains state-of-the-art performance on three WSOL benchmarks. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 46s "}} +{"id": "ddf8f49c-342a-4cd0-8b3b-b588af08ed0d", "displayed_text": "Title: On the effectiveness of evaluating retrieval systems in the absence of relevance judgments\n\nAbstract: Soboroff, Nicholas and Cahan recently proposed a method for evaluating the performance of retrieval systems without relevance judgments. They demonstrated that the system evaluations produced by their methodology are correlated with actual evaluations using relevance judgments in the TREC competition. In this work, we propose an explanation for this phenomenon. We devise a simple measure for quantifying the similarity of retrieval systems by assessing the similarity of their retrieved results. Then, given a collection of retrieval systems and their retrieved results, we use this measure to assess the average similarity of a system to the other systems in the collection. We demonstrate that evaluating retrieval systems according to average similarity yields results quite similar to the methodology proposed by Soboroff et~al., and we further demonstrate that these two techniques are in fact highly correlated. Thus, the techniques are effectively evaluating and ranking retrieval systems by popularity\" as opposed to performance.", "label_annotations": {"Multi-aspect Summary": {"Context": "Soboroff, Nicholas and Cahan proposed a method for evaluating the performance of retrieval systems without relevance judgments and demonstrated that the system evaluations produced by their methodology are correlated with actual evaluations using relevance judgments in the TREC competition.", "Key idea": "The authors propose an explanation for phenomenon, the system evaluations are correlated with actual evaluations using relevance judgments in the TREC competition.", "Method": "The authors devise a simple measure for quantifying the similarity of retrieval systems by assessing the similarity of their retrieved results and they use this measure to assess the average similarity of a system to the other systems in the collection. ", "Outcome": "The authors prove that evaluating retrieval systems according to average similarity yields results are quite similar to the methodology proposed by Soboroff et~al and these two techniques are in fact highly correlated. ", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 32s "}} +{"id": "ddf8f49c-342a-4cd0-8b3b-b588af08ed0d", "displayed_text": "Title: On the effectiveness of evaluating retrieval systems in the absence of relevance judgments\n\nAbstract: Soboroff, Nicholas and Cahan recently proposed a method for evaluating the performance of retrieval systems without relevance judgments. They demonstrated that the system evaluations produced by their methodology are correlated with actual evaluations using relevance judgments in the TREC competition. In this work, we propose an explanation for this phenomenon. We devise a simple measure for quantifying the similarity of retrieval systems by assessing the similarity of their retrieved results. Then, given a collection of retrieval systems and their retrieved results, we use this measure to assess the average similarity of a system to the other systems in the collection. We demonstrate that evaluating retrieval systems according to average similarity yields results quite similar to the methodology proposed by Soboroff et~al., and we further demonstrate that these two techniques are in fact highly correlated. Thus, the techniques are effectively evaluating and ranking retrieval systems by popularity\" as opposed to performance.", "label_annotations": {"Multi-aspect Summary": {"Context": "Soboroff, Nicholas and Cahan recently proposed a method for evaluating the performance of retrieval systems without relevance judgments. They demonstrated that the system evaluations produced by their methodology are correlated with actual evaluations using relevance judgments in the TREC competition.", "Key idea": "The authors propose an explanation for the phenomenon that the system evaluations produced by the methodology introduced by Soboroff et al. are correlated with actual evaluations using relevance judgments in the TREC competition.", "Method": "The authors devise a simple measure for quantifying the similarity of retrieval systems by assessing the similarity of their retrieved results. Then, they compare the measured similarity and compare it the the results yielded by the methodology proposed by Soboroff et al.", "Outcome": "The authors demonstrate that evaluating retrieval systems according to average similarity yields results quite similar to the methodology proposed by Soboroff et al., and they further demonstrate that these two techniques are in fact highly correlated.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 7m 17s "}} +{"id": "ded9a095-d94d-4ccd-8825-283ec4bb7093", "displayed_text": "Title: Evolution of design competence in UX practice\n\nAbstract: There has been increasing interest in the adoption of UX within corporate environments, and what competencies translate into effective UX design. This paper addresses the space between pedagogy and UX practice through the lens of competence, with the goal of understanding how students are initiated into the practice community, how their perception of competence shifts over time, and what factors influence this shift. A 12-week longitudinal data collection, including surveys and interviews, documents this shift, with participants beginning internships and full-time positions in UX. Students and early professionals were asked to assess their level of competence and factors that influenced competence. A co-construction of identity between the designer and their environment is proposed, with a variety of factors relating to tool and representational knowledge, complexity, and corporate culture influencing perceptions of competence in UX over time. Opportunities for future research, particularly in building an understanding of competency in UX based on this preliminary framing of early UX practice are addressed.", "label_annotations": {"Multi-aspect Summary": {"Context": "There has been increasing interest in the adoption of UX within corporate environments, and what competencies translate into effective UX design.", "Key idea": "The author addresses the space between pedagogy and UX practice through the lens of competence, in order to understand how students are initiated into the practice community, how their perception of competence shifts over time, and what factors influence this shift. After the data collection on the observation, a co-construction of identity between the designer and their environment is proposed.", "Method": "The author conducts a 12-week longitudinal data collection, including surveys and interviews, documents this shift, with participants beginning internships and full-time positions in UX. Students and early professionals were asked to assess their level of competence and factors that influenced competence.", "Outcome": "N/A", "Future Impact": "Opportunities for future research, particularly in building an understanding of competency in UX based on this preliminary framing of early UX practice are addressed."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 26s "}} +{"id": "ded9a095-d94d-4ccd-8825-283ec4bb7093", "displayed_text": "Title: Evolution of design competence in UX practice\n\nAbstract: There has been increasing interest in the adoption of UX within corporate environments, and what competencies translate into effective UX design. This paper addresses the space between pedagogy and UX practice through the lens of competence, with the goal of understanding how students are initiated into the practice community, how their perception of competence shifts over time, and what factors influence this shift. A 12-week longitudinal data collection, including surveys and interviews, documents this shift, with participants beginning internships and full-time positions in UX. Students and early professionals were asked to assess their level of competence and factors that influenced competence. A co-construction of identity between the designer and their environment is proposed, with a variety of factors relating to tool and representational knowledge, complexity, and corporate culture influencing perceptions of competence in UX over time. Opportunities for future research, particularly in building an understanding of competency in UX based on this preliminary framing of early UX practice are addressed.", "label_annotations": {"Multi-aspect Summary": {"Context": "There has been increasing interest in the adoption of UX within corporate environments, and what competencies translate into effective UX design.", "Key idea": "The authors address the space between pedagogy and UX practice through the lens of competence, with the goal of understanding how students are initiated into the practice community, how their perception of competence shifts over time, and what factors influence this shift.", "Method": "The authors collect data from surveys and interviews. Students and early professionals were asked to assess their level of competence and factors that influenced competence.", "Outcome": "A co-construction of identity between the designer and their environment is proposed, with a variety of factors relating to tool and representational knowledge, complexity, and corporate culture influencing perceptions of competence in UX over time.", "Future Impact": "The authors address opportunities for future research, particularly in building an understanding of competency in UX based on this preliminary framing of early UX practice."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 59s "}} +{"id": "dfba01c5-a632-4394-8607-9a32f20e526c", "displayed_text": "Title: Modeling geometric structure and illumination variation of a scene from real images\n\nAbstract: We present in this paper a system which automatically builds, from real images, a scene model containing both 3D geometric information of the scene structure and its photometric information under various illumination conditions. The geometric structure is recovered from images taken from distinct viewpoints. Structure-from-motion and correlation-based stereo techniques are used to match pixels between images of different viewpoints and to reconstruct the scene in 3D space. The photometric property is extracted from images taken under different illumination conditions (orientation, position and intensity of the light sources). This is achieved by computing a low-dimensional linear space of the spatio-illumination volume, and is represented by a set of basis images. The model that has been built can be used to create realistic renderings from different viewpoints and illumination conditions. Applications include object recognition, virtual reality and product advertisement.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author presents a system that automatically builds, from real images, a scene model containing both 3D geometric information of the scene structure and its photometric information under various illumination conditions. The geometric structure is recovered from images taken from distinct viewpoints and processed by matching pixels between images of different viewpoints, and the photometric property is extracted from images taken under different illumination conditions and processed by computing a low-dimensional linear space of the spatio-illumination volume.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "Applications include object recognition, virtual reality and product advertisement."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 2s "}} +{"id": "dfba01c5-a632-4394-8607-9a32f20e526c", "displayed_text": "Title: Modeling geometric structure and illumination variation of a scene from real images\n\nAbstract: We present in this paper a system which automatically builds, from real images, a scene model containing both 3D geometric information of the scene structure and its photometric information under various illumination conditions. The geometric structure is recovered from images taken from distinct viewpoints. Structure-from-motion and correlation-based stereo techniques are used to match pixels between images of different viewpoints and to reconstruct the scene in 3D space. The photometric property is extracted from images taken under different illumination conditions (orientation, position and intensity of the light sources). This is achieved by computing a low-dimensional linear space of the spatio-illumination volume, and is represented by a set of basis images. The model that has been built can be used to create realistic renderings from different viewpoints and illumination conditions. Applications include object recognition, virtual reality and product advertisement.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a system which automatically builds, from real images, a scene model containing both 3D geometric information of the scene structure and its photometric information under various illumination conditions.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The model that has been built can be used to create realistic renderings from different viewpoints and illumination conditions. Applications include object recognition, virtual reality and product advertisement."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "e10b883a-ded6-4b4e-9934-6daba2d2f2b2", "displayed_text": "Title: Semantic hashing using tags and topic modeling\n\nAbstract: It is an important research problem to design efficient and effective solutions for large scale similarity search. One popular strategy is to represent data examples as compact binary codes through semantic hashing, which has produced promising results with fast search speed and low storage cost. Many existing semantic hashing methods generate binary codes for documents by modeling document relationships based on similarity in a keyword feature space. Two major limitations in existing methods are: (1) Tag information is often associated with documents in many real world applications, but has not been fully exploited yet; (2) The similarity in keyword feature space does not fully reflect semantic relationships that go beyond keyword matching. This paper proposes a novel hashing approach, Semantic Hashing using Tags and Topic Modeling (SHTTM), to incorporate both the tag information and the similarity information from probabilistic topic modeling. In particular, a unified framework is designed for ensuring hashing codes to be consistent with tag information by a formal latent factor model and preserving the document topic/semantic similarity that goes beyond keyword matching. An iterative coordinate descent procedure is proposed for learning the optimal hashing codes. An extensive set of empirical studies on four different datasets has been conducted to demonstrate the advantages of the proposed SHTTM approach against several other state-of-the-art semantic hashing techniques. Furthermore, experimental results indicate that the modeling of tag information and utilizing topic modeling are beneficial for improving the effectiveness of hashing separately, while the combination of these two techniques in the unified framework obtains even better results.", "label_annotations": {"Multi-aspect Summary": {"Context": "Designing efficient and effective solutions for large scale similarity search is an important research problem, and one popular strategy is to represent data examples as compact binary codes through semantic hashing. Many existing semantic hashing methods generate binary codes for documents by modeling document relationships based on similarity in a keyword feature space, which has two major limitations.", "Key idea": "The authors propose a novel hashing approach, Semantic Hashing using Tags and Topic Modeling (SHTTM), to incorporate both the tag information and the similarity information from probabilistic topic modeling.", "Method": "The authors evaluate the proposed SHTTM on four different datasets and compare it with several other state-of-the-art semantic hashing techniques.", "Outcome": "An extensive set of empirical studies on four different datasets has been conducted to demonstrate the advantages of the proposed SHTTM approach against several other state-of-the-art semantic hashing techniques. Furthermore, experimental results indicate that the modeling of tag information and utilizing topic modeling are beneficial for improving the effectiveness of hashing separately, while the combination of these two techniques in the unified framework obtains even better results.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 30s "}} +{"id": "e10b883a-ded6-4b4e-9934-6daba2d2f2b2", "displayed_text": "Title: Semantic hashing using tags and topic modeling\n\nAbstract: It is an important research problem to design efficient and effective solutions for large scale similarity search. One popular strategy is to represent data examples as compact binary codes through semantic hashing, which has produced promising results with fast search speed and low storage cost. Many existing semantic hashing methods generate binary codes for documents by modeling document relationships based on similarity in a keyword feature space. Two major limitations in existing methods are: (1) Tag information is often associated with documents in many real world applications, but has not been fully exploited yet; (2) The similarity in keyword feature space does not fully reflect semantic relationships that go beyond keyword matching. This paper proposes a novel hashing approach, Semantic Hashing using Tags and Topic Modeling (SHTTM), to incorporate both the tag information and the similarity information from probabilistic topic modeling. In particular, a unified framework is designed for ensuring hashing codes to be consistent with tag information by a formal latent factor model and preserving the document topic/semantic similarity that goes beyond keyword matching. An iterative coordinate descent procedure is proposed for learning the optimal hashing codes. An extensive set of empirical studies on four different datasets has been conducted to demonstrate the advantages of the proposed SHTTM approach against several other state-of-the-art semantic hashing techniques. Furthermore, experimental results indicate that the modeling of tag information and utilizing topic modeling are beneficial for improving the effectiveness of hashing separately, while the combination of these two techniques in the unified framework obtains even better results.", "label_annotations": {"Multi-aspect Summary": {"Context": "There are two major limitations in existing semantic hashing methods: (1) Tag information is often associated with documents in many real world applications, but has not been fully exploited yet; (2) The similarity in keyword feature space does not fully reflect semantic relationships that go beyond keyword matching.", "Key idea": "The authors proposes a novel hashing approach, Semantic Hashing using Tags and Topic Modeling (SHTTM), to incorporate both the tag information and the similarity information from probabilistic topic modeling. ", "Method": "The authors conduct an extensive set of empirical studies on four different datasets to evaluate the proposed SHTTM approach against several other state-of-the-art semantic hashing techniques.", "Outcome": "Experimental results indicate that the modeling of tag information and utilizing topic modeling are beneficial for improving the effectiveness of hashing separately, while the combination of these two techniques in the unified framework obtains even better results.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 13s "}} +{"id": "e1a191db-6463-4be9-883d-dfc7fb05c5f4", "displayed_text": "Title: Dependency Parsing with Bounded Block Degree and Well-nestedness via Lagrangian Relaxation and Branch-and-Bound\n\nAbstract: We present a novel dependency parsing method which enforces two structural properties on dependency trees: bounded block degree and well-nestedness. These properties are useful to better represent the set of admissible dependency structures in treebanks and connect dependency parsing to context-sensitive grammatical formalisms. We cast this problem as an Integer Linear Program that we solve with Lagrangian Relaxation from which we derive a heuristic and an exact method based on a Branch-and-Bound search. Experimentally, we see that these methods are efficient and competitive compared to a baseline unconstrained parser, while enforcing structural properties in all cases.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a novel dependency parsing method which enforces two structural properties on dependency trees: bounded block degree and well-nestedness.", "Method": "N/A", "Outcome": "The authors show that the proposed methods are efficient and competitive compared to a baseline unconstrained parser, while enforcing structural properties in all cases.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 58s "}} +{"id": "e1a191db-6463-4be9-883d-dfc7fb05c5f4", "displayed_text": "Title: Dependency Parsing with Bounded Block Degree and Well-nestedness via Lagrangian Relaxation and Branch-and-Bound\n\nAbstract: We present a novel dependency parsing method which enforces two structural properties on dependency trees: bounded block degree and well-nestedness. These properties are useful to better represent the set of admissible dependency structures in treebanks and connect dependency parsing to context-sensitive grammatical formalisms. We cast this problem as an Integer Linear Program that we solve with Lagrangian Relaxation from which we derive a heuristic and an exact method based on a Branch-and-Bound search. Experimentally, we see that these methods are efficient and competitive compared to a baseline unconstrained parser, while enforcing structural properties in all cases.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors propose a dependency parsing method that enforces two structural properties, bounded block degree and well-nestedness on dependency trees. The problem is treated as an Integer Linear Program and solved with Lagrangian Relaxation.", "Method": "The authors compare the proposed dependency parsing method with a baseline unconstrained parser.", "Outcome": "The proposed methods are efficient and competitive compared to a baseline unconstrained parser, while enforcing structural properties in all cases.", "Future Impact": "The two enforced structural properties of the proposed method are useful to better represent the set of admissible dependency structures in treebanks and connect dependency parsing to context-sensitive grammatical formalisms."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 12m 10s "}} +{"id": "e2466d98-98f9-41a8-8a38-e862adc2ed47", "displayed_text": "Title: A conceptual model for IT Governance: A case study research\n\nAbstract: The purpose of this paper is to explore the importance of Information Technology (IT) Governance models for public organizations and presenting an IT Governance model that can be adopted by both practitioners and researchers. A review of the literature in IT Governance has been initiated to shape the intended theoretical background of this study. The systematic literature review formalizes a richer context for the IT Governance concept. An empirical survey, using a questionnaire based on COBIT 4.1 maturity model used to investigate IT Governance practice in multiple case studies from Kingdom of Bahrain. This method enabled the researcher to gain insights to evaluate IT Governance practices. The results of this research will enable public sector organizations to adopt an IT Governance model in a simple and dynamic manner. The model provides a basic structure of a concept; for instance, this allows organizations to gain a better perspective on IT Governance processes and provides a clear focus for decision-making attention. IT Governance model also forms as a basis for further research in IT Governance adoption models and bridges the gap between conceptual frameworks, real life and functioning governance.", "label_annotations": {"Multi-aspect Summary": {"Context": "Information Technology (IT) Governance models for public organizations is important.", "Key idea": "The author presents an IT Governance model that is constructed from review of literature and an empirical survey that uses a questionnaire based on COBIT 4.1 maturity model.", "Method": "N/A", "Outcome": "The model provides a basic structure of a concept; for instance, this allows organizations to gain a better perspective on IT Governance processes and provides a clear focus for decision-making attention.", "Future Impact": "The results of this research will enable public sector organizations to adopt an IT Governance model in a simple and dynamic manner. IT Governance model also forms as a basis for further research in IT Governance adoption models and bridges the gap between conceptual frameworks, real life and functioning governance."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 29s "}} +{"id": "e2466d98-98f9-41a8-8a38-e862adc2ed47", "displayed_text": "Title: A conceptual model for IT Governance: A case study research\n\nAbstract: The purpose of this paper is to explore the importance of Information Technology (IT) Governance models for public organizations and presenting an IT Governance model that can be adopted by both practitioners and researchers. A review of the literature in IT Governance has been initiated to shape the intended theoretical background of this study. The systematic literature review formalizes a richer context for the IT Governance concept. An empirical survey, using a questionnaire based on COBIT 4.1 maturity model used to investigate IT Governance practice in multiple case studies from Kingdom of Bahrain. This method enabled the researcher to gain insights to evaluate IT Governance practices. The results of this research will enable public sector organizations to adopt an IT Governance model in a simple and dynamic manner. The model provides a basic structure of a concept; for instance, this allows organizations to gain a better perspective on IT Governance processes and provides a clear focus for decision-making attention. IT Governance model also forms as a basis for further research in IT Governance adoption models and bridges the gap between conceptual frameworks, real life and functioning governance.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors explore the importance of Information Technology (IT) Governance models for public organizations and presenting an IT Governance model that can be adopted by both practitioners and researchers.", "Method": "The authors initiate a systematic literature review.", "Outcome": "N/A", "Future Impact": "The authors show that IT Governance model allows organizations to gain a better perspective on IT Governance processes and provides a clear focus for decision-making attention and forms as a basis for further research in IT Governance adoption models and bridges the gap between conceptual frameworks, real life and functioning governance."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 28m 33s "}} +{"id": "e58b9947-7a3a-414d-a0e8-d6cf02ed7127", "displayed_text": "Title: Exploring recommendations in internet of things\n\nAbstract: With recent advances in radio-frequency identification (RFID), wireless sensor networks, and Web-based services, physical things are becoming an integral part of the emerging ubiquitous Web. In this paper, we focus on the things recommendation problem in Internet of Things (IoT). In particular, we propose a unified probabilistic based framework by fusing information across relationships between users (i.e., usersu0027social network) and things (i.e., things correlations) to make more accurate recommendations. The proposed approach not only inherits the advantages of the matrix factorization, but also exploits the merits of social relationships and thing-thing correlations. We validate our approach based on an Internet of Things platform and the experimental results demonstrate its feasibility and effectiveness.", "label_annotations": {"Multi-aspect Summary": {"Context": "With recent advances in radio-frequency identification (RFID), wireless sensor networks, and Web-based services, physical things are becoming an integral part of the emerging ubiquitous Web.", "Key idea": "Authors propose a unified probabilistic based framework by fusing information across relationships between users (i.e., social network) and things (i.e., things correlations) to make more accurate recommendations. The proposed approach not only inherits the advantages of the matrix factorization, but also exploits the merits of social relationships and thing-thing correlations", "Method": "Authors design experiments on an Internet of Things platform.", "Outcome": "Authors validate their approach based on an Internet of Things platform and the experimental results demonstrate its feasibility and effectiveness.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 21s "}} +{"id": "e58b9947-7a3a-414d-a0e8-d6cf02ed7127", "displayed_text": "Title: Exploring recommendations in internet of things\n\nAbstract: With recent advances in radio-frequency identification (RFID), wireless sensor networks, and Web-based services, physical things are becoming an integral part of the emerging ubiquitous Web. In this paper, we focus on the things recommendation problem in Internet of Things (IoT). In particular, we propose a unified probabilistic based framework by fusing information across relationships between users (i.e., usersu0027social network) and things (i.e., things correlations) to make more accurate recommendations. The proposed approach not only inherits the advantages of the matrix factorization, but also exploits the merits of social relationships and thing-thing correlations. We validate our approach based on an Internet of Things platform and the experimental results demonstrate its feasibility and effectiveness.", "label_annotations": {"Multi-aspect Summary": {"Context": "With recent advances in radio-frequency identification (RFID), wireless sensor networks, and Web-based services, physical things are becoming an integral part of the emerging ubiquitous Web.", "Key idea": "The authors propose a unified probabilistic based framework by fusing information across relationships between users (i.e., users' social network) and things (i.e., things correlations) to make more accurate recommendations.", "Method": "The authors validate their approach based on an Internet of Things platform.", "Outcome": "The proposed approach not only inherits the advantages of the matrix factorization, but also exploits the merits of social relationships and thing-thing correlations. The experimental results demonstrate its feasibility and effectiveness.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 4s "}} +{"id": "e827ee51-aafd-4f3b-99ca-60a5e289a555", "displayed_text": "Title: Recognizing textual entailment via atomic propositions\n\nAbstract: This paper describes Macquarie Universityu0027s Centre for Language Technology contribution to the PASCAL 2005 Recognizing Textual Entailment challenge. Our main aim was to test the practicability of a purely logical approach. For this, atomic propositions were extracted from both the text and the entailment hypothesis and they were expressed in a custom logical notation. The text entails the hypothesis if every proposition of the hypothesis is entailed by some proposition in the text. To extract the propositions and encode them into a logical notation the system uses the output of Link Parser. To detect the independent entailment relations the system relies on the use of Otter and WordNet.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The author aims to test the practicability of a purely logical approach by extracting atomic propositions from both the text and the entailment hypothesis. The system uses the output of Link Parser to extract the propositions and encode them into a logical notation and uses Otter and WordNet to detect the independent entailment relations.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 16s "}} +{"id": "e827ee51-aafd-4f3b-99ca-60a5e289a555", "displayed_text": "Title: Recognizing textual entailment via atomic propositions\n\nAbstract: This paper describes Macquarie Universityu0027s Centre for Language Technology contribution to the PASCAL 2005 Recognizing Textual Entailment challenge. Our main aim was to test the practicability of a purely logical approach. For this, atomic propositions were extracted from both the text and the entailment hypothesis and they were expressed in a custom logical notation. The text entails the hypothesis if every proposition of the hypothesis is entailed by some proposition in the text. To extract the propositions and encode them into a logical notation the system uses the output of Link Parser. To detect the independent entailment relations the system relies on the use of Otter and WordNet.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors test the practicability of a purely logical approach in the PASCAL 2005 Recognizing Textual Entailment challenge.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 41s "}} +{"id": "ea414544-c89c-4039-8227-23b11e9a1239", "displayed_text": "Title: A Practically Unsupervised Learning Method to Identify Single-Snippet Answers to Definition Questions on the Web\n\nAbstract: We present a practically unsupervised learning method to produce single-snippet answers to definition questions in question answering systems that supplement Web search engines. The method exploits on-line encyclopedias and dictionaries to generate automatically an arbitrarily large number of positive and negative definition examples, which are then used to train an SVM to separate the two classes. We show experimentally that the proposed method is viable, that it outperforms the alternative of training the system on questions and news articles from TREC, and that it helps the search engine handle definition questions significantly better.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a practically unsupervised learning method to produce single-snippet answers to definition questions in question answering systems that supplement Web search engines. ", "Method": "The authors conduct experiments to evaluate new practically unsupervised learning method and compare it with the alternative of training the system on questions and news articles from TREC.", "Outcome": "Experiments prove that new practically unsupervised learning method is viable and it outperforms the alternative of training the system on questions and news articles from TREC, and that it helps the search engine handle definition questions significantly better.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 3s "}} +{"id": "ea414544-c89c-4039-8227-23b11e9a1239", "displayed_text": "Title: A Practically Unsupervised Learning Method to Identify Single-Snippet Answers to Definition Questions on the Web\n\nAbstract: We present a practically unsupervised learning method to produce single-snippet answers to definition questions in question answering systems that supplement Web search engines. The method exploits on-line encyclopedias and dictionaries to generate automatically an arbitrarily large number of positive and negative definition examples, which are then used to train an SVM to separate the two classes. We show experimentally that the proposed method is viable, that it outperforms the alternative of training the system on questions and news articles from TREC, and that it helps the search engine handle definition questions significantly better.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors present a practically unsupervised learning method to produce single-snippet answers to definition questions in question answering systems that supplement Web search engines.", "Method": "The authors compare the proposed methods with an alternative of training the question answering system on questions and news articles from TREC", "Outcome": "The authors show experimentally that the proposed method is viable, that it outperforms the alternative of training the system on questions and news articles from TREC, and that it helps the search engine handle definition questions significantly better.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 16s "}} +{"id": "eb15ebe7-aa58-4a98-8f9e-939967c6359f", "displayed_text": "Title: Synchronization of Group-labelled Multi-graphs.\n\nAbstract: Synchronization refers to the problem of inferring the unknown values attached to vertices of a graph where edges are labelled with the ratio of the incident vertices, and labels belong to a group. This paper addresses the synchronization problem on multi-graphs, that are graphs with more than one edge connecting the same pair of nodes. The problem naturally arises when multiple measures are available to model the relationship between two vertices. This happens when different sensors measure the same quantity, or when the original graph is partitioned into sub-graphs that are solved independently. In this case, the relationships among sub-graphs give rise to multi-edges and the problem can be traced back to a multi-graph synchronization. The baseline solution reduces multi-graphs to simple ones by averaging their multi-edges, however this approach falls short because: i) averaging is well defined only for some groups and ii) the resulting estimator is less precise and accurate, as we prove empirically. Specifically, we present MultiSynch, a synchronization algorithm for multi-graphs that is based on a principled constrained eigenvalue optimization. MultiSynch is a general solution that can cope with any linear group and we show to be profitably usable both on synthetic and real problems.", "label_annotations": {"Multi-aspect Summary": {"Context": "Many papers try to address the synchronization problem on multi-graphs, that are graphs with more than one edge connecting the same pair of nodes. The baseline solution reduces multi-graphs to simple ones by averaging their multi-edges, however this approach falls short.", "Key idea": "The author present a synchronization algorithm, named MultiSynch, for multi-graphs that is based on a principled constrained eigenvalue optimization.", "Method": "The authors apply MultiSynch algorithm on any linear group and multi-graphs synthetic problems and real problems.", "Outcome": "The experiment proves that MultiSynch is a general solution that can cope with any linear group and this algorithm may be profitably usable both on synthetic and real problems.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "eb15ebe7-aa58-4a98-8f9e-939967c6359f", "displayed_text": "Title: Synchronization of Group-labelled Multi-graphs.\n\nAbstract: Synchronization refers to the problem of inferring the unknown values attached to vertices of a graph where edges are labelled with the ratio of the incident vertices, and labels belong to a group. This paper addresses the synchronization problem on multi-graphs, that are graphs with more than one edge connecting the same pair of nodes. The problem naturally arises when multiple measures are available to model the relationship between two vertices. This happens when different sensors measure the same quantity, or when the original graph is partitioned into sub-graphs that are solved independently. In this case, the relationships among sub-graphs give rise to multi-edges and the problem can be traced back to a multi-graph synchronization. The baseline solution reduces multi-graphs to simple ones by averaging their multi-edges, however this approach falls short because: i) averaging is well defined only for some groups and ii) the resulting estimator is less precise and accurate, as we prove empirically. Specifically, we present MultiSynch, a synchronization algorithm for multi-graphs that is based on a principled constrained eigenvalue optimization. MultiSynch is a general solution that can cope with any linear group and we show to be profitably usable both on synthetic and real problems.", "label_annotations": {"Multi-aspect Summary": {"Context": "Synchronization refers to the problem of inferring the unknown values attached to vertices of a graph where edges are labelled with the ratio of the incident vertices, and labels belong to a group. ", "Key idea": "Authors show that baseline solution reduces multi-graphs to simple ones by averaging their multi-edges, however this approach falls short because: i) averaging is well defined only for some groups and ii) the resulting estimator is less precise and accurate.\r\nAuthors present MultiSynch, a synchronization algorithm for multi-graphs that is based on a principled constrained eigenvalue optimization.\r\n", "Method": "Authors conduct experiments on synthetic and real problems.", "Outcome": "Authors argue that MultiSynch is a general solution that can cope with any linear group and is show to be profitably usable both on synthetic and real problems.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "f23bdd28-b2d6-4a42-a56c-c9774f6451b5", "displayed_text": "Title: Learning Algorithm for Tracking Hypersonic Targets in Near Space\n\nAbstract: With the development of hypersonic vehicles in near space such as X-51A, HTV-2 and so on, tracking for them is becoming a new task and hotspot. In this paper, a learning tracking algorithm is introduced for hypersonic targets, especially for the sliding jump maneuver. Firstly the algorithm uses the Sine model, which makes the tracking model more close to the particular maneuver, next two Sine models different in angular velocity are used into IMM algorithm, and it learns the target tracking error characteristics to adjust the sampling rate adaptively. The algorithm is compared with the single accurate model algorithm and general IMM algorithms with fixed sampling rate. Through simulation experiments it is proved that the algorithm in this paper can improve the tracking accuracy effectively.", "label_annotations": {"Multi-aspect Summary": {"Context": "Tracking for hypersonic vehicles in near space is becoming a new task and hotspot.", "Key idea": "The authors introduce a learning tracking algorithm for hypersonic targets, especially for the sliding jump maneuver.", "Method": "The authors compare the proposed algorithm with the single accurate model algorithm and general IMM algorithms with fixed sampling rate on simulation experiments.", "Outcome": "Through simulation experiments it is proved that the algorithm in this paper can improve the tracking accuracy effectively.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 20s "}} +{"id": "f23bdd28-b2d6-4a42-a56c-c9774f6451b5", "displayed_text": "Title: Learning Algorithm for Tracking Hypersonic Targets in Near Space\n\nAbstract: With the development of hypersonic vehicles in near space such as X-51A, HTV-2 and so on, tracking for them is becoming a new task and hotspot. In this paper, a learning tracking algorithm is introduced for hypersonic targets, especially for the sliding jump maneuver. Firstly the algorithm uses the Sine model, which makes the tracking model more close to the particular maneuver, next two Sine models different in angular velocity are used into IMM algorithm, and it learns the target tracking error characteristics to adjust the sampling rate adaptively. The algorithm is compared with the single accurate model algorithm and general IMM algorithms with fixed sampling rate. Through simulation experiments it is proved that the algorithm in this paper can improve the tracking accuracy effectively.", "label_annotations": {"Multi-aspect Summary": {"Context": "Tracking hypersonic vehicles in near space such as X-51A, HTV-2 is becoming a new task and hotspot.", "Key idea": "The authors introduce a learning tracking algorithm for hypersonic targets, especially for the sliding jump maneuver. ", "Method": "The authors conduct simulation experiments and compare new algorithm with the single accurate model algorithm and general IMM algorithms with fixed sampling rate.", "Outcome": "Simulation experiments prove that the algorithm in this paper can improve the tracking accuracy effectively.\r\n", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 13s "}} +{"id": "f2c9f8c3-f9d1-4cae-b7ae-a919ada1daaf", "displayed_text": "Title: Targeted Mismatch Adversarial Attack: Query With a Flower to Retrieve the Tower\n\nAbstract: Access to online visual search engines implies sharing of private user content -- the query images. We introduce the concept of targeted mismatch attack for deep learning based retrieval systems to generate an adversarial image to conceal the query image. The generated image looks nothing like the user intended query, but leads to identical or very similar retrieval results. Transferring attacks to fully unseen networks is challenging. We show successful attacks to partially unknown systems, by designing various loss functions for the adversarial image construction. These include loss functions, for example, for unknown global pooling operation or unknown input resolution by the retrieval system. We evaluate the attacks on standard retrieval benchmarks and compare the results retrieved with the original and adversarial image.", "label_annotations": {"Multi-aspect Summary": {"Context": "Access to online visual search engines implies sharing of private user content -- the query images.", "Key idea": "The authors introduce the concept of targeted mismatch attack for deep learning based retrieval systems to generate an adversarial image to conceal the query image.", "Method": "The authors evaluate the attacks on standard retrieval benchmarks and compare the results retrieved with the original and adversarial image.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 6m 33s "}} +{"id": "f2c9f8c3-f9d1-4cae-b7ae-a919ada1daaf", "displayed_text": "Title: Targeted Mismatch Adversarial Attack: Query With a Flower to Retrieve the Tower\n\nAbstract: Access to online visual search engines implies sharing of private user content -- the query images. We introduce the concept of targeted mismatch attack for deep learning based retrieval systems to generate an adversarial image to conceal the query image. The generated image looks nothing like the user intended query, but leads to identical or very similar retrieval results. Transferring attacks to fully unseen networks is challenging. We show successful attacks to partially unknown systems, by designing various loss functions for the adversarial image construction. These include loss functions, for example, for unknown global pooling operation or unknown input resolution by the retrieval system. We evaluate the attacks on standard retrieval benchmarks and compare the results retrieved with the original and adversarial image.", "label_annotations": {"Multi-aspect Summary": {"Context": "Transferring attacks to fully unseen networks is challenging. ", "Key idea": "The authors introduce the concept of targeted mismatch attack for deep learning based retrieval systems to generate an adversarial image to conceal the query image and try to design various loss functions for the adversarial image construction.", "Method": "The authors evaluate the attacks on standard retrieval benchmarks and compare the results retrieved with the original and adversarial image.\r\n", "Outcome": "After designing various loss functions for the adversarial image construction, attacks to partially unknown systems are successful.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 43s "}} +{"id": "f32d6bc3-d75e-4e84-8bfa-3c83578281dc", "displayed_text": "Title: Monitoring and evaluation of information systems via transaction log analysis\n\nAbstract: Transaction log analysis represents a powerful methodology which allows examination of both user commands and system responses when conducting an online information search. Machine-readable transaction log tapes from online catalogs are obtained and subsequently analyzed using stochastic pattern developments within parsed user sessions, mathematical models utilizing Markov chain analysis and the development of state transition probability matrices, which illustrate the probability of proceeding from one user or system state to another state. The objective of monitoring information systems and using transaction log analysis is to discover the extent to which systems are used and to determine the actual user patterns when conducting an information search. This in turn can aid in the evaluation of such systems and assist in the improvement of existing and future systems. Such analysis can assist in system design, while the predictive power of the methodology may allow real-time aids to be developed.As examples of the use of the methodology, patron use and system response patterns from several online public catalogs have been obtained by transaction log tapes. This paper presents an overview of the methodology, results obtained, and efforts that are being conducted within OCLCu0027s Office of Research.", "label_annotations": {"Multi-aspect Summary": {"Context": "Transaction log analysis is powerful in examining both user commands and system responses when conducting an online information search. Machine-readable transaction log tapes from online catalogs are obtained and subsequently analyzed using various methods.", "Key idea": "The author proposed monitoring information systems and using transaction log analysis to discover the extent to which systems are used and to determine the actual user patterns when conducting an information search. This in turn can aid in the evaluation of such systems and assist in the improvement of existing and future systems.", "Method": "The author uses the proposed methodology to obtain patron use and system response patterns from several online public catalogs.", "Outcome": "N/A", "Future Impact": "Such analysis can assist in system design, while the predictive power of the methodology may allow real-time aids to be developed."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 34s "}} +{"id": "f32d6bc3-d75e-4e84-8bfa-3c83578281dc", "displayed_text": "Title: Monitoring and evaluation of information systems via transaction log analysis\n\nAbstract: Transaction log analysis represents a powerful methodology which allows examination of both user commands and system responses when conducting an online information search. Machine-readable transaction log tapes from online catalogs are obtained and subsequently analyzed using stochastic pattern developments within parsed user sessions, mathematical models utilizing Markov chain analysis and the development of state transition probability matrices, which illustrate the probability of proceeding from one user or system state to another state. The objective of monitoring information systems and using transaction log analysis is to discover the extent to which systems are used and to determine the actual user patterns when conducting an information search. This in turn can aid in the evaluation of such systems and assist in the improvement of existing and future systems. Such analysis can assist in system design, while the predictive power of the methodology may allow real-time aids to be developed.As examples of the use of the methodology, patron use and system response patterns from several online public catalogs have been obtained by transaction log tapes. This paper presents an overview of the methodology, results obtained, and efforts that are being conducted within OCLCu0027s Office of Research.", "label_annotations": {"Multi-aspect Summary": {"Context": "Transaction log analysis represents a powerful methodology which allows examination of both user commands and system responses when conducting an online information search.", "Key idea": "The authors propose a way that obtain machine-readable transaction log tapes from online catalogs and subsequently analyze using stochastic pattern developments within parsed user sessions, mathematical models utilizing Markov chain analysis and the development of state transition probability matrices.", "Method": "The authors apply this way that obtain machine-readable transaction log tapes from online catalogs on several online public catalogs.", "Outcome": "The probability of proceeding from one user or system state to another state has been proved.\r\nUsing this methodology, patron use and system response patterns from several online public catalogs can be obtained by transaction log tapes.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 8m 19s "}} +{"id": "f32e53d5-c7f4-407e-a5cb-26fef230b5fd", "displayed_text": "Title: Dynamic Bayesian Logistic Matrix Factorization for Recommendation with Implicit Feedback\n\nAbstract: Matrix factorization has been widely adopted for recommendation by learning latent embeddings of users and items from observed user-item interaction data. However, previous methods usually assume the learned embeddings are static or homogeneously evolving with the same diffusion rate. This is not valid in most scenarios, where usersu0027 preferences and item attributes heterogeneously drift over time. To remedy this issue, we have proposed a novel dynamic matrix factorization model, named Dynamic Bayesian Logistic Matrix Factorization (DBLMF), which aims to learn heterogeneous user and item embeddings that are drifting with inconsistent diffusion rates. More specifically, DBLMF extends logistic matrix factorization to model the probability a user would like to interact with an item at a given timestamp, and a diffusion process to connect latent embeddings over time. In addition, an efficient Bayesian inference algorithm has also been proposed to make DBLMF scalable on large datasets. The effectiveness of the proposed method has been demonstrated by extensive experiments on real datasets, compared with the state-of-the-art methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Matrix factorization has been widely adopted for recommendation by learning latent embeddings of users and items from observed user-item interaction data. Previous methods usually wrongly assume the learned embeddings are static or homogeneously evolving with the same diffusion rate, while users' preferences and item attributes heterogeneously drift over time.\r\n", "Key idea": "The author proposes a novel dynamic matrix factorization model that learns heterogeneous user and item embeddings that are drifting with inconsistent diffusion rates by extending logistic matrix factorization to model the probability a user would like to interact with an item at a given timestamp, and a diffusion process to connect latent embeddings over time. An efficient Bayesian inference algorithm has also been proposed to make the model scalable on large datasets.", "Method": "The author compares the effectiveness of the proposed method with the state-of-the-art methods on extensive experiments on real datasets.", "Outcome": "The effectiveness of the proposed method has been demonstrated by extensive experiments on real datasets, compared with the state-of-the-art methods.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 7m 18s "}} +{"id": "f32e53d5-c7f4-407e-a5cb-26fef230b5fd", "displayed_text": "Title: Dynamic Bayesian Logistic Matrix Factorization for Recommendation with Implicit Feedback\n\nAbstract: Matrix factorization has been widely adopted for recommendation by learning latent embeddings of users and items from observed user-item interaction data. However, previous methods usually assume the learned embeddings are static or homogeneously evolving with the same diffusion rate. This is not valid in most scenarios, where usersu0027 preferences and item attributes heterogeneously drift over time. To remedy this issue, we have proposed a novel dynamic matrix factorization model, named Dynamic Bayesian Logistic Matrix Factorization (DBLMF), which aims to learn heterogeneous user and item embeddings that are drifting with inconsistent diffusion rates. More specifically, DBLMF extends logistic matrix factorization to model the probability a user would like to interact with an item at a given timestamp, and a diffusion process to connect latent embeddings over time. In addition, an efficient Bayesian inference algorithm has also been proposed to make DBLMF scalable on large datasets. The effectiveness of the proposed method has been demonstrated by extensive experiments on real datasets, compared with the state-of-the-art methods.", "label_annotations": {"Multi-aspect Summary": {"Context": "Since users preferences and item attributes heterogeneously drift over time, previous methods, which usually assume the learned embeddings are static or homogeneously evolving with the same diffusion rate, is not valid in most scenarios.", "Key idea": "The authors proposed a novel dynamic matrix factorization model, named Dynamic Bayesian Logistic Matrix Factorization (DBLMF), which aims to learn heterogeneous user and item embeddings that are drifting with inconsistent diffusion rates. And the authors proposed an efficient Bayesian inference algorithm.", "Method": "The authors conduct extensive experiments on real datasets to evaluate effectiveness of Dynamic Bayesian Logistic Matrix Factorization (DBLMF) and other state-of-the-art methods.", "Outcome": "With an efficient Bayesian inference algorithm, Dynamic Bayesian Logistic Matrix Factorization (DBLMF) method is scalable on large dataset. Compared with other state-of-the-art methods, DBLMF is more effective on real datasets.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 1s "}} +{"id": "f3cef657-1887-42a7-9e11-8fdf54f8fa90", "displayed_text": "Title: Kernel Continual Learning\n\nAbstract: This paper introduces kernel continual learning, a simple but effective variant of continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting. We deploy an episodic memory unit that stores a subset of samples for each task to learn task-specific classifiers based on kernel ridge regression. This does not require memory replay and systematically avoids task interference in the classifiers. We further introduce variational random features to learn a data-driven kernel for each task. To do so, we formulate kernel continual learning as a variational inference problem, where a random Fourier basis is incorporated as the latent variable. The variational posterior distribution over the random Fourier basis is inferred from the coreset of each task. In this way, we are able to generate more informative kernels specific to each task, and, more importantly, the coreset size can be reduced to achieve more compact memory, resulting in more efficient continual learning based on episodic memory. Extensive evaluation on four benchmarks demonstrates the effectiveness and promise of kernels for continual learning.", "label_annotations": {"Multi-aspect Summary": {"Context": "Catastrophic forgetting may be a significant challenge in continual learning.", "Key idea": "The authors introduce a simple but effective variant of continual learning named kernel continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting. This method not requires memory replay and systematically avoids task interference in the classifiers.", "Method": "The authors apply kernel continual learning on different tasks and evaluate results on four benchmarks.", "Outcome": "Using kernel continual learning, more informative kernels specific to each task can be generated, and the coreset size can be reduced to achieve more compact memory, resulting in more efficient continual learning based on episodic memory. Kernel continual learning demonstrates the effectiveness and promise of kernels for continual learning on four benchmarks.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "f3cef657-1887-42a7-9e11-8fdf54f8fa90", "displayed_text": "Title: Kernel Continual Learning\n\nAbstract: This paper introduces kernel continual learning, a simple but effective variant of continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting. We deploy an episodic memory unit that stores a subset of samples for each task to learn task-specific classifiers based on kernel ridge regression. This does not require memory replay and systematically avoids task interference in the classifiers. We further introduce variational random features to learn a data-driven kernel for each task. To do so, we formulate kernel continual learning as a variational inference problem, where a random Fourier basis is incorporated as the latent variable. The variational posterior distribution over the random Fourier basis is inferred from the coreset of each task. In this way, we are able to generate more informative kernels specific to each task, and, more importantly, the coreset size can be reduced to achieve more compact memory, resulting in more efficient continual learning based on episodic memory. Extensive evaluation on four benchmarks demonstrates the effectiveness and promise of kernels for continual learning.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors introduce kernel continual learning, a simple but effective variant of continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting. They further introduce variational random features to learn a data-driven kernel for each task.", "Method": "The authors evaluated their proposed methods on four benchmarks.", "Outcome": "Extensive evaluation on four benchmarks demonstrates the effectiveness and promise of kernels for continual learning.", "Future Impact": "The proposed method allows generating more informative kernels specific to each task, and, more importantly, the coreset size can be reduced to achieve more compact memory, resulting in more efficient continual learning based on episodic memory."}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 53s "}} +{"id": "f41bfd75-a6b0-44d7-a842-b117797ebd7e", "displayed_text": "Title: Studying Feature Generation from Various Data Representations for Answer Extraction\n\nAbstract: In this paper, we study how to generate features from various data representations, such as surface texts and parse trees, for answer extraction. Besides the features generated from the surface texts, we mainly discuss the feature generation in the parse trees. We propose and compare three methods, including feature vector, string kernel and tree kernel, to represent the syntactic features in Support Vector Machines. The experiment on the TREC question answering task shows that the features generated from the more structured data representations significantly improve the performance based on the features generated from the surface texts. Furthermore, the contribution of the individual feature will be discussed in detail.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "Authors propose and compare three methods, including feature vector, string kernel and tree kernel, to represent the syntactic features in Support Vector Machines.", "Method": "Authors conduct experiments on the TREC question answering task.", "Outcome": "The experiment on the TREC question answering task shows that the features generated from the more structured data representations significantly improve the performance based on the features generated from the surface texts. Authors also discuss the contribution of the individual feature.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 26s "}} +{"id": "f41bfd75-a6b0-44d7-a842-b117797ebd7e", "displayed_text": "Title: Studying Feature Generation from Various Data Representations for Answer Extraction\n\nAbstract: In this paper, we study how to generate features from various data representations, such as surface texts and parse trees, for answer extraction. Besides the features generated from the surface texts, we mainly discuss the feature generation in the parse trees. We propose and compare three methods, including feature vector, string kernel and tree kernel, to represent the syntactic features in Support Vector Machines. The experiment on the TREC question answering task shows that the features generated from the more structured data representations significantly improve the performance based on the features generated from the surface texts. Furthermore, the contribution of the individual feature will be discussed in detail.", "label_annotations": {"Multi-aspect Summary": {"Context": "N/A", "Key idea": "The authors study how to generate features from various data representations, such as surface texts and parse trees, for answer extraction.\r\nThe authors propose and compare three methods, including feature vector, string kernel and tree kernel, to represent the syntactic features in Support Vector Machines. ", "Method": "The authors compares the performance of three feature representation methods (feature vector, string kernel, tree kernel) for syntactic features derived from parse trees, and evaluates their effectiveness in a Support Vector Machine model on the TREC question answering task.\r\n\r\n", "Outcome": "The experiment result of the TREC question answering task demonstrates that features generated from the more structured data representations significantly improve the performance based on the features generated from the surface texts.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 3s "}} +{"id": "f4f1bbaf-c1a2-44d5-8305-27235fa69d62", "displayed_text": "Title: Modeling and Mining Feature-Rich Networks\n\nAbstract: In the field of web mining and web science, as well as data science and data mining there has been a lot of interest in the analysis of (social) networks. With the growing complexity of heterogeneous data, feature-rich networks have emerged as a powerful modeling approach: They capture data and knowledge at different scales from multiple heterogeneous data sources, and allow the mining and analysis from different perspectives. The challenge is to devise novel algorithms and tools for the analysis of such networks.\n\nThis tutorial provides a unified perspective on feature-rich networks, focusing on different modeling approaches, in particular multiplex and attributed networks. It outlines important principles, methods, tools and future research directions in this emerging field.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "In the field of web mining and web science, as well as data science and data mining there has been a lot of interest in the analysis of (social) networks. With the growing complexity of heterogeneous data, feature-rich networks have emerged as a powerful modeling approach, but the challenge is to devise novel algorithms and tools for the analysis of such networks.", "Key idea": "The authors provide a unified perspective on feature-rich networks, focusing on different modeling approaches, in particular multiplex and attributed networks. They also outline important principles, methods, tools and future research directions in this emerging field.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 59s "}} +{"id": "f4f1bbaf-c1a2-44d5-8305-27235fa69d62", "displayed_text": "Title: Modeling and Mining Feature-Rich Networks\n\nAbstract: In the field of web mining and web science, as well as data science and data mining there has been a lot of interest in the analysis of (social) networks. With the growing complexity of heterogeneous data, feature-rich networks have emerged as a powerful modeling approach: They capture data and knowledge at different scales from multiple heterogeneous data sources, and allow the mining and analysis from different perspectives. The challenge is to devise novel algorithms and tools for the analysis of such networks.\n\nThis tutorial provides a unified perspective on feature-rich networks, focusing on different modeling approaches, in particular multiplex and attributed networks. It outlines important principles, methods, tools and future research directions in this emerging field.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "It's challenging to devise novel algorithms and tools for the analysis of feature-rich networks.", "Key idea": "This tutorial provides a unified perspective on feature-rich networks, focusing on different modeling approaches, in particular multiplex and attributed networks.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 16s "}} +{"id": "f64fdfde-7e93-411b-865a-1e29d71c95b2", "displayed_text": "Title: Scalable training of hierarchical topic models\n\nAbstract: Large-scale topic models serve as basic tools for feature extraction and dimensionality reduction in many practical applications. As a natural extension of flat topic models, hierarchical topic models (HTMs) are able to learn topics of different levels of abstraction, which lead to deeper understanding and better generalization than their flat counterparts. However, existing scalable systems for flat topic models cannot handle HTMs, due to their complicated data structures such as trees and concurrent dynamically growing matrices, as well as their susceptibility to local optima.\r\n\r\nIn this paper, we study the hierarchical latent Dirichlet allocation (hLDA) model which is a powerful nonparametric Bayesian HTM. We propose an efficient partially collapsed Gibbs sampling algorithm for hLDA, as well as an initialization strategy to deal with local optima introduced by tree-structured models. We also identify new system challenges in building scalable systems for HTMs, and propose efficient data layout for vectorizing HTM as well as distributed data structures including dynamic matrices and trees. Empirical studies show that our system is 87 times more efficient than the previous open-source implementation for hLDA, and can scale to thousands of CPU cores. We demonstrate our scalability on a 131-million-document corpus with 28 billion tokens, which is 4--5 orders of magnitude larger than previously used corpus. Our distributed implementation can extract 1,722 topics from the corpus with 50 machines in just 7 hours.", "label_annotations": {"Multi-aspect Summary": {"Context": "Large-scale topic models serve as basic tools for feature extraction and dimensionality reduction and hierarchical topic models (HTMs) are able to learn topics of different levels of abstraction. However, existing scalable systems for flat topic models cannot handle HTMs, due to their complicated data structures.", "Key idea": "The authors propose an efficient partially collapsed Gibbs sampling algorithm for hLDA, as well as an initialization strategy to deal with local optima introduced by tree-structured models. They further identify new system challenges in building scalable systems for HTMs, and propose efficient data layout.", "Method": "The authors evaluate the proposed method on a 131-million-document corpus with 28 billion tokens, which is 4--5 orders of magnitude larger than previously used corpus.", "Outcome": "Empirical studies show that the proposed system is 87 times more efficient than the previous open-source implementation for hLDA, and can scale to thousands of CPU cores. The distributed implementation is also scalable on a 131-million-document corpus with 28 billion tokens, extracting 1,722 topics from the corpus with 50 machines in just 7 hours.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 3m 0s "}} +{"id": "f64fdfde-7e93-411b-865a-1e29d71c95b2", "displayed_text": "Title: Scalable training of hierarchical topic models\n\nAbstract: Large-scale topic models serve as basic tools for feature extraction and dimensionality reduction in many practical applications. As a natural extension of flat topic models, hierarchical topic models (HTMs) are able to learn topics of different levels of abstraction, which lead to deeper understanding and better generalization than their flat counterparts. However, existing scalable systems for flat topic models cannot handle HTMs, due to their complicated data structures such as trees and concurrent dynamically growing matrices, as well as their susceptibility to local optima.\r\n\r\nIn this paper, we study the hierarchical latent Dirichlet allocation (hLDA) model which is a powerful nonparametric Bayesian HTM. We propose an efficient partially collapsed Gibbs sampling algorithm for hLDA, as well as an initialization strategy to deal with local optima introduced by tree-structured models. We also identify new system challenges in building scalable systems for HTMs, and propose efficient data layout for vectorizing HTM as well as distributed data structures including dynamic matrices and trees. Empirical studies show that our system is 87 times more efficient than the previous open-source implementation for hLDA, and can scale to thousands of CPU cores. We demonstrate our scalability on a 131-million-document corpus with 28 billion tokens, which is 4--5 orders of magnitude larger than previously used corpus. Our distributed implementation can extract 1,722 topics from the corpus with 50 machines in just 7 hours.", "label_annotations": {"Multi-aspect Summary": {"Context": "Large-scale topic models serve as basic tools for feature extraction and dimensionality reduction in many practical applications. As a natural extension of flat topic models, hierarchical topic models (HTMs) are able to learn topics of different levels of abstraction, which lead to deeper understanding and better generalization than their flat counterparts. However, existing scalable systems for flat topic models cannot handle HTMs, due to their complicated data structures such as trees and concurrent dynamically growing matrices, as well as their susceptibility to local optima.", "Key idea": "Authors propose an efficient partially collapsed Gibbs sampling algorithm for hLDA, as well as an initialization strategy to deal with local optima introduced by tree-structured models. Authors also identify new system challenges in building scalable systems for HTMs, and propose efficient data layout for vectorizing HTM as well as distributed data structures including dynamic matrices and trees.", "Method": "Authors setup experiments to compare the proposed method with previous open-source implementation for hierarchical latent Dirichlet allocation (hLDA).", "Outcome": "Empirical studies show that proposed system is 87 times more efficient than the previous open-source implementation for hLDA, and can scale to thousands of CPU cores. Authors demonstrate the scalability on a 131-million-document corpus with 28 billion tokens, which is 4--5 orders of magnitude larger than previously used corpus.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 48s "}} +{"id": "f82f2e26-5437-4397-b781-50668ce5851b", "displayed_text": "Title: Discriminative Alignment Training without Annotated Data for Machine Translation\n\nAbstract: In present Statistical Machine Translation (SMT) systems, alignment is trained in a previous stage as the translation model. Consequently, alignment model parameters are not tuned in function of the translation task, but only indirectly. In this paper, we propose a novel framework for discriminative training of alignment models with automated translation metrics as maximization criterion. In this approach, alignments are optimized for the translation task. In addition, no link labels at the word level are needed. This framework is evaluated in terms of automatic translation evaluation metrics, and an improvement of translation quality is observed.", "label_annotations": {"Multi-aspect Summary": {"Context": "In present Statistical Machine Translation (SMT) systems, alignment is trained in a previous stage as the translation model. Consequently, alignment model parameters are not tuned in function of the translation task, but only indirectly.", "Key idea": "The authors propose a novel framework for discriminative training of alignment models with automated translation metrics as maximization criterion.", "Method": "The authors evaluate the proposed framework in terms of automatic translation evaluation metrics", "Outcome": "The evaluation of the proposed framework in terms of automatic translation evaluation metrics shows an improvement of translation quality is observed.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 1m 59s "}} +{"id": "f82f2e26-5437-4397-b781-50668ce5851b", "displayed_text": "Title: Discriminative Alignment Training without Annotated Data for Machine Translation\n\nAbstract: In present Statistical Machine Translation (SMT) systems, alignment is trained in a previous stage as the translation model. Consequently, alignment model parameters are not tuned in function of the translation task, but only indirectly. In this paper, we propose a novel framework for discriminative training of alignment models with automated translation metrics as maximization criterion. In this approach, alignments are optimized for the translation task. In addition, no link labels at the word level are needed. This framework is evaluated in terms of automatic translation evaluation metrics, and an improvement of translation quality is observed.", "label_annotations": {"Multi-aspect Summary": {"Context": "Since in present Statistical Machine Translation (SMT) systems, alignment is trained in a previous stage as the translation model, so, alignment model parameters are not tuned in function of the translation task, but only indirectly.", "Key idea": "The authors propose a novel framework for discriminative training of alignment models with automated translation metrics as maximization criterion. ", "Method": "The authors evaluate novel framework in terms of automatic translation evaluation metrics.", "Outcome": "This method optimize alignments in translation task and no link labels at the word level are needed.\r\nAn improvement of translation quality is observed when evaluating novel framework in terms of automatic translation evaluation metrics.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 35s "}} +{"id": "fb538ce2-abf4-4bd8-b35c-1bfe3ab9f48e", "displayed_text": "Title: JU_NLP at SemEval-2016 Task 11: Identifying Complex Words in a Sentence.\n\nAbstract: The complex word identification task refers to the process of identifying difficult words in a sentence from the perspective of readers belonging to a specific target audience. This task has immense importance in the field of lexical simplification. Lexical simplification helps in improving the readability of texts consisting of challenging words. As a participant of the SemEval-2016: Task 11 shared task, we developed two systems using various lexical and semantic features to identify complex words, one using Naive Bayes and another based on Random Forest Classifiers. The Naive Bayes classifier based system achieves the maximum G-score of 76.7% after incorporating rule based post-processing techniques.", "label_annotations": {"Multi-aspect Summary": {"Context": "The complex word identification task refers to the process of identifying difficult words in a sentence from the perspective of readers belonging to a specific target audience. Lexical simplification helps in improving the readability of texts consisting of challenging words.", "Key idea": "The author develops two systems using various lexical and semantic features to identify complex words, one using Naive Bayes and another based on Random Forest Classifiers.", "Method": "The author tests the perfromance Naive Bayes classifier based system by incorporating rule based post-processing techniques and using G-score as an indicator.", "Outcome": "The Naive Bayes classifier based system achieves the maximum G-score of 76.7% after incorporating rule based post-processing techniques.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 26s "}} +{"id": "fb538ce2-abf4-4bd8-b35c-1bfe3ab9f48e", "displayed_text": "Title: JU_NLP at SemEval-2016 Task 11: Identifying Complex Words in a Sentence.\n\nAbstract: The complex word identification task refers to the process of identifying difficult words in a sentence from the perspective of readers belonging to a specific target audience. This task has immense importance in the field of lexical simplification. Lexical simplification helps in improving the readability of texts consisting of challenging words. As a participant of the SemEval-2016: Task 11 shared task, we developed two systems using various lexical and semantic features to identify complex words, one using Naive Bayes and another based on Random Forest Classifiers. The Naive Bayes classifier based system achieves the maximum G-score of 76.7% after incorporating rule based post-processing techniques.", "label_annotations": {"Multi-aspect Summary": {"Context": "The complex word identification task refers to the process of identifying difficult words in a sentence from the perspective of readers belonging to a specific target audience. Lexical simplification helps in improving the readability of texts consisting of challenging words.", "Key idea": "The authors develope two systems using various lexical and semantic features to identify complex words, one using Naive Bayes and another based on Random Forest Classifiers.", "Method": "N/A", "Outcome": "The Naive Bayes classifier based system achieves the maximum G-score of 76.7% after incorporating rule based post-processing techniques.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 2m 2s "}} +{"id": "feb75e1f-7838-48ca-9a78-cc31b717e5bf", "displayed_text": "Title: Lightning Talk - Think Outside the Dataset: Finding Fraudulent Reviews using Cross-Dataset Analysis\n\nAbstract: Many crowd-sourced review platforms, such as Yelp, TripAdvisor, and Foursquare, have sprung up to provide a shared space for people to write reviews and rate local businesses. With the substantial impact of businesses\u2019 online ratings on their selling\u00a0[2], many businesses add themselves to multiple websites to more easily be discovered. Some might also engage in reputation management, which could range from rewarding their customers for a favorable review, or a complex review campaign, where armies of accounts post reviews to influence a business\u2019 average review score.\n\nMost of previous work use supervised machine learning, and only focus on textual and stylometry features\u00a0[1, 3, 4, 7]. Their obtained ground truth data is not large and comprehensive\u00a0[4, 5, 6, 7, 8, 10]. These works also assume a limited threat model, e.g., an adversary\u2019s activity is assumed to be found near sudden shifts in the data\u00a0[8], or focused on positive campaigns.\n\nWe propose OneReview , a system for finding fraudulent content on a crowd-sourced review site, leveraging correlations with other independent review sites, and the use of textual and contextual features. We assume that an attacker may not be able to exert the same influence over a business\u2019 reputation on several websites, due to increased cost. OneReview focuses on isolating anomalous changes in a business\u2019 reputation across multiple review sites, to locate malicious activity without relying on specific patterns. Our intuition is that a business\u2019s reputation should not be very different in multiple review sites; e.g., if a restaurant changes its chef or manager, then the impact of these changes should appear on reviews across all the websites. OneReview utilizes Change Point Analysis method on the reviews of every business independently on every website, and then uses our proposed Change Point Analyzer to evaluate change-points, detect those that do not match across the websites, and identify them as suspicious. Then, it uses supervised machine learning, utilizing a combination of textual and metadata features to locate fraudulent reviews among the suspicious reviews.\n\nWe evaluated our approach, using data from two reviewing websites, Yelp and TripAdvisor, to find fraudulent activity on Yelp. We obtained Yelp reviews, through the Yelp Data Challenge\u00a0[9], and used our Change Point Analyzer to correlate this with data crawled from TripAdvisor. Since realistic and varied ground truth data is not currently available, we used a combination of our change point analysis and crowd-labeling to create a set of 5,655 labeled reviews. We used k-cross validation (k=5) on our ground truth and obtained 97% (+/- 0.01) accuracy, 91% (+/- 0.03) precision and 90% (+/- 0.06) recall. The model was used on the suspicious reviews, which classified 61,983 reviews, about 8% of all reviews, as fraudulent.\n\nWe further detected fraudulent campaigns that are actively initiated by or targeted toward specific businesses. We identified 3,980 businesses with fraudulent reviews, as well as, 14,910 suspected spam, where at least 40% of their reviews are classified as fraudulent. We also used community detection algorithms to locate several large astroturfing campaigns. These results show the effectiveness of OneReview in detecting fraudulent campaigns.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Many business engaged in reputation management by using armies of accounts post reviews to influence a business\u2019 average review score. Unfortunately, most of previous work use supervised machine learning, and only focus on textual and stylometry features, and ground truth data they gained is not large and comprehensive .", "Key idea": "The authors propose a system for finding fraudulent content on a crowd-sourced review site named OneReview, leveraging correlations with other independent review sites, and the use of textual and contextual features. OneReview focuses on isolating anomalous changes in a business\u2019 reputation across multiple review sites, to locate malicious activity without relying on specific patterns. And it uses supervised machine learning, utilizing a combination of textual and metadata features to locate fraudulent reviews among the suspicious reviews. ", "Method": "The authors evaluate OneReview approach with data from two reviewing websites, Yelp and TripAdvisor, to find fraudulent activity on Yelp. The authors obtain Yelp reviews through the Yelp Data Challenge and use our Change Point Analyzer to correlate this with data crawled from TripAdvisor. And finally use a combination of our change point analysis and crowd-labeling to create a set of 5,655 labeled reviews as evaluation dataset.", "Outcome": "The authors used k-cross validation(k=5) on ground truth and obtained 97% accuracy, 91% precision and 90% recall. Besides, authors use this model to identify 3,980 businesses with fraudulent reviews and 14,910 suspected spam, where at least 40% of their reviews are classified as fraudulent.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 4s "}} +{"id": "feb75e1f-7838-48ca-9a78-cc31b717e5bf", "displayed_text": "Title: Lightning Talk - Think Outside the Dataset: Finding Fraudulent Reviews using Cross-Dataset Analysis\n\nAbstract: Many crowd-sourced review platforms, such as Yelp, TripAdvisor, and Foursquare, have sprung up to provide a shared space for people to write reviews and rate local businesses. With the substantial impact of businesses\u2019 online ratings on their selling\u00a0[2], many businesses add themselves to multiple websites to more easily be discovered. Some might also engage in reputation management, which could range from rewarding their customers for a favorable review, or a complex review campaign, where armies of accounts post reviews to influence a business\u2019 average review score.\n\nMost of previous work use supervised machine learning, and only focus on textual and stylometry features\u00a0[1, 3, 4, 7]. Their obtained ground truth data is not large and comprehensive\u00a0[4, 5, 6, 7, 8, 10]. These works also assume a limited threat model, e.g., an adversary\u2019s activity is assumed to be found near sudden shifts in the data\u00a0[8], or focused on positive campaigns.\n\nWe propose OneReview , a system for finding fraudulent content on a crowd-sourced review site, leveraging correlations with other independent review sites, and the use of textual and contextual features. We assume that an attacker may not be able to exert the same influence over a business\u2019 reputation on several websites, due to increased cost. OneReview focuses on isolating anomalous changes in a business\u2019 reputation across multiple review sites, to locate malicious activity without relying on specific patterns. Our intuition is that a business\u2019s reputation should not be very different in multiple review sites; e.g., if a restaurant changes its chef or manager, then the impact of these changes should appear on reviews across all the websites. OneReview utilizes Change Point Analysis method on the reviews of every business independently on every website, and then uses our proposed Change Point Analyzer to evaluate change-points, detect those that do not match across the websites, and identify them as suspicious. Then, it uses supervised machine learning, utilizing a combination of textual and metadata features to locate fraudulent reviews among the suspicious reviews.\n\nWe evaluated our approach, using data from two reviewing websites, Yelp and TripAdvisor, to find fraudulent activity on Yelp. We obtained Yelp reviews, through the Yelp Data Challenge\u00a0[9], and used our Change Point Analyzer to correlate this with data crawled from TripAdvisor. Since realistic and varied ground truth data is not currently available, we used a combination of our change point analysis and crowd-labeling to create a set of 5,655 labeled reviews. We used k-cross validation (k=5) on our ground truth and obtained 97% (+/- 0.01) accuracy, 91% (+/- 0.03) precision and 90% (+/- 0.06) recall. The model was used on the suspicious reviews, which classified 61,983 reviews, about 8% of all reviews, as fraudulent.\n\nWe further detected fraudulent campaigns that are actively initiated by or targeted toward specific businesses. We identified 3,980 businesses with fraudulent reviews, as well as, 14,910 suspected spam, where at least 40% of their reviews are classified as fraudulent. We also used community detection algorithms to locate several large astroturfing campaigns. These results show the effectiveness of OneReview in detecting fraudulent campaigns.\n\n", "label_annotations": {"Multi-aspect Summary": {"Context": "Many crowd-sourced review platforms, such as Yelp, TripAdvisor, and Foursquare, have sprung up to provide a shared space for people to write reviews and rate local businesses. Some might also engage in reputation management, which could range from rewarding their customers for a favorable review, or a complex review campaign, where armies of accounts post reviews to influence a business\u2019 average review score.", "Key idea": "Authors propose OneReview , a system for finding fraudulent content on a crowd-sourced review site, leveraging correlations with other independent review sites, and the use of textual and contextual features. ", "Method": "OneReview utilizes Change Point Analysis method on the reviews of every business independently on every website, and then uses our proposed Change Point Analyzer to evaluate change-points, detect those that do not match across the websites, and identify them as suspicious. Then, it uses supervised machine learning, utilizing a combination of textual and metadata features to locate fraudulent reviews among the suspicious reviews.", "Outcome": "Authors used k-cross validation (k=5) on our ground truth and obtained 97% (+/- 0.01) accuracy, 91% (+/- 0.03) precision and 90% (+/- 0.06) recall.\r\nOneReview identified 3,980 businesses with fraudulent reviews, as well as, 14,910 suspected spam, where at least 40% of their reviews are classified as fraudulent.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 4m 34s "}} +{"id": "fed7302a-43a7-412e-8ace-d07905e38c3c", "displayed_text": "Title: One-class collaborative filtering with random graphs\n\nAbstract: The bane of one-class collaborative filtering is interpreting and modelling the latent signal from the missing class. In this paper we present a novel Bayesian generative model for implicit collaborative filtering. It forms a core component of the Xbox Live architecture, and unlike previous approaches, delineates the odds of a user disliking an item from simply being unaware of it. The latent signal is treated as an unobserved random graph connecting users with items they might have encountered. We demonstrate how large-scale distributed learning can be achieved through a combination of stochastic gradient descent and mean field variational inference over random graph samples. A fine-grained comparison is done against a state of the art baseline on real world data.", "label_annotations": {"Multi-aspect Summary": {"Context": "The bane of one-class collaborative filtering is interpreting and modelling the latent signal from the missing class. ", "Key idea": "Authors present a novel Bayesian generative model for implicit collaborative filtering. It forms a core component of the Xbox Live architecture, and unlike previous approaches, delineates the odds of a user disliking an item from simply being unaware of it. \r\nAuthors demonstrate how large-scale distributed learning can be achieved through a combination of stochastic gradient descent and mean field variational inference over random graph samples.", "Method": "Authors design fine-grained comparison of the proposed method against a state of the art baseline on real world data..", "Outcome": "The proposed method is better than state of the art baseline on real world data in fine-grained comparison.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 5m 41s "}} +{"id": "fed7302a-43a7-412e-8ace-d07905e38c3c", "displayed_text": "Title: One-class collaborative filtering with random graphs\n\nAbstract: The bane of one-class collaborative filtering is interpreting and modelling the latent signal from the missing class. In this paper we present a novel Bayesian generative model for implicit collaborative filtering. It forms a core component of the Xbox Live architecture, and unlike previous approaches, delineates the odds of a user disliking an item from simply being unaware of it. The latent signal is treated as an unobserved random graph connecting users with items they might have encountered. We demonstrate how large-scale distributed learning can be achieved through a combination of stochastic gradient descent and mean field variational inference over random graph samples. A fine-grained comparison is done against a state of the art baseline on real world data.", "label_annotations": {"Multi-aspect Summary": {"Context": "The bane of one-class collaborative filtering is interpreting and modelling the latent signal from the missing class.", "Key idea": "The authors present a novel Bayesian generative model for implicit collaborative filtering. ", "Method": "The authors demonstrate a fine-grained comparison is done against a state of the art baseline on real world data and novel Bayesian generative model.", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 2s "}} +{"id": "ffd14676-a525-479f-a74e-2c5d3a85c510", "displayed_text": "Title: Shape representation in parallel systems\n\nAbstract: There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units [1 2 3 4]. At the early stages of visual processing, individual units can represent hypotheses about how small local fragments of the visual input should be interpreted, and interactions between units can encode knowledge about the constraints between local interpretations. Higher up in the visual system, the representational issues are more complex. This paper considers the difficulties involved in representing shapes in parallel systems, and suggests ways of overcoming them. In doing so, it provides a mechanism for shape perception and visual attention which allows a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.", "label_annotations": {"Multi-aspect Summary": {"Context": "In parallel systems, visual processing by having individual network units represent hypotheses about how small local fragments of the visual input should be interpreted and encode knowledge about the constraints between local interpretations. But the representational issues are more complex higher up in the visual system.", "Key idea": "The author provides a mechanism for shape perception and visual attention which allows a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 0m 26s "}} +{"id": "ffd14676-a525-479f-a74e-2c5d3a85c510", "displayed_text": "Title: Shape representation in parallel systems\n\nAbstract: There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units [1 2 3 4]. At the early stages of visual processing, individual units can represent hypotheses about how small local fragments of the visual input should be interpreted, and interactions between units can encode knowledge about the constraints between local interpretations. Higher up in the visual system, the representational issues are more complex. This paper considers the difficulties involved in representing shapes in parallel systems, and suggests ways of overcoming them. In doing so, it provides a mechanism for shape perception and visual attention which allows a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.", "label_annotations": {"Multi-aspect Summary": {"Context": "There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units. However, higher up than small local fragments, the representational issues are complex in the visual system.", "Key idea": "The authors suggest ways of overcoming the difficulties involved in representing shapes in parallel systems. They provide a mechanism for shape perception and visual attention.", "Method": "N/A", "Outcome": "The proposed mechanism allows a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.", "Future Impact": "N/A"}}, "span_annotations": {}, "behavioral_data": {"time_string": "Time spent: 0d 0h 7m 10s "}} \ No newline at end of file diff --git a/data/benchmark_0531.jsonl b/data/benchmark_0531.jsonl new file mode 100644 index 0000000..e09ea1b --- /dev/null +++ b/data/benchmark_0531.jsonl @@ -0,0 +1,1020 @@ +{"pid": "3f06487d-85a0-4ba9-b0b9-fe4ea2fe74cb", "context": "The collective spatial keyword query (CoSKQ) selects a set of objects from a database, such that it collectively covers a set of given keywords and has the smallest cost. Existing exact algorithms for CoSKQ have scalability problems and approximate algorithms, while scalable, do not guarantee near-optimal solutions.", "key_idea": "The authors propose a distance owner-driven approach for CoSKQ. They also introduce two new cost measurements: the maximum sum cost (MaxSum-CoSKQ) and the diameter cost (Dia-CoSKQ). The approach features exact and approximate algorithms for both cost measurements.", "method": "The exact and approximate distance owner-driven algorithms were tested on real datasets. The exact algorithm was compared to the existing best algorithm for speed and the approximate algorithm was assessed for its closeness to an optimal solution.", "outcome": "Experimentation on real datasets show that the proposed exact algorithms are scalable and the approximate algorithms return near-to-optimal solutions. The exact algorithm runs faster than the best-known existing algorithm by several orders of magnitude, and the approximate algorithm improves the best-known constant approximation factor from 2 to 1.375.", "future_impact": "The successful implementation of distance owner-driven approaches in tackling the CoSKQ problem can be extended to other queries and keyword search problems, potentially improving scalability and accuracy of results.", "venue": "SIGMOD", "year": 2013, "title": "Collective spatial keyword queries: a distance owner-driven approach"} +{"pid": "5f156a9d91e011d7db223aae", "context": "Given a bipartite graph, the maximum balanced biclique (MBB) problem, which involves finding mutually connected, disjoint sets of equal size with the maximum cardinality, is significant for mining the bipartite graph and has multiple applications.", "key_idea": "The authors propose two exact algorithms for small dense and large sparse bipartite graphs respectively, to solve the MBB problem. Key techniques include transforming a large sparse bipartite graph into a limited number of dense subgraphs and solving MBB on each of these subgraphs and finding the MBB in small dense graphs by making the search converge fast to sufficiently dense bipartite graphs.", "method": "The paper includes extensive experiments on synthetic and real large bipartite graphs to demonstrate the efficiency and effectiveness of the proposed algorithms and techniques.", "outcome": "The results show that the proposed algorithms can find an exact solution to the MBB problem extremely fast in various bipartite graphs, including small dense bipartite graphs common in applications like VLSI design and large sparse bipartite graphs typical for applications like biological data analysis. The algorithm can even discover MBB within a few seconds for bipartite graphs with millions of vertices.", "future_impact": "The proposed techniques for solving the MBB problem could improve how we handle challenges in numerous application areas, including VLSI design and biological data analysis.", "venue": "SIGMOD", "year": 2021, "title": "Efficient Exact Algorithms for Maximum Balanced Biclique Search in Bipartite Graphs"} +{"pid": "614f9694-1041-4091-87f9-6d6c88baa61f", "context": "Label-Constrained Reachability (LCR) queries are core fragments of practical graph query languages and play an important role in graph analytics. However, existing solutions for LCR evaluation do not scale well to large graphs which are becoming increasingly common across various application domains.", "key_idea": "The authors propose the first practical solution for efficient LCR evaluation by leveraging landmark-based indexes for large graphs.", "method": "The authors compare their landmark-based indexes with state-of-the-art LCR indexing techniques through extensive experiments.", "outcome": "The landmark-based indexes proposed by the authors are significantly smaller than state-of-the-art LCR indexing techniques, while supporting faster query evaluation times.", "future_impact": "The authors have made their complete C++ codebase available as open source for further research, potentially inspiring more scalable and efficient solutions in LCR evaluation and graph analytics.", "venue": "SIGMOD", "year": 2017, "title": "Landmark Indexing for Evaluation of Label-Constrained Reachability Queries"} +{"pid": "bd3f0249-c170-40c6-8da0-2a607c5af5d9", "context": "Deep learning models perform well in tasks like language translation and object recognition. However, understanding the high-level logic of these models is unclear. Current research uses statistical methods to identify hidden units that behave similarly to human logic but require considerable manual effort.", "key_idea": "The authors propose DeepBase, a system designed to inspect the behaviors of neural networks through a unified interface. The high-level logic is modeled with user-provided hypothesis functions that annotate the data with high-level labels.", "method": "DeepBase is used to identify individual or groups of units that have strong statistical dependencies with desired hypotheses. It is used to express existing analyses and to reproduce recent studies from the NLP literature.", "outcome": "DeepBase speeds up a standard Python implementation by up to 72x and can reproduce recent studies from the NLP literature.", "future_impact": "This work has the potential to improve transparency, better model design, and faster experimentation in deep learning models by providing a tool to understand the high-level logic more easily.", "venue": "SIGMOD", "year": 2019, "title": "DeepBase: Deep Inspection of Neural Networks"} +{"pid": "1e298cf7-f075-4bd6-ada1-1af55856ba65", "context": "The rise of live campaigns as a new marketing strategy is stymied by challenges in identifying optimal times and locations, which can make or break the success of these campaigns.", "key_idea": "This paper proposes a solution to find suitable locations for arranging live campaigns by using online location-based services and examining various spatio-temporal mining features.", "method": "Authors use a dataset collected from Foursquare in New York City, from which they develop models to predict the expected audience based on spatio-temporal mining features.", "outcome": "The study finds that individual feature-based approach achieves an accuracy of 50.46% and a Support Vector Machine (SVM) regression model achieves an accuracy of 72.6%.", "future_impact": "This work could form a basis for improving the effectiveness of live campaigns by taking into consideration optimal locations and times based on data from location-based services.", "venue": "SIGMOD", "year": 2017, "title": "Finding suitable places for live campaigns using location-based services"} +{"pid": "2b6ea227-fda9-4ec1-804e-6635ef844259", "context": "The past decade has seen a genomics revolution benefiting greatly from information technology tools, assisting in tasks like large-scale sequence assembly, automated gene identification, and efficient database searching. However, the upcoming focus on comprehensive, systematic studies of proteins and other biomolecules, with a particular interest in high-resolution imaging of molecules and cells, is set to generate an even larger volume of data.", "key_idea": "The researchers are developing sophisticated information technologies and methods for collecting and interpreting the increasing volume of biological image data.", "method": "Their work involves the creation of a distributed digital library of biomolecular image data, enabling global access and comprehensive understanding of the data.", "outcome": "The major output of the research will be a unique, fully operational, distributed digital library of biomolecular image data, accessible to researchers worldwide.", "future_impact": "Such searchable databases will optimize understanding and interpretation of the data, contributing to a more complete and integrated understanding of cellular structure, function, and regulation. It will also help understand how different properties change in response to various factors like stress, injury, aging, and disease.", "venue": "SIGMOD", "year": 2004, "title": "A distributed database for bio-molecular images"} +{"pid": "04df172c-f423-441c-9c10-5bdabfe8c5c9", "context": "At the time of the study, distributed processing and database management were major research areas in computer science and there was a need for a practical multiprocessor database machine suitable for large database, high transaction-rate applications.", "key_idea": "The authors describe Jasmin, a multiprocessor database machine implemented as part of a research project. Key features include its configurable performance, use of off-the-shelf parts, and ability to handle distributed databases.", "method": "The authors develop a uniprocessor prototype of Jasmin and plan to develop a multiprocessor version.", "outcome": "A uniprocessor prototype of Jasmin has been completed, demonstrating feasibility.", "future_impact": "The completion of the multiprocessor version of Jasmin, planned for the future, could show the practicality of a multiprocessor database machine for high transaction-rate applications.", "venue": "SIGMOD", "year": 1984, "title": "Overview of the Jasmin database machine"} +{"pid": "5e2c15df-7938-4672-8f0c-33d93e6d864e", "context": "In a federated database environment, different constituents within the federation can use different temporal models or different physical representations of temporal information, causing differences or mismatches.", "key_idea": "This paper introduces a solution to resolve these mismatches in the form of a new concept named a 'Temporal Module'. The temporal module conceals the implementation details of a temporal relation by showing its information through two 'windowing functions'.", "method": "The authors build a calculus-style language for forming queries on temporal modules. The authors then extend temporal modules to reconcile another mismatch type involving different time units used to record temporal data, relying on 'information conversions' provided by each constituent.", "outcome": "Authors manage to define and study time units formally, propose a federated temporal database model, and extend a calculus-style language as its query language.", "future_impact": "The approach of using temporal modules to reconcile mismatches in federated databases can streamline data processing and improve the efficiency of querying in such environments.", "venue": "SIGMOD", "year": 1993, "title": "Temporal modules: an approach toward federated temporal databases"} +{"pid": "a4dc0e5b-a373-41b7-bd03-b7a253fecd53", "context": "Banks are recognizing the importance of better understanding their customers due to growing competition. Currently, customer interactions such as emails and phone calls are not linked with customer and account profiles stored in their data warehouse.", "key_idea": "The authors present LIPTUS, a tool that links customer interactions such as emails and transcribed phone calls with customer and account profiles in a data warehouse to enable more comprehensive customer analytics.", "method": "LIPTUS operates by bridging the gap between structured (customer and account profiles) and unstructured (customer interactions) data using text mining techniques. The functionality and benefits of this tool are demonstrated through specific customer intelligence applications.", "outcome": "LIPTUS is successfully being used extensively in a large bank in India, illustrating the practical application and value of this tool across the banking sector.", "future_impact": "The tool could be of immense value in deriving customer intelligence in other banks and similar environments, paving the way for improved customer understanding and relationship management through enhanced customer analytics.", "venue": "SIGMOD", "year": 2007, "title": "LIPTUS: associating structured and unstructured information in a banking environment"} +{"pid": "6dc075ed-26eb-4359-90b1-ffe24515a6fd", "context": "Database developers use data access APIs like ADO.NET to execute SQL queries, however, there is a significant gap in program analysis and profiling tools for database applications due to their lack of understanding of data access APIs and DBMS.", "key_idea": "The authors developed tools that enhance static program analysis by understanding database APIs to identify security, correctness, and performance issues and to extend profiling infrastructure for correlation of application events with DBMS events.", "method": "The authors built tools that enhance traditional static analysis of programs by leveraging understanding of database APIs and extend the existing DBMS and application profiling infrastructure. The effectiveness of these tools was showcased through analysis, tuning and profiling tasks.", "outcome": "Their tools enable a rich class of analysis, tuning, and profiling tasks that were not possible before.", "future_impact": "This approach could serve as a standard for improving program analysis and profiling for application and database development, leading to early detection of issues and better performance.", "venue": "SIGMOD", "year": 2009, "title": "Bridging the application and DBMS divide using static analysis and dynamic profiling"} +{"pid": "c6352212-de41-40af-aaf4-b70a40961aae", "context": "Materialized views in relational and deductive database systems currently lack efficient algorithms for incremental evaluation in response to changes (insertions, deletions, and updates) to the relations.", "key_idea": "The authors propose two incremental evaluation algorithms - a counting algorithm and the Delete and Rederive algorithm (DRed). The counting algorithm tracks the number of alternative derivations (counts) for each derived tuple in a view, while DRed is designed for incremental maintenance of recursive views where negation and aggregation are permitted.", "method": "The authors demonstrate the counting algorithm through the use of nonrecursive views (with negation and aggregation), showing that the count for a tuple can be computed with little or no additional cost beyond the cost of deriving the tuple. The Delete and Rederive algorithm is presented for incremental maintenance of recursive views, performing by first deleting a superset of the tuples that need to be deleted, and then rederiving some of them.", "outcome": "The paper shows that the proposed algorithms are optimal as the counting algorithm computes exactly those view tuples that are inserted or deleted and only the number of derivations are stored, not the derivations themselves. The DRed algorithm is capable of handling when the view definition itself is altered.", "future_impact": "The algorithms proposed in this paper could potentially improve the efficiency of incremental evaluation of materialized views in response to changes, providing a new methodology for maintaining views incrementally in both relational and deductive database systems.", "venue": "SIGMOD", "year": 1993, "title": "Maintaining views incrementally"} +{"pid": "13644cc7-66e2-482b-b6af-acbb27cd79af", "context": "Relational database tasks, such as index selection, histogram tuning, approximate query processing, and statistics selection, often encounter scalability issues due to the large size of workloads presented as sets of SQL DML statements.", "key_idea": "The paper introduces the novel problem of workload compression, intended to improve the scalability of SQL workload-driven tasks.", "method": "The authors propose a solution to workload compression, broadly applicable to a variety of tasks, while accommodating task-specific knowledge. The solution is implemented and tested in the context of two tasks: index selection and approximate query processing.", "outcome": "The experiments confirm the effectiveness of the proposed workload compression solution in improving the scalability of index selection and approximate query processing tasks.", "future_impact": "The introduced concept of workload compression, along with the proposed solution, has the potential to significantly improve the scalability of a range of SQL workload-driven tasks, thus making them more efficient and effective.", "venue": "SIGMOD", "year": 2002, "title": "Compressing SQL workloads"} +{"pid": "f31dc6ec-babc-4ec2-b816-13708f19f1fc", "context": "Traditional database query optimizers use statistics built over base tables and assume independence between attributes while propagating statistical information through the query plan. This approach may produce large estimation errors, leading to inefficient execution plans.", "key_idea": "This study proposes an extension to a generic optimizer that leverages statistics built on expressions corresponding to intermediate nodes of query plans, as this can improve the quality of the resulting plans.", "method": "The authors proposed approach was implemented in Microsoft SQL Server 2000, and a workload-driven technique was introduced to isolate a small subset of statistics that could offer significant benefits.", "outcome": "Experimental results showed that in some cases, the proposed approach led to significantly better plan quality than when only using base-table statistics.", "future_impact": "The proposed optimization could enhance database query efficiency by exploiting statistics on query expressions, potentially changing how optimizers are developed and used.", "venue": "SIGMOD", "year": 2002, "title": "Exploiting statistics on query expressions for optimization"} +{"pid": "53468f4e-89d3-4601-98c2-c782e77d0f6f", "context": "Domain adaptation techniques come in handy when developing models with data from varying sources, of different qualities, or time ranges, but are often sensitive to the choice of a weight hyperparameter that is expensive to fine-tune.", "key_idea": "The authors propose CrossTrainer, a system that facilitates practical domain adaptation by exploiting loss reweighting, which has shown consistently high model accuracy across different datasets.", "method": "The authors use loss reweighting from [6] and apply optimizations by leveraging unique properties of loss reweighting to build CrossTrainer.", "outcome": "Through the development of CrossTrainer, the authors were able to improve training time compared to naive hyperparameter search, while still maintaining accurate models.", "future_impact": "The CrossTrainer system can provide substantial improvements in the deployment of domain adaptation techniques, significantly reducing training times and potential user challenges in developing models with different data sources.", "venue": "SIGMOD", "year": 2019, "title": "CrossTrainer: Practical Domain Adaptation with Loss Reweighting"} +{"pid": "fefda679-4526-4449-b60c-ec9c3094792a", "context": "The Cypher property graph query language was originally designed and implemented as part of the Neo4j graph database and has been used by several commercial database products and researchers.", "key_idea": "The authors describe Cypher 9, the first version of the language governed by the openCypher Implementers Group, and its usage in industry, providing a formal semantic definition of its core read-query features.", "method": "The authors introduce the language through examples, conduct a comparison of Cypher's features to other property graph query languages, and describe extensions at an advanced stage of development for the upcoming Cypher 10.", "outcome": "The authors provide a detailed understanding of Cypher 9's core read-query features including its variant of the property graph data model and its ASCII Art graph pattern matching mechanism.", "future_impact": "The description of extensions for Cypher 10, including support for graph projections and multiple named graphs, suggests that the language will evolve into a compositional language, likely impacting future data manipulation in graph databases.", "venue": "SIGMOD", "year": 2018, "title": "Cypher: An Evolving Query Language for Property Graphs"} +{"pid": "60d3093491e01182ead65d43", "context": "The integration of multiple GPUs into a single machine and the introduction of higher bandwidth interconnects like NVLink 2.0 has opened up new opportunities for relational query processing on multiple GPUs. Yet, due to unique characteristics of GPUs and the interconnects, existing hash join implementations spend up to 66% of their execution time moving data between GPUs with less than 50% utilization of the bandwidth, which leads to poor scalability.", "key_idea": "The authors propose MG-Join, a scalable partitioned hash join implementation on multiple GPUs of a single machine that employs a novel multi-hop routing for cross-GPU communication that adaptively chooses the efficient route for each data flow to minimize congestion.", "method": "The authors develop and test MG-Join on a DGX-1 machine, measuring its bandwidth utilization, scalability, and overall performance in comparison with state-of-the-art hash join implementations and database systems.", "outcome": "MG-Join is found to significantly reduce communication overhead and attains up to 97% utilization of the bisection bandwidth of the interconnects, outperforming state-of-the-art hash join implementations by up to 2.5x and enhancing the overall performance of TPC-H queries by up to 4.5x over the multi-GPU version of the Omnisci GPU database.", "future_impact": "MG-Join could revolutionize the field of relational query processing on multiple GPUs, offering superior performance, scalability and utilization of bandwidth.", "venue": "SIGMOD", "year": 2021, "title": "MG-Join: A Scalable Join for Massively Parallel Multi-GPU Architectures"} +{"pid": "605b1f3691e0119ebe7e5d9e", "context": "Distributed systems that manage graph-structured data solve a graph partitioning problem to minimize communication overhead and query run-time. Problematically, optimal graph partitioning is not only NP-hard but also incurs considerable memory overhead. Streaming partitioning algorithms are commonly used to reduce memory overhead but may lack competitiveness with in-memory partitioners.", "key_idea": "This paper proposes a new system, Hybrid Edge Partitioner (HEP), which partitions graphs that only partly fit into memory while maintaining high partitioning quality. This is done by separating the graph's edge set into two subsets: one subset is partitioned using NE++, a new efficient in-memory algorithm, while the other subset uses a streaming approach.", "method": "The authors evaluated the performance of HEP on large real-world graphs, comparing it against both in-memory partitioning and streaming partitioning.", "outcome": "HEP was found to outperform both in-memory and streaming partitioning in many cases. Furthermore, using HEP led to a significant speedup of distributed graph processing jobs on Spark/GraphX compared to state-of-the-art partitioning algorithms.", "future_impact": "HEP is an attractive alternative to existing solutions, especially for systems that cannot fine-tune their memory overheads. It may also contribute to enhancing the efficiency of distributed graph processing.", "venue": "SIGMOD", "year": 2021, "title": "Hybrid Edge Partitioner: Partitioning Large Power-Law Graphs under Memory Constraints"} +{"pid": "30345aed-0eb0-4f4a-9bfb-2bdef8e21881", "context": "Existing metrics used to measure the influence degree of researchers may not be adequately capturing the connections between communities that researchers establish.", "key_idea": "The authors propose the 3c-index, a new metric designed to measure the influence of researchers by evaluating the links they create between different communities.", "method": "The performance of the proposed 3c-index is evaluated against well-known existing metrics for measuring researcher influence.", "outcome": "Results show that the 3c-index outperforms most existing metrics and can serve as a supplementary metric for assessing researcher productivity.", "future_impact": "The 3c-index could potentially be used as a supplementary metric to assess researchers' productivity in the future.", "venue": "SIGMOD", "year": 2016, "title": "Research Contribution as a Measure of Influence"} +{"pid": "0c34f7ed-f534-427c-aab1-0b9822a7edfc", "context": "Regular Expression (RE) matching has important applications in areas such as XML content distribution and network security. Both Deterministic Finite Automata (DFA) and Non-deterministic Finite Automata (NFA) have been used, but these systems have struggled to efficiently scale to hundreds of REs.", "key_idea": "The authors propose a high-performance RE matching system that combines the processing efficiency of DFA with the space efficiency of NFA. They exploit the fact that a bulk of the DFA transitions are concentrated around a few DFA states to cache only the frequent core of each DFA in memory, rather than the entire DFA.", "method": "This proposal is tested with real-life RE data on data streams. Additional techniques, such as clustering REs that cause exponential increases in state number into separate groups, and a clever combination of RE grouping, static and dynamic caching, are applied to enhance performance.", "outcome": "The authors show that their RE matching system convincingly outperforms a state-of-the-art Network Intrusion Detection tool with support for efficient RE matching. The system can match REs at high speeds and in their full generality, even with limited memory.", "future_impact": "The performance and scalability of the presented RE matching system could significantly improve tasks in the areas of XML content distribution and network security. This opens the door for more efficient and effective solutions in these areas.", "venue": "SIGMOD", "year": 2008, "title": "Scalable regular expression matching on data streams"} +{"pid": "0d8a6519-5b36-43ad-82f7-d775dbb2c568", "context": "Worldwide, a large number of data sources on environment, energy, and natural resources are available. However, users face several problems when they wish to search for and utilize this information.", "key_idea": "The Ecobase project was initiated to address the mentioned problems in the context of numerous environmental applications in Brazil and Europe. The project introduces a distributed architecture for environmental information systems (EIS) based on the Le Select middleware.", "method": "The authors present the newly introduced architecture and its capabilities in this paper, providing insights into its functionality.", "outcome": "The authors share their observations and learning from the project, serving as a tangible output to the Ecobase project.", "future_impact": "Open issues have been identified during the Ecobase project which would require future research to resolve and further improve the architecture.", "venue": "SIGMOD", "year": 2001, "title": "The Ecobase project: database and web technologies for environmental information systems"} +{"pid": "e77d7628-6a8f-42e3-b89b-17d501de0394", "context": "Current data governance techniques are labor-intensive, requiring data stewards to transform business policies into governance rules. This approach is not scalable due to increased complexity and variety in the data ecosystem of enterprises with numerous data requirements, use cases, applications, tools, and systems.", "key_idea": "The paper proposes a 'contextual intelligence' framework, which entails collecting and integrating contextual metadata from a variety of sources to establish a unified repository of contextual data use across users and applications. The framework also involves implementing machine learning and AI techniques on this metadata to improve data governance practices.", "method": "An architecture is proposed that unifies governance across several systems, with a graph serving as a core repository of contextual metadata. This architecture aims to accurately represent data usage across the enterprise and facilitate machine learning. The application of this approach to enable ML-based recommendations in support of governance best practices is demonstrated.", "outcome": "The paper presents an architecture that leverages artificial intelligence and machine learning to improve data discoverability, quality, and compliance. The effectiveness of this approach is shown through its ability to facilitate ML-based recommendations to support best practices in data governance.", "future_impact": "Although not specifically mentioned in the abstract, it is implicit that the proposed model has the potential to streamline and scale data governance practices in increasingly data-driven enterprises, possibly shaping future techniques and strategies in data governance.", "venue": "SIGMOD", "year": 2018, "title": "Contextual Intelligence for Unified Data Governance"} +{"pid": "674b03cf-d0ae-4bd4-9d14-b1aabd6c14a6", "context": "Current solutions for developing distributed applications rely on imperative programming languages and multi-tiered application servers.", "key_idea": "The Demaq project introduces a new perspective on developing distributed applications based on the asynchronous exchange of XML messages, using a declarative language to implement the application logic as a set of rules.", "method": "A rule compiler is used to transform application specifications into execution plans against the message history, which are then evaluated using an optimized runtime engine.", "outcome": "This novel approach allows the leveraging of existing knowledge about declarative query processing for optimizing distributed applications.", "future_impact": "This new method of developing distributed applications may potentially optimize the process and improve the use of existing knowledge in declarative query processing.", "venue": "SIGMOD", "year": 2008, "title": "The Demaq system: declarative development of distributed applications"} +{"pid": "6b3cda9b-9b34-4fb1-8b45-d980e672d4c2", "context": "Previous approaches to rule support in object-oriented databases lacked a clear separation between event detection and rules.", "key_idea": "The paper proposes a novel approach to support reactive capability in an object-oriented database by introducing an 'event interface' that extends the conventional object semantics to include the role of an event generator.", "method": "The event interface approach uses a runtime subscription mechanism between rules and objects to selectively monitor particular objects dynamically, capable of supporting class level as well as instance level rules.", "outcome": "The proposed mechanism enabled objects to react to their own changes as well as to those of other objects, with both events and rules treated as first-class objects.", "future_impact": "The introduced event interface represents a new approach to rule support in object-oriented databases, and its successful utilization may inspire similar mechanisms in related domain.", "venue": "SIGMOD", "year": 1993, "title": "A new perspective on rule support for object-oriented databases"} +{"pid": "606c44b291e0114248cd013d", "context": "The cloud infrastructure promotes the disaggregation of monolithic data stores into separate components assembled per an application's workload. The interaction between components usually happens through Remote Direct Memory Access (RDMA).", "key_idea": "The authors investigate the disaggregation of an LSM-tree key-value store into components separating storage from processing, which introduces shared bandwidth and space. The system, Nova-LSM, enhances performance through dynamic range construction to parallelize compaction, and utilizes power-of-d for load balancing across an arbitrary number of storage components.", "method": "The authors compare the performance of the resulting system Nova-LSM with its monolithic counterparts, LevelDB and RocksDB, using workloads that showcase a skewed pattern of data access.", "outcome": "Nova-LSM outperforms its monolithic counterparts, LevelDB and RocksDB, by several orders of magnitude with workloads exhibiting skewed data access patterns.", "future_impact": "The conceptualisation of LSM-tree key-value store as separate components communicating through RDMA could set a new precedent in future dispersed data storage creation and could improve performance for applications with skewed data access patterns.", "venue": "SIGMOD", "year": 2021, "title": "Nova-LSM: A Distributed, Component-based LSM-tree Key-value Store"} +{"pid": "148d638b-9e87-4e1e-9bd0-c2b5a0287ddf", "context": "The emergence of differential privacy as a primary standard for privacy protection has led to the development of many algorithms, but the deployment of these techniques has been slowed by the complexity of the algorithms and an incomplete understanding of the accuracy cost involved in adopting differential privacy.", "key_idea": "The authors present DPComp, a publicly-accessible web-based system designed to support a broad community of users in assessing the accuracy of privacy algorithms, and allow for qualitative and quantitative understanding of the error introduced by the algorithms.", "method": "Users can interact with DPComp and contribute new algorithms or datasets, which are automatically incorporated into a benchmark based on a rigorous evaluation methodology.", "outcome": "The demonstration of DPComp allows users to interactively assess the accuracy of state-of-the-art privacy algorithms and understand the error introduced by the algorithms. It also embraces user-contributions for algorithms and datasets.", "future_impact": "Through DPComp, users can better interactively understand and explore accuracy-cost trade-offs in differential privacy algorithms, advancing privacy research and contributing to a clearer understanding of the impact of differential privacy.", "venue": "SIGMOD", "year": 2016, "title": "Exploring Privacy-Accuracy Tradeoffs using DPComp"} +{"pid": "624a61af5aee126c0f90c20e", "context": "Local differential privacy (LDP) has been utilized for private streaming data collection and analysis. However, existing LDP studies over streams are either limited to finite streams only or suffer from insufficient protection, high noise, and costly communication.", "key_idea": "The authors present LDP-IDS, a novel w-event LDP paradigm, designed to provide practical privacy guarantee for infinite streams, by constructing a unified error analysis and adapting the existing budget division framework in centralized differential privacy (CDP) for LDP-IDS.", "method": "They propose a novel and extensible framework of population division and recycling, involving online adaptive population division algorithms for LDP-IDS, and validate their approach through extensive experiments on synthetic and real-world datasets.", "outcome": "The proposed framework and methods demonstrate a significant reduction in utility loss and communication overhead. The results showed that LDP-IDS's effectiveness, efficiency, and flexibility are high in practice.", "future_impact": "The authors predict that their framework will be compatible with various analytic tasks. It displays flexibility in incorporating ideas of many existing stream algorithms, implying potential for broader application in the field.", "venue": "SIGMOD", "year": 2022, "title": "LDP-IDS: Local Differential Privacy for Infinite Data Streams"} +{"pid": "552c9da0-9fba-4d86-b992-5fb339c471a8", "context": "Subgraph listing is a fundamental operation in graph and network analysis. However, existing computational solutions, including recent parallel approaches, struggle to scale effectively. Specifically, they often rely on computationally expensive join operations, impeding performance for large graphs.", "key_idea": "The authors propose a novel parallel subgraph listing framework, PSgL. The framework leverages an iterative divide-and-conquer process for subgraph enumeration and avoids the explicit join operation, instead relying on graph traversal.", "method": "The authors implement the prototype of PSgL and conduct extensive experiments of various graph listing operations on diverse large graphs. They evaluate performance balance, reduction of intermediate results, and overall efficiency.", "outcome": "The experiments indicate that the PSgL framework can achieve performance gains over existing solutions of up to 90%, demonstrating robust performance in diverse large-scale graph contexts.", "future_impact": "Although explicit future impacts are not detailed in the abstract, the creation of a robust and efficient parallel subgraph listing framework could significantly improve future large-scale graph analysis tasks.", "venue": "SIGMOD", "year": 2014, "title": "Parallel subgraph listing in a large-scale graph"} +{"pid": "63718695-efc0-4b0d-a6ea-1ec38ef5b227", "context": "Many online databases are hidden behind web forms, where a small sample of the database can be retrieved through specific queries. Sampling these hidden databases is important for understanding the quality and nature of the data stored within them, but a practical method for such sampling is lacking.", "key_idea": "The authors have developed HDSampler, which is, to their knowledge, the first practical system for sampling structured hidden web databases, enabling efficient sampling and accurate aggregate query answering.", "method": "The authors demonstrated the capabilities of HDSampler by applying it to Google Base, a content-rich hidden web database maintained by Google.", "outcome": "HDSampler was able to provide a snapshot of the marginal distribution of various attributes of Google Base in a few minutes.", "future_impact": "The development of HDSampler could enable more efficient data analytics and help power a multitude of third-party applications, such as web-mashups and meta-search engines.", "venue": "SIGMOD", "year": 2009, "title": "HDSampler: revealing data behind web form interfaces"} +{"pid": "0bf5cffa-b73a-4dea-bba2-2df11bf38dff", "context": "Microblogging sites provide a large amount of manually-generated content. Entity linking is crucial for understanding and maintaining such information. However, traditional entity linking methods for documents are largely inapplicable for tweets due to their short, informal, and error-prone nature. Existing methods for entity linking in tweets utilise information from other tweets and link entities in a batch manner, which is computationally complex and not suitable for real-time applications due to high arrival rate of tweets.", "key_idea": "The authors propose an efficient solution for linking entities in tweets by analyzing their social and temporal context. The proposed framework leverages entity popularity, entity recency, and user interest information embedded in social interactions to assist the entity linking task.", "method": "The authors develop effective indexing structures and incremental algorithms to reduce the computation and maintenance costs of the approach. They test the effectiveness and efficiency of their approach using real tweet datasets.", "outcome": "Experimental results based on real tweet datasets verify the effectiveness and efficiency of the proposed solution for linking entities in tweets.", "future_impact": "The proposed solution can facilitate many real-world applications such as tweet clustering and classification, personalized microblog search, and it is suitable for high-arrival rate, real-time applications.", "venue": "SIGMOD", "year": 2015, "title": "Microblog Entity Linking with Social Temporal Context"} +{"pid": "11f9ff97-7d6f-4c6a-9c80-e09318e3c66d", "context": "Training machine learning (ML) models on large datasets is a major computational cost in the enterprise. Many analysts use a small sample of their entire data during their initial stage of analysis, which often results in ad-hoc sampling without capturing precise quality of the model.", "key_idea": "BlinkML, a system for fast, quality-guaranteed ML training has been proposed. It allows users to make error-computation tradeoffs by training an approximate model with quality guarantees using a sample instead of training a model on their full data.", "method": "BlinkML is tested on ML models that rely on maximum likelihood estimation (MLE), which include Generalized Linear Models and PPCA (Probabilistic Principal Component Analysis).", "outcome": "Experiments show that BlinkML speeds up the training of large-scale ML tasks by 6.26\u00d7 to 629\u00d7 while guaranteeing the same predictions, with 95% probability, as the full model.", "future_impact": "BlinkML allows for a more efficient and accurate method of ML model training, which can significantly optimize the decision-making process during the tuning phase in enterprise settings.", "venue": "SIGMOD", "year": 2019, "title": "BlinkML: Efficient Maximum Likelihood Estimation with Probabilistic Guarantees"} +{"pid": "bcf6c9a1-c384-41a8-93d5-efd6cf61e70e", "context": "Previously, there was no solution to correlate continuous streams of XML data through approximate (structure and content) matching as defined by a general tree-edit distance metric in small space.", "key_idea": "The authors propose a novel algorithm for obliviously embedding tree-edit distance metrics into an L1 vector space with a guaranteed worst-case upper bound on distance distortion.", "method": "The authors apply their embedding algorithm in conjunction with known random sketching techniques to build a compact synopsis of a massive, streaming XML data tree and to approximate the result of tree-edit-distance similarity joins over continuous XML document streams. Experimental study was conducted with both synthetic and real-life XML data trees.", "outcome": "Experimental results validate the proposed approach, demonstrating that the average-case behavior of the embedding techniques is much better than the predicted theoretical worst-case distortion bounds.", "future_impact": "The authors' work could revolutionize the approach to correlating XML data in the streaming model through similarity joins and other methods or provide a stepping stone to further algorithmic results on low-distortion embeddings for tree-edit distance metrics.", "venue": "SIGMOD", "year": 2005, "title": "XML stream processing using tree-edit distance embeddings"} +{"pid": "60d3093491e01182ead65d1c", "context": "Natural-language interfaces are gaining popularity due to their potential to democratize access to data, insights, and make the interaction with data more natural and accessible for a wide range of business users. However, for complete democratization, effective and continuous guidance is needed during data exploration.", "key_idea": "Boomerang is proposed, a system that recommends data-driven insights to guide exploration of datasets through a conversational interface. It aggregates recommendations from statistical, collaborative, and content-based recommenders, selecting insights that closely match the user's current state of data exploration.", "method": "Boomerang combines metrics like relevance, interestingness, and timeliness to rank insights and recommends them based on the current conversational context. The demo showcases its application for guided data exploration on a sales dataset.", "outcome": "The system successfully guides data exploration on a sales dataset, providing data-driven insights to the user through a conversational interface.", "future_impact": "The Boomerang system could make data exploration more accessible and intuitive for a wide range of business users, democratizing access to data and insights.", "venue": "SIGMOD", "year": 2021, "title": "Boomerang: Proactive Insight-Based Recommendations for Guiding Conversational Data Analysis"} +{"pid": "582bda1f-9059-40a1-a38d-4756c196ac3c", "context": "Traditional query processors generate full, accurate query results, either in batch or in pipelined fashion, a model that may be too rigid for exploratory queries over diverse and distributed data sources, such as sources on the Internet.", "key_idea": "A more flexible model of querying is proposed, where a user submits a broad initial query outline, and the system continually generates partial result tuples that may contain values for only some of the output fields allowing the user to refine the query over time.", "method": "The authors describe the querying model and user interface and present a query processing architecture for this model which is implemented in the Telegraph dataflow system. They also study reordering policies that optimize for the quality of partial results delivered over time.", "outcome": "Experimental evaluation demonstrated the benefits of the new querying architecture, especially in terms of generating partial results quickly and adjusting query execution to changing user interests.", "future_impact": "The architecture may change the way querying is performed over diverse and distributed data sources, allowing for more user-friendly and adaptable query processing.", "venue": "SIGMOD", "year": 2002, "title": "Partial results for online query processing"} +{"pid": "c5867b57-c79b-4c88-970c-fd9c0eb082ac", "context": "Much attention has been given to improving the node fanout of B-tree internal nodes to minimize tree height and the importance of B-tree page size. However, less focus has been given on understanding the architecture of B-tree internal nodes.", "key_idea": "The author seeks to describe the evolution of internal node architecture and techniques in B-trees, considering each problem that was addressed during the various incremental steps that have led to improved node organizations.", "method": "This paper appears to leverage a review and historical analysis of the evolution of B-tree node organizations and the problems tackled during each step of the improvement.", "outcome": "The paper provides a detailed account of the evolution of B-tree internal node architecture, shedding light on problems tackled at each stage of improvement towards better node organizations.", "future_impact": "The understanding garnered from this historical account could guide future optimization efforts related to B-tree architecture and node organization techniques.", "venue": "SIGMOD", "year": 2001, "title": "The evolution of effective B-tree: page organization and techniques: a personal account"} +{"pid": "329e934b-f724-4ebf-9914-5c907124bce4", "context": "Database systems for real-time applications must balance maintaining data consistency, fulfilling timing constraints of transactions, and ensuring security. The introduction of multilevel security requirements makes transaction processing even more complicated in real-time database systems.", "key_idea": "The authors propose that trade-offs between security and timeliness need to be made because of the complexities involved, presenting an adaptive method to balance these factors, depending on the current system state.", "method": "The authors discuss and use secure two-phase locking protocol as the method to experiment their adaptive approach.", "outcome": "The performance of the adapted secure two-phase locking protocol shows improved timeliness in real-time secure database systems.", "future_impact": "The authors indicate a direction for future research to further improve the timeliness of secure database systems.", "venue": "SIGMOD", "year": 1996, "title": "Improving timeliness in real-time secure database systems"} +{"pid": "1ea6b1d1-d1b4-47f8-8a49-c609a8d64ff8", "context": "Building personal data management applications often rely on centralized data and protection management, requiring account administration and presenting challenges for efficient data organization and sharing.", "key_idea": "HomeViews, a new peer-to-peer middleware system for building personal data management applications, seamlessly integrates views and queries from databases, a capability-based protection model from operating systems, and a peer-to-peer distributed architecture.", "method": "The authors developed a prototype of HomeViews, deployed it on a small network of Linux machines, and utilized it to develop two distributed data-sharing applications: a peer-to-peer version of the Gallery photo-sharing application and a simple read-only shared file system. The practicality and performance of this approach were then tested.", "outcome": "The HomeViews prototype was successfully implemented and used to develop two distributed data-sharing applications. Measurements from these applications demonstrated the practicality and performance of the HomeViews approach.", "future_impact": "The novel combination of database, operating system and peer-to-peer concepts in HomeViews can simplify and enhance the building of personal data management applications, potentially influencing the way such systems are designed and implemented.", "venue": "SIGMOD", "year": 2007, "title": "Homeviews: peer-to-peer middleware for personal data sharing applications"} +{"pid": "dc3a4509-325e-4f02-9f78-72789d14493c", "context": "The current use of machine learning processes often involves moving data out of the database systems for computations, which can be inefficient.", "key_idea": "The paper introduces a dedicated machine learning language, MLearn, designed to move computations inside the core of database systems. It also proposes the integration of ArrayQL as stored procedures to unify the relational and array perspectives.", "method": "The authors outline the requirements of their ML2SQL compiler that allows MLearn to be run on different target architectures, and explain how the compiler translates into target specific user-defined-functions for the PostgreSQL and HyPer database systems.", "outcome": "The paper explains the necessary tensor datatypes and the corresponding functions based on an accompanying example of linear regression, demonstrating how to effectively use MLearn.", "future_impact": "The development of MLearn, which enables moving computations inside the database systems, and the proposed integration of ArrayQL, may influence the design of future machine learning language and database systems for more efficient data processing.", "venue": "SIGMOD", "year": 2019, "title": "MLearn: A Declarative Machine Learning Language for Database Systems"} +{"pid": "418a21f2-2d28-4a65-a763-b820dc14151a", "context": "The Living In a Database system (LID) is an entity-relationship database interface with its ideas similar to Cattellu0027s PDB [Cattell 80]. However, interfaces like these often lack the ability to operate on sets of data instances in the same way as individual data instances, presenting an implementation challenge.", "key_idea": "LID uses a bit-mapped graphics terminal with a mouse pointer to create an attractive interaction environment and introduces dynamic graphic displays that change as the data they present change.", "method": "The development and implementation of LID, using a dynamic graphics interface on a bit-mapped graphics terminal, was undertaken to examine its practicality and identify challenges.", "outcome": "The implementation indicated that dynamic graphic displays are an important feature in user interfaces but are difficult to implement given the current technology. It also underlined a critical inadequacy in the PDB/LID idea: the inability to operate on sets of data instances similarly to individual data instances.", "future_impact": "An extension suggestion to LID is provided which aims to alleviate the problem of not being able to operate on sets of data instances in the same way as individual instances, which could influence the development of future graphical query interfaces.", "venue": "SIGMOD", "year": 1984, "title": "Lessons from a \"living in a database\" graphical query interface"} +{"pid": "af944a96-e175-4419-aeb8-306bef765c13", "context": "Many researchers in the Earth and space sciences are becoming aware of the International Directory Network (IDN), a federation of international directories to Earth and space science data, which can potentially become distributed nodes of a single, virtual master data directory.", "key_idea": "The authors emphasize on the potential and existence of Earth-and-space-sciece-relevant information systems that can be accessed automatically from the directories, enriching the utility of the IDN by providing detailed information about the correlative possibilities of these data sets.", "method": "The approach to gathering detailed information and capabilities about the data sets involves automatic accessing of connected guide/catalog and inventory systems via the IDN directories.", "outcome": "The utilization of IDN directories now enables capabilities such as data set browse, subsetting, and analysis.", "future_impact": "The usefulness and convenience of accessing detailed information about Earth and space science data sets through the IDN will continue to improve in the future.", "venue": "SIGMOD", "year": 1993, "title": "The international directory network and connected data information systems for research in the earth and space sciences"} +{"pid": "8b2c81bc-1dfa-49a3-b3e2-df75ddbdb901", "context": "As the amount of motion history, or trajectory, data being collected from devices such as GPS-enabled mobiles, surveillance cameras, and social networks increases, its storage and management become a challenge for traditional databases due to the variable lengths and asynchronous sampling rates of this data.", "key_idea": "The authors developed a novel trajectory storage system, leveraging the successful techniques from column store and in-memory databases to boost query processing performance for trajectory data.", "method": "The authors structured and partitioned the trajectories into frames in a column-oriented in-memory storage method. This technique aligns the sample points of a moving object by the time interval, and these frames are highly compressed and structured to enhance memory utilization and reduce CPU-cache miss.", "outcome": "The novel storage design allows for increased memory utilization, reduced CPU-cache missing, and facilitates parallel data processing on a multi-core server as the data frames are mutually independent.", "future_impact": "The proposed trajectory storage system could make it easier to manage and store the growing amount of trajectory data generated from various sources, potentially impacting tasks such as real-time tracking, anomaly detection, and other applications dealing with large-scale trajectory data.", "venue": "SIGMOD", "year": 2015, "title": "SharkDB: An In-Memory Storage System for Massive Trajectory Data"} +{"pid": "5c64d33d-b2be-43dc-aa5d-bf93849e6e5a", "context": "Data transformation, a crucial step in data analysis, involves bringing diverse data into a uniform format that can be analyzed. This process is usually tedious, time-consuming, and often requires programming skills that may exceed the expertise of data analysts.", "key_idea": "The authors propose a technique for synthesizing data transformation programs through examples, with the analyst providing a small input-output example pair, without needing to describe the transformation steps. In addition, a system has been implemented called FOOFAH using this technique.", "method": "The authors developed the FOOFAH system, which efficiently explores the space of possible data transformation operations to generate a program that performs the target transformation. FOOFAH's performance was then compared with the widely recognized WRANGLER system.", "outcome": "The experiments show that data transformation programs can be created quickly using FOOFAH in a variety of instances, requiring 60% less user effort than the well-known WRANGLER system.", "future_impact": "The technique proposed in this paper may reduce the burden of data transformation on data analysts in real-world applications, by allowing them to define transformations via examples without needing to program the detailed steps, resulting in more accessible and efficient data analysis.", "venue": "SIGMOD", "year": 2017, "title": "Foofah: Transforming Data By Example"} +{"pid": "bf21ec5a-896b-4b8e-a0d6-2e102f30c6ee", "context": "Creating a large-scale, highly accurate knowledge base of common facts with convenient searchability is a challenging task. Traditional approaches often neglect to reconcile individual entity facts from sources like Wikipedia with consistent class systems like WordNet's taxonomy.", "key_idea": "The paper presents the YAGO-NAGA method for information extraction, which combines the fact-harvesting capabilities of Wikipedia infoboxes and category names with the taxonomic categorization of WordNet. This ensures proper classification for all entities and consistency in the class system.", "method": "The authors present the architecture of the YAGO extractor toolkit, with a focus on consistency checking, ongoing maintenance and growth mechanisms, and a query engine named NAGA. The paper also discusses the accuracy estimation of YAGO based on intensive sampling.", "outcome": "YAGO currently contains around 19 million instances of binary relations for approximately 1.95 million entities and has an accuracy estimated to be above 95 percent.", "future_impact": "The YAGO-NAGA approach provides new avenues for further work in integrating fact candidates extracted from natural-language text sources, which may yield even more comprehensive knowledge bases in the future.", "venue": "SIGMOD", "year": 2009, "title": "The YAGO-NAGA approach to knowledge discovery"} +{"pid": "9facccf3-ae14-457d-87a9-99e77a558345", "context": "Most probabilistic database systems have complex language definitions, and it is often difficult to identify the core query algebra within these definitions.", "key_idea": "The authors intend to define a core query algebra for probabilistic databases, which captures the most essential parts of query languages in most existing probabilistic database systems.", "method": "Authors formally define the proposed algebra and illustrate it with examples. They also survey the current knowledge regarding the core's expressive power and complexity.", "outcome": "The authors have defined a core query algebra for probabilistic databases and have surveyed its expressive power and complexity.", "future_impact": "This core query algebra has the potential to guide the development and modification of probabilistic database languages, simplifying the process by providing a clearer and simpler foundation.", "venue": "SIGMOD", "year": 2009, "title": "On Query Algebras for Probabilistic Databases"} +{"pid": "e644866e-de5b-4b57-a175-e49fbe6851ca", "context": "PointCast Inc is renowned as the inventor and leader in delivering broadcast news through the Internet and corporate intranets, using a client/server system.", "key_idea": "The PointCast Network is constructed with an integrated client/server system. It is devised to offer users control over selecting the kind of information retrieved and how often this occurs. The network harnesses a system of servers called PointServers at geographically disparate locations, which can serve high volumes of requests shoring up to 200 million in a day.", "method": "The PointCast network receives content from over 100 different sources via satellite links or the internet. After assimilation, the data from these sources is indexed, formatted, and stored in multiple content databases.", "outcome": "The PointCast network successfully handles close to 200 million requests in a day from its servers, located across the United States, by assimilating data from over 100 sources and serving it to customers.", "future_impact": "This paper lays out the challenges and associated solutions for creating a high-traffic data center, focusing on fault tolerance, load balancing, achieving scalability through pre-caching on servers, packaging information to optimize internet bandwidth usage, and minimizing data latency, which could enlighten similar future endeavors.", "venue": "SIGMOD", "year": 1998, "title": "The PointCast network (abstract)"} +{"pid": "c2271b61-c57c-439b-8b93-ade0db3aeb5b", "context": "The concept of normalisation and the definition of third normal form hold central importance in the relational model of a database, and no generally recognized data definition language for the relational model exists yet.", "key_idea": "This paper proposes a constructive definition of the third normal form, using rules for building normalised relations from simple domains. The method applies two ideas derived from the study of programming languages -- the notion of data type, and the notion of a selector.", "method": "The paper aims at achieving a concrete definition of the third normal form using the ideas of data type and selector. The abstract syntax of relations created is then compared with the standard definition of normalisation.", "outcome": "The paper results in an abstract syntax of relations. The proposed definition was compared with the standard definition of normalisation and its completeness was explored.", "future_impact": "The constructive definition of the third normal form could be a crucial step towards creating a generally accepted data definition language for the relational model.", "venue": "SIGMOD", "year": 1976, "title": "A constructive definition of third normal form"} +{"pid": "7cf755ec-3026-40e4-a099-3ad758259b0b", "context": "Most existing protocols for the 1-out-of-n single-server Computationally-Private Information Retrieval (CPIR) problem exploit the trade-off between communication and computation, and have high complexity measures.", "key_idea": "The authors propose new protocols for the CPIR problem that trade accuracy for reduced complexity measures, suggesting that it is possible to decrease these complexities at the cost of results' precision.", "method": "The work is purely theoretical and does not present a specific experiment or methodology to validate the proposed idea.", "outcome": "The proposed protocols achieve polylogarithmic extra storage, communication, and client and server local computational complexities, enhancing the best current bounds.", "future_impact": "The indirect (practical) implication of this work is to prove that CPIR is viable for very large databases, potentially influencing the design of future CPIR protocols. The authors also present some open questions and future research directions.", "venue": "SIGMOD", "year": 2008, "title": "Approximate private information retrieval"} +{"pid": "dacd7012-ff98-4e35-8bfd-106bcfd9ff90", "context": "The problem of k nearest neighbor (k-NN) querying is well-studied for static datasets; however, traditional methods do not extend well to dynamic environments where multiple continuous queries require real-time results, and both objects and queries are subject to frequent location updates.", "key_idea": "This paper proposes the use of Conceptual Partitioning (CPM), an efficient technique for continuous nearest neighbor (NN) query monitoring in dynmaic environments, which works by handling location updates only from objects that fall in the vicinity of some query.", "method": "The performance of CPM was analysed against current state-of-the-art algorithms for various problem settings. Additionally, the framework was extended to aggregate NN (ANN) queries.", "outcome": "The analysis showed that CPM outperforms the current state-of-the-art algorithms across all evaluated settings.", "future_impact": "While not mentioned in the abstract, it can be inferred that the successful application of CPM could lead to more efficient real-time location-based services and high update-rate applications.", "venue": "SIGMOD", "year": 2005, "title": "Conceptual partitioning: an efficient method for continuous nearest neighbor monitoring"} +{"pid": "bfec08da-2d0e-451c-9c03-7b2118d21937", "context": "The Linked Data Benchmark Council (LDBC) is two years into its mission to establish benchmarks and benchmarking practices for evaluating graph data management systems. There is a lack of a unified framework for testing graph query functionality, graph generation techniques, and graph algorithm workloads.", "key_idea": "The LDBC introduced a choke-point driven methodology for developing benchmark workloads, incorporating both user input and expert systems architects' advice. The paper introduces the LDBC Social Network Benchmark (SNB) which tests graph query functionality, uses correlated graph generation techniques, and has a scalable benchmark driver managing a workload with complex graph dependencies.", "method": "The authors present the SNB Interactive Workload in great detail and perform early trials to demonstrate its functionality.", "outcome": "The benchmark has yielded some early results from the Interactive Workload, but specific details are not provided in the abstract.", "future_impact": "The two other workloads of the SNB, the Business Intelligence, and Graph Algorithms, would be further developed with outlined goals.", "venue": "SIGMOD", "year": 2015, "title": "The LDBC Social Network Benchmark: Interactive Workload"} +{"pid": "95520d88-6703-4402-bee1-dde458278f52", "context": "Existing methods of describing and comparing database concurrency control algorithms are likely complex or inconsistent, making it difficult to compare relative storage and CPU overheads.", "key_idea": "The paper presents an abstract model of concurrency control algorithms that enables implementation-independent descriptions and clear comparison based on required information, blocking/restart conditions, and request processing.", "method": "Single-site versions of two-phase locking, basic timestamp ordering, and serial validation are evaluated utilizing the proposed model, highlighting its effectiveness.", "outcome": "Using the proposed abstract model, the authors were able to distinguish and compare the relative storage and CPU overheads of various concurrency control algorithms accurately.", "future_impact": "The paper also discusses potential extensions to the abstract model that could facilitate comparison of multiple version and distributed algorithms, suggesting it may be a foundation for future research.", "venue": "SIGMOD", "year": 1983, "title": "An abstract model of database concurrency control algorithms"} +{"pid": "7427ea94-96cb-4c7a-88f0-59ada2dce7fe", "context": "Database administrators construct secondary indexes on frequently queried columns to speed up query processing in relational database management systems (RDBMSs). However, maintaining multiple secondary indexes can consume a lot of space and cause significant performance degradation due to potential exhaustion of memory space.", "key_idea": "The authors propose HERMIT, a succinct secondary indexing mechanism for modern RDBMSs that exploits column correlations to accelerate data access. Instead of storing every single entry in the key columns, HERMIT uses an existing index built on correlated columns and navigates any incoming key access queries to it.", "method": "HERMIT uses the Tiered Regression Search Tree (TRS-Tree), a succinct, ML-enhanced data structure for fast curve fitting to dynamically capture both column correlations and outliers. The authors conducted an extensive experimental study in two different RDBMSs to validate the method.", "outcome": "Experimental results demonstrate that HERMIT significantly reduces space consumption with limited performance overhead, especially when it comes to supporting complex range queries.", "future_impact": "HERMIT is anticipated to enhance the efficiency of modern RDBMSs by reducing the space consumption of secondary indexes, which could influence future developments in the field of database management.", "venue": "SIGMOD", "year": 2019, "title": "Designing Succinct Secondary Indexing Mechanism by Exploiting Column Correlations"} +{"pid": "359b416d-eed6-4bd0-984e-09580f43575c", "context": "The existing literature lacks adequate characterization and evaluation of restart and recovery policies in a transaction-oriented information processing system.", "key_idea": "The authors present a model to understand the relationships between errors in information processing systems, the necessary procedures to recover from these errors, and the actions needed to make recovery procedures feasible.", "method": "The authors define and discuss restart and recovery policies, and they apply an objective function to evaluate these policies.", "outcome": "The authors present results obtained through the evaluation of recovery and restart policies using a particular objective function, demonstrating the feasibility of marginal cost analysis.", "future_impact": "The methodology presented could help an operations manager perform marginal cost analysis. The model could also serve as groundwork for further research into recovery procedures in information systems.", "venue": "SIGMOD", "year": 1974, "title": "Restart and recovery in a transaction-oriented information processing system"} +{"pid": "8c205ca5-5eb8-4f6b-a1db-41cf1bdc9bb6", "context": "In a co-space environment, physical and virtual spaces co-exist and interact continuously. Technologies have to be developed to facilitate users' seamless manipulation of real and digital spaces.", "key_idea": "This paper introduces the concept of co-space applications, which provide opportunities unavailable in separate physical or virtual spaces, and posit that the database community can significantly contribute to this emerging field.", "method": "The authors conducted a discussion about the potential applications of co-space and the challenges their management brings.", "outcome": "The paper sets the premise that the database community can meet challenges posed by co-space environments, providing no quantitative outcomes.", "future_impact": "The emergence of co-space applications could bring new challenges for the database community to address, creating opportunities for further research and development in managing co-spaces.", "venue": "SIGMOD", "year": 2010, "title": "Sense the physical, walkthrough the virtual, manage the co (existing) spaces: a database perspective"} +{"pid": "5ed384dd9fced0a24b65700c", "context": "Spreadsheet systems are used for storing and analyzing data across domains by programmers and non-programmer alike. Despite supporting increasingly large datasets, these systems are prone to hanging and freezing while performing computations even on smaller datasets.", "key_idea": "The authors present a benchmarking study designed to evaluate and compare the performance of three popular spreadsheet systems: Microsoft Excel, LibreOffice Calc, and Google Sheets, on a range of spreadsheet computation operations.", "method": "The authors perform evaluation and comparison of Microsoft Excel, LibreOffice Calc, and Google Sheets on various spreadsheet computation operations. They also evaluate whether spreadsheet systems adopt database optimization techniques like indexing, intelligent data layout, and so on.", "outcome": "Their research shows that spreadsheet systems lack interactivity for several operations on data sets significantly below their advertised scalability limits. They also found that these spreadsheet systems do not adequately adopt optimization techniques.", "future_impact": "The authors outline several ways in which future spreadsheet systems could be redesigned to offer interactive response times on large datasets, potentially improving their usability.", "venue": "SIGMOD", "year": 2020, "title": "Benchmarking Spreadsheet Systems"} +{"pid": "8c5e7b83-7928-4215-b546-49d3145e7edb", "context": "The semi-join is a relational operator used in processing queries involving binary operations in databases, focusing on the initial selection of relevant data to reduce the size of operand relations and computational cost. However, optimization algorithms for executing semi-joins in a multiprocessor environment are not fully explored.", "key_idea": "This paper presents and analyzes new algorithms for executing semi-joins within a multiprocessor database system for increased efficiency, including two equi-semi-join algorithms and one inequi-semi-join algorithm.", "method": "The authors initially describe their architecture model for a multiprocessor system, which factors in I-O, CPU, and message transmission costs. The costs of executing their proposed algorithms are then evaluated using this model, and methods using semi-joins are compared to those joining two relations.", "outcome": "The execution cost of these algorithms is found to generally be linearly proportional to the size of the operand and result relations and inversely proportional to the number of processors. It is shown that the method using semi-joins is generally better than the method using two relation joins.", "future_impact": "These algorithms and their approving evaluation could lead to developments in improving the efficiency of executing semi-joins in multiprocessor database systems.", "venue": "SIGMOD", "year": 1982, "title": "Semi-join algorithms for multiprocessor systems"} +{"pid": "a98df90d-42f9-474a-b63b-47b3199bf795", "context": "Active database systems are in widespread use, but the interaction between triggers, transactions and application programs can lead to costly redundant computations in rule conditions and actions.", "key_idea": "The authors propose a solution to eliminate redundant computations of SQL triggers in active relational database systems when these computations are costly.", "method": "The methodology involves defining a model to describe programs, rules and their interactions, providing algorithms to extract invariant subqueries from trigger's condition and action, and developing a rewriting technique to generate and execute the optimized code of SQL triggers.", "outcome": "The authors have successfully devised a solution to eliminate the costly redundant computations of SQL triggers.", "future_impact": "Their solution could optimize the use of active relational database systems by eliminating inefficiencies caused by redundant computations.", "venue": "SIGMOD", "year": 1997, "title": "Eliminating costly redundant computations from SQL trigger executions"} +{"pid": "60d3093491e01182ead65d77", "context": "Most database systems delegate scheduling decisions to the operating system, which simplifies the overall database design but presents challenges for adaptive resource allocation and incorporating domain knowledge to improve query scheduling. Modern systems often use task-based parallelism, breaking a single query into small, independent tasks for fine-grained scheduling. However, little work has investigated the optimization opportunities created by this execution model.", "key_idea": "The authors present a lock-free, self-tuning stride scheduler for database systems that optimizes query latencies for analytical workloads, with the ability to adaptively manage query priorities and task granularity for high scheduling elasticity, and incorporation of domain knowledge into scheduling decisions.", "method": "The paper's authors compare their novel stride scheduler to traditional database systems under various workloads, focusing on query latencies and workload management.", "outcome": "The self-tuning stride scheduler retains near-optimal latencies for short-running queries even at high load, and often improves tail latencies by over 10 times compared to traditional database systems.", "future_impact": "The proposed scheduler may serve as a strong solution for database systems having difficulty with certain workloads, making query processing more efficient in analytical environments.", "venue": "SIGMOD", "year": 2021, "title": "Self-Tuning Query Scheduling for Analytical Workloads"} +{"pid": "090bc21c-0188-4b67-99b6-a5c7352b63ed", "context": "Query optimizers of current relational database systems use several statistics maintained by the system on the contents of the database to decide on the most efficient access plan for a given query. These statistics contain errors that transitively affect many estimates derived by the optimizer.", "key_idea": "The authors present a formal framework to study the principles of error propagation in query optimization based on system's statistics.", "method": "Within this formal framework, the authors derive several analytic results on how the error propagates in general, as well as in the extreme and average cases. They also analyze the guarantees that the database system can make based on the statistics that it maintains and discuss some promising approaches to controlling the error propagation.", "outcome": "The authors provide analytic results on error propagation and deriving properties of approaches controlling error propagation.", "future_impact": "This study could direct future research toward controlling error propagation and improving the efficiency and accuracy of query optimization in database systems.", "venue": "SIGMOD", "year": 1991, "title": "On the propagation of errors in the size of join results"} +{"pid": "33204577-f9b2-49de-8f64-1a12fd1ad7a5", "context": "Correlated queries are common and pivotal in decision support systems, but traditional nested iteration methods for such queries can be time-consuming. Query rewriting techniques may provide a more efficient alternative, but they're not always applicable.", "key_idea": "This paper introduces a new method to evaluate correlated queries efficiently through use of an 'invariant' technique, where portions of the subquery not related to outer references are recognized and the results are cached after the first execution for reuse in combination with changing parts of the subquery in each iteration.", "method": "Three algorithms were proposed to recognize the invariant part of a data flow tree and restructure the evaluation plan to reuse stored intermediate results. The authors also suggested a method to teach an existing join optimizer to comprehend the invariant feature, enabling it to generate better join plans. Their techniques were implemented on an existing real commercial database system, and their efficiency was evaluated experimentally.", "outcome": "The experimental evaluation demonstrated the superiority of the invariant technique over traditional nested iteration methods when query rewriting isn't possible, and sometimes over query rewriting methods as well.", "future_impact": "The invariant technique is poised to fill a key gap in the evaluation of correlated queries not covered by query rewriting techniques, implying it could emerge as an indispensable tool in correlated query handling.", "venue": "SIGMOD", "year": 1998, "title": "Reusing invariants: a new strategy for correlated queries"} +{"pid": "9497abb5-b3ba-405b-ba18-da3c9f1a9ac3", "context": "Due to the rapid growth of the Semantic Web, there is a lot of RDF data and ontologies created, leading to a need for efficient RDF data management on top of relational databases. This need has gained attention from both the Semantic Web and database communities.", "key_idea": "In this paper, the authors propose an effective and efficient method for Semantic Web data management over DB2, including a schema and indexes design for storage, practical ontology reasoning support, and a SPARQL-to-SQL translation method for RDF query.", "method": "The authors showcase the performance and scalability of their proposed system by comparing it with well-known RDF stores.", "outcome": "The paper does not provide specific details about the outcomes of the comparisons performed or any other measurable results in the abstract.", "future_impact": "While no explicit impacts are mentioned, the authors do mention that they will be discussing future work.", "venue": "SIGMOD", "year": 2008, "title": "Effective and efficient semantic web data management over DB2"} +{"pid": "85dbcd06-b3e4-4bb4-80a7-eaf73d44be6a", "context": "Certain answers are a principled method for handling uncertainty in data management tasks. However, this method is expensive and may exclude useful but uncertain answers causing users to resort to less principled approaches.", "key_idea": "The proposed Uncertainty Annotated Databases (UA-DBs) combine an under- and over-approximation of certain answers to achieve the reliability of certain answers, with the performance of a classical database system. UA-DBs are based on incomplete K-relations, a concept introduced in this paper to generalize the classical set-based notion of incomplete databases and certain answers.", "method": "An implementation of the proposed approach is used for a set of experiments to produce tight approximations of certain answers.", "outcome": "Experimental results show the efficiency of the proposed approach in producing tight approximations of certain answers that have high utility.", "future_impact": "The uncertainty-annotated databases and subsequent methodology opens up the door for more efficient and nuanced handling of uncertain data, leading to increases in data reliability and performance. Further research could measure the real-world effectiveness of this system when handling complex and uncertain data streams.", "venue": "SIGMOD", "year": 2019, "title": "Uncertainty Annotated Databases - A Lightweight Approach for Approximating Certain Answers"} +{"pid": "f0b63bec-ca54-4b77-8d6b-53bb14c0784b", "context": "The alignment of documents and speech in multimodal meetings is typically challenging, particularly when there are explicit verbal references to documents and parts of documents.", "key_idea": "The authors propose a method for document/speech alignment based on the detection of expressions referring to documents in transcribed speech and the recognition of the documents and document elements they refer to.", "method": "The authors evaluated the implemented modules of their proposed method, first separately and then in a pipeline.", "outcome": "The evaluation of the proposed method shows that results are well above baseline values.", "future_impact": "The authors suggest the integration of this method with other techniques for document/speech alignment, which might pave the way for more effective multimodal communication systems.", "venue": "ICML", "year": 2005, "title": "Detection and resolution of references to meeting documents"} +{"pid": "9f4fe417-d29c-46ac-9cdd-03bb047a3201", "context": "Principal component regression (PCR) is a computational problem typically solved through calling ridge regression repeatedly. Previous methods require O(\u03b3-2) calls, which is computationally expensive and may not be suitable for large-scale PCR instances.", "key_idea": "The authors propose a novel approach to solving PCR with reduced computational cost, by reducing it to O(\u03b3-1) black-box calls of ridge regression. The solution does not require any explicit construction of the top principal components.", "method": "The authors developed a stable recurrence formula for matrix Chebyshev polynomials and a degree-optimal polynomial approximation to the matrix sign function to achieve these results.", "outcome": "The proposed method achieved a multiplicative accuracy of 1 + \u03b3 in solving PCR tasks, proving it does not require explicit construction of top principal components and is suitable for large-scale PCR instances.", "future_impact": "The developed techniques for matrix Chebyshev polynomials and degree-optimal polynomial approximation may be of independent interest and could be useful in designing other iterative methods.", "venue": "ICML", "year": 2017, "title": "Faster principal component regression and stable matrix Chebyshev approximation"} +{"pid": "7b42efeb-6db7-456b-8ea9-5956bff5dd89", "context": "The interest in relational reinforcement learning and the need for an effective method for application in relational domains motivate this study.", "key_idea": "The study introduces REBEL, a novel relational Bellman update operator for Markov decision processes, which employs a constraint logic programming language for compact representation.", "method": "The authors developed a novel value iteration algorithm using REBEL, with abstraction playing a major role. They present convergence results and conduct experiments to validate the effectiveness of REBEL.", "outcome": "Convergence results and experimental data are produced, demonstrating the benefits of the novel relational Bellman update operator REBEL.", "future_impact": "This new framework provides insights into relational reinforcement learning which may help future investigations and applications in the same areas.", "venue": "ICML", "year": 2004, "title": "Bellman goes relational"} +{"pid": "86761497-145b-4278-af87-a15a19bfda79", "context": "In machine learning, improving performance is often achieved by balancing efficiency and predictive accuracy. However, previous work in inductive and explanation-based learning has primarily focused on improving only the predictive accuracy.", "key_idea": "The authors extend the work in explanation-based and inductive learning to introduce a framework which aims at simultaneously improving both efficiency and accuracy of a learning system without compromising one for the other.", "method": "This framework is applied to inductive learning, specifically learning from examples that build concept descriptions based on instances labeled with their class information. The efficiency of the learning system is improved through a concept hierarchy used to classify instances.", "outcome": "The framework aims to ensure that learning becomes more efficient as the system matches the concept multiple times, reducing workload for the matcher.", "future_impact": "The proposed framework, by balancing efficiency and accuracy, could lead to next-generation learning systems that offer better performance. This work suggests further research into optimization of machine learning systems.", "venue": "ICML", "year": 1989, "title": "A framework for improving efficiency and accuracy"} +{"pid": "60bdde338585e32c38af50a6", "context": "Generative learning tasks typically involve interpolating between a reference and a target distribution, often based on the Kullback-Leibler divergence. Most approaches currently use Generative Adversarial Networks (GANs) and similar methods.", "key_idea": "The authors propose to learn a generative model through entropy interpolation via a Schr\u00f6dinger Bridge, an approach that involves using a deep score estimator and a deep density ratio estimator within the Euler-Maruyama method.", "method": "The proposed Schr\u00f6dinger Bridge algorithm is theoretically analyzed for consistency under certain smoothness assumptions for the target distribution. Experimental tests are carried out on multimodal synthetic data and benchmark data, and the algorithm's usefulness is demonstrated in image interpolation and image inpainting.", "outcome": "Theoretical results support the consistency of the approach and experimental results suggest that it is comparable to state-of-the-art GANs in terms of its performance on test data. Additionally, the model proves useful in image inpainting and image interpolation tasks.", "future_impact": "The proposed Schrodinger Bridge approach provides a novel formulation for generative learning, which could serve as an alternative to GANs and other existing methods in generative modelling.", "venue": "ICML", "year": 2021, "title": "Deep Generative Learning via Schrodinger Bridge"} +{"pid": "629ec1f95aee126c0fb6ffd8", "context": "Bayesian neural networks (BNNs) show improved generalization under covariate shift, and node-based BNNs have been proposed as a scalable alternative to weight-based BNNs, but the implicit data perturbations caused by these latent noise variables have not been thoroughly explored.", "key_idea": "The paper posits that the latent noise variables in node-based BNNs represent implicit data perturbations during training, and suggests a novel approach to increase the entropy of these variables to improve performance under covariate shifts.", "method": "The introduced method is tested on out-of-distribution image classification benchmarks, focusing on performance under covariate shift due to input perturbations and noisy training labels.", "outcome": "They found that increased entropy of the latent variables improved uncertainty estimation of node-based BNNs under covariate shift due to input perturbations while also providing robustness against noisy training labels.", "future_impact": "This approach to dealing with covariate shift could inspire improved uncertainty estimation methods in neural networks, potentially leading to more robust models in cases of noisy labels and out-of-distribution inputs.", "venue": "ICML", "year": 2022, "title": "Tackling covariate shift with node-based Bayesian neural networks."} +{"pid": "62c28ae65aee126c0f8a21af", "context": "Top-$k$ classification is a generalization of multiclass classification, used in a variety of scenarios. Current surrogate solutions for the problem are either non-convex or inconsistent, suggesting a need for better solutions.", "key_idea": "The paper proposes the application of the embedding framework from Finocchiaro et al. to polyhedral surrogates, which leads to a consistent surrogate for the top-$k$ classification problem.", "method": "The authors analyze hinge-like surrogates that are convex using the embedding framework. They then use this analysis to derive conditions (constraints on the conditional label distributions) under which these surrogates become consistent for top-$k$.", "outcome": "The authors successfully derive a model for convex hinge-like surrogates that form consistent solutions for top-$k$ classification problems.", "future_impact": "The proposed consistent polyhedral surrogate for top-$k$ classification can provide a better alignment with some applications and may improve the efficiency and quality of data retrieval, image classification, and other scenarios.", "venue": "ICML", "year": 2022, "title": "Consistent Polyhedral Surrogates for Top-k Classification and Variants."} +{"pid": "60140a2791e01120e4ab58eb", "context": "The design of rotating detection regression loss has been hampered by boundary discontinuity and its inconsistency with the final detection metric.", "key_idea": "This paper proposes a novel regression loss based on Gaussian Wasserstein distance, by converting the rotated bounding box into a 2-D Gaussian distribution. This allows approximation of the indifferntiable rotational IoU induced loss by the Gaussian Wasserstein distance (GWD), which can be learned efficiently via gradient back-propagation.", "method": "The authors apply the proposed GWD to five datasets by using different detectors to show its effectiveness.", "outcome": "The proposed GWD is effective in solving the boundary discontinuity and square-like problem as evidenced by its application on five datasets using different detectors.", "future_impact": "Not mentioned in the abstract, therefore it is marked as 'N/A'", "venue": "ICML", "year": 2021, "title": "Rethinking Rotated Object Detection with Gaussian Wasserstein Distance Loss"} +{"pid": "654f6802-3049-434a-b653-1524ed904459", "context": "Traditional inspection and statistical processing methods for monitoring injection molding machines may not be accurate and efficient for total quality control.", "key_idea": "The study proposes a wireless operation monitoring and remote control system using a smartphone-based video camera. This system allows the injection molding machine operation to be monitored and defective products to be identified by comparing the injection product patterns.", "method": "The paper describes the development of a product that utilizes the proposed smartphone-based video camera system, which can be customized according to the characteristics of the production equipment and respective company.", "outcome": "The product developed in this study improves productivity and reduces cost by using existing facilities.", "future_impact": "This approach can be used for the development of production facilities capable of real-time monitoring and control, especially using mobile devices such as smartphones and tablet PCs.", "venue": "ICML", "year": 2017, "title": "Wireless Operation Monitoring and Remote Control System Implementation Using the Smartphone-Based Video Camera"} +{"pid": "6257bf9e5aee126c0f42f2c4", "context": "Designing protein sequences with a specific biological function is a long-standing challenge in protein engineering. Current machine-learning-guided approaches utilize a surrogate sequence-function model to alleviate the need for costly lab experiments.", "key_idea": "The authors leverage the natural property of protein fitness landscapes, where a concise set of mutations on the wild-type sequence suffices to enhance the desired function. They propose the Proximal Exploration (PEX) algorithm that prioritizes the evolutionary search for high-fitness mutants with low mutation counts and the Mutation Factorization Network (MuFacNet) for predicting low-order mutational effects.", "method": "The authors evaluate their method through a series of in-silico protein sequence design experiments, comparing with baseline algorithms.", "outcome": "The proposed method, which uses both Proximal Exploration (PEX) and Mutation Factorization Network (MuFacNet), shows significant improvement over baseline algorithms in in-silico protein sequence design tasks.", "future_impact": "The Proximal Exploration (PEX) algorithm and Mutation Factorization Network (MuFacNet) can increase the efficiency of model-guided evolution and potentially advance the field of protein engineering.", "venue": "ICML", "year": 2022, "title": "Proximal Exploration for Model-guided Protein Sequence Design."} +{"pid": "5e85c28491e0114016e82126", "context": "Embedding computation in molecular contexts is expected to have wide-ranging impact in various fields, but a key remaining challenge is to develop programming paradigms for molecular computation that are well-aligned with the underlying chemical hardware, as opposed to trying to fit ill-suited electronic paradigms.", "key_idea": "The authors reveal a connection between a class of binary-weight ReLU neural networks (BinaryConnect) and a class of coupled chemical reactions that are absolutely robust to reaction rates, and propose using rate-independent chemical computation as a promising target for bioengineering implementation.", "method": "The authors train a BinaryConnect neural network in silico using well-founded deep learning optimization techniques and compile it into an equivalent chemical reaction network. This is demonstrated on the IRIS and MNIST datasets, as well as on a task to discriminate between different viruses based on gene expression levels.", "outcome": "The authors demonstrate that a BinaryConnect neural network can be compiled to an equivalent chemical reaction network, and apply this to tasks such as classifying data from the IRIS and MNIST datasets and discriminating between different virus types based on gene expression levels.", "future_impact": "The work sets the stage for knowledge transfer between neural network and molecular programming communities, potentially leading to more effective computation in molecular contexts.", "venue": "ICML", "year": 2020, "title": "Deep Molecular Programming: A Natural Implementation of Binary-Weight ReLU Neural Networks"} +{"pid": "5ede0553e06a4c1b26a83e48", "context": "Existing researches using margin based comparison loss demonstrate the effectiveness of penalizing the distance between face features and the corresponding class centers. However, these approaches do not explicitly encourage the generic embedding learning for an open set recognition problem.", "key_idea": "This paper analyses margin based softmax loss from a probabilistic perspective, and proposes two general principles - monotonic decreasing and margin probability penalty - for designing new margin loss functions. It also treats open set face recognition as an information transmission problem.", "method": "The authors propose an auto-encoder architecture called Linear-Auto-TS-Encoder (LATSE) and validate this via extensive experiments on several benchmarks, effectively using the more generalized face embedding to boost single model performance.", "outcome": "The results demonstrate that LATSE boosts the single model performance with an open training dataset to more than 99% on the MegaFace test, showing it helps face embedding to gain more generalization capability.", "future_impact": "Based on the results achieved, the new probabilistic principles and the LATSE architecture have the potential to substantially improve the performance of face recognition systems in open set scenarios.", "venue": "ICML", "year": 2020, "title": "More Information Supervised Probabilistic Deep Face Embedding Learning"} +{"pid": "3e27f18a-f2db-4bee-95d4-607f08f9cb32", "context": "Working set methods, which reduce optimization to a sequence of small subproblems, achieve fast convergence times for many problems. However, theoretical understanding of working sets is limited, and implementations often resort to heuristics to determine subproblem size, makeup, and stopping criteria.", "key_idea": "The authors propose BLITZ, a fast working set algorithm, which makes no assumptions on data and introduces a theoretical relationship between subproblem size and progress toward convergence. The authors also propose methods for optimizing algorithmic parameters and discarding irrelevant variables as iterations progress.", "method": "The newly proposed algorithm, BLITZ, is applied to l1-regularized learning and compared to existing solvers in sequential, limited-memory, and distributed settings.", "outcome": "When applied to l1-regularized learning, BLITZ convincingly outperforms existing solvers in sequential, limited-memory, and distributed settings.", "future_impact": "As BLITZ is not specific to l1-regularized learning, the algorithm could be relevant to many applications involving sparsity or constraints, paving the way for its use in a range of optimization contexts.", "venue": "ICML", "year": 2015, "title": "Blitz: A Principled Meta-Algorithm for Scaling Sparse Optimization"} +{"pid": "6bc26026-cdbb-444a-8095-8eb3e0834471", "context": "Graph clustering is extensively used in the study of relational data sets, but traditional methods like the k-medoids algorithm and the Girvan-Newman method are computationally expensive, making them impractical for modern, large-scale relational data sets.", "key_idea": "The paper introduces a strategy to incorporate Network Structure Indices (NSIs), an efficient method for indexing network structure and finding short paths, into graph clustering algorithms to overcome their complexity limitations.", "method": "The authors adapt the k-medoids algorithm and Girvan-Newman method to incorporate NSIs and evaluate these modified algorithms on synthetic and real data sets.", "outcome": "The modified algorithms incorporating NSIs, applied to synthetic and real data sets, demonstrated promising quantitative and qualitative results.", "future_impact": "The technique of incorporating NSIs into graph clustering algorithms can potentially revolutionize the handling of large-scale relational data sets in graph clustering, by efficiently overcoming complexity limitations.", "venue": "ICML", "year": 2007, "title": "Graph clustering with network structure indices"} +{"pid": "6189e6d05244ab9dcb76e71a", "context": "Pretrained language models, while delivering strong performance for many NLP tasks, are very expensive to train in resources and time.", "key_idea": "The authors propose a learning framework named TLM that avoids the need for large-scale pretraining. Given some labeled task data and a large general corpus, TLM uses the task data as queries to retrieve a small subset from the general corpus and jointly optimizes the task objective and the language modeling objective from scratch.", "method": "The TLM is tested on eight classification datasets across four domains, and its results are compared to pretrained models like RoBERTa-Large.", "outcome": "On the tested datasets, TLM achieves results that are better or similar to pretrained language models such as RoBERTa-Large, while reducing the training FLOPs by two orders of magnitude.", "future_impact": "The authors believe the efficiency and effectiveness of TLM might contribute to democratizing NLP and expediting its development.", "venue": "ICML", "year": 2022, "title": "NLP From Scratch Without Large-Scale Pretraining: A Simple and Efficient Framework."} +{"pid": "61f360465aee126c0f7d4bb8", "context": "Optimization forms a critical part of various AI tasks, but the existing algorithms might get stuck in a local minimum or progress slower in shallow valleys.", "key_idea": "The authors propose a new optimization framework based on energy-conserving Hamiltonian dynamics characterized by chaotic regime. The approach uses a discretization of Born-Infeld dynamics with a squared relativistic speed limit dependent on the objective function.", "method": "The authors evaluate the proposed approach using tasks related to machine learning and PDE-solving tasks.", "outcome": "The proposed approach exhibits better performance, not stopping at high local minima and moving faster than GD+momentum in shallow loss valleys.", "future_impact": "This optimizer's capability of not stopping at local minima and traversing faster in shallow loss valleys could potentially improve the efficiency of machine learning and PDE-solving tasks.", "venue": "ICML", "year": 2022, "title": "Born-Infeld (BI) for AI: Energy-Conserving Descent (ECD) for Optimization."} +{"pid": "34c684cf-2ea5-4653-a68e-b4fa78a6270b", "context": "The presence of hidden state or partial observability in reinforcement learning (RL) problems often makes popular RL methods such as Q-learning and Sarsa ineffective. Most recent research has focused on using memory to estimate state to overcome this, but these methods are computationally very expensive and thus have limited applicability.", "key_idea": "The authors propose to use Sarsa(\u03bb), a family of reinforcement learning algorithms that use eligibility traces, on RL problems with hidden states that have good memoryless policies. This approach bypasses the need for computationally expensive state estimation.", "method": "The authors empirically validate this approach by applying conventional Sarsa(\u03bb) to four test problems previously explored by other researchers.", "outcome": "In all four test cases, Sarsa(\u03bb) was successfully able to find the best or a very good memoryless policy, reducing the computational expense of state estimation and demonstrating its applicability in reinforcement learning problems with hidden states.", "future_impact": "The successful demonstration of Sarsa(\u03bb) in the context of hidden-state reinforcement learning presents an alternate solution that can widen the applicability of RL algorithms on problems with hidden states and/or poor observability.", "venue": "ICML", "year": 1998, "title": "Using Eligibility Traces to Find the Best Memoryless Policy in Partially Observable Markov Decision Processes"} +{"pid": "cd6f9d25-0c46-49e7-bc44-2ae8773c5be8", "context": "Autism spectrum disorder (ASD) is a variety of developmental disorders that cause life-long communication and social deficits, but is typically only diagnosed as early as 2 years of age, despite early signs often emerging within the first year.", "key_idea": "The authors introduce a novel multi-kernel support vector machine (SVM) framework for identifying infants at high-risk for ASD at 6-months old, which is based on diffusion parameters derived from a hierarchical set of white matter (WM) connectomes.", "method": "The authors employ the proposed multi-kernel SVM framework to identify infants at high-risk for ASD and compare their results with those achieved by the best single connectome method.", "outcome": "The proposed method attains an accuracy of 76% in comparison to the 70% accuracy achieved with the best single-connectome method, suggesting that the multi-kernel SVM framework is more effective.", "future_impact": "The proposed framework provides essential imaging connectomic markers that can contribute to the evaluation of ASD risks as early as 6 months, which could potentially enable earlier intervention and treatment.", "venue": "ICML", "year": 2015, "title": "Identification of Infants at Risk for Autism Using Multi-parameter Hierarchical White Matter Connectomes"} +{"pid": "62997c085aee126c0f77cb78", "context": "Ensembles of deep neural networks significantly enhance model performance. For increased diversity of ensemble members, particle-based inference is a potentially promising approach from a Bayesian perspective. However, the application of these methods to neural networks often results in inefficiency due to over-parametrization or underfitting.", "key_idea": "The authors propose to optimize particles in the feature space, where activations of a certain intermediate layer reside, so as to mitigate the difficulties posed by over-parametrization or underfitting.", "method": "The proposed method was extensively evaluated on real-world datasets.", "outcome": "The model significantly outperforms the gold-standard Deep Ensembles in various metrics, including accuracy, calibration, and robustness.", "future_impact": "This methodology can enhance the robustness of the ensemble prediction by encouraging each ensemble member to capture distinct features.", "venue": "ICML", "year": 2022, "title": "Feature Space Particle Inference for Neural Network Ensembles."} +{"pid": "609a6b88e4510cd7c88b9a53", "context": "Multi-view clustering (MVC) aims to partition a set of multi-source data into their underlying groups. The performance of MVC can be improved by exploring better representation.", "key_idea": "The authors propose a deep matrix factorization model, incorporating features fusion and embedding neighbouring constraint to deal with sequential multi-view clustering problems.", "method": "The proposed model was tested with experiments to observe its performance in improving clustering, particularly in applications such as motion segmentation.", "outcome": "The experiments confirmed that the proposed model greatly improved the clustering performance.", "future_impact": "The proposed model can be applied in other applications such as motion segmentation, indicating its potential future utility in different data analysis fields.", "venue": "ICML", "year": 2020, "title": "Neighbouring Constraint Deep Matrix Factorization for Sequential Multi-view Clustering"} +{"pid": "1dedac8f-e7c9-49ae-913f-3e7cd5a96341", "context": "Recently, string kernels used for text classification have been shown to provide good results. However, they typically represent documents as individual letters or as a bag-of-words.", "key_idea": "This paper introduces the use of the string kernel in conjunction with syllables for text document representation, aiming to provide a more natural and less computationally intensive text representation. It also proposes new extensions to the string kernel: a lambda-weighting scheme and a method for 'soft matching'.", "method": "The authors incorporate syllables in the string kernel representation, and experiment with a novel lambda-weighting scheme and 'soft matching'. They test these approaches on text categorisation problems.", "outcome": "Experimental results demonstrate that syllables can be effectively used in text categorization problems, and the proposed lambda-weighting and 'soft-matching' techniques are successful.", "future_impact": "The proposed approaches potentially provide an effective and efficient means of text representation, influencing the field of text categorization and other applications where the significance of certain symbols is known to be less.", "venue": "ICML", "year": 2002, "title": "Syllables and other String Kernel Extensions"} +{"pid": "9922039b-4f31-4d48-a09c-efb8ac9b0809", "context": "Inductive Logic Programming (ILP) involves the construction of first-order definite clause theories from examples and background knowledge. Traditional ILP systems have successful applications across various domains but have comparatively weaker PAC-learning results, which has been argued to be due to distributional assumptions made in application domains.", "key_idea": "The paper introduces the idea of modeling these assumptions as a Bayesian prior probability representing subjective degrees of belief. Along with this, it incorporates Bayesian prior distributions over time-bounded hypotheses in PAC, leading to a new model called U-learnability.", "method": "The author proposes testing this idea through time-bounded logic programs, enforcing decidability and allowing a unified characterisation of speed-up learning and inductive learning.", "outcome": "The paper concludes that time-bounded logic programs are polynomially U-learnable under certain distributions. The proposed model U-learnability includes PAC and Natarajan's model of speed-up learning as special cases.", "future_impact": "The concept of U-learnability is argued to be more appropriate than PAC for Universal (Turing computable) languages, potentially changing how future learnability is approached in ILP.", "venue": "ICML", "year": 1994, "title": "Bayesian inductive logic programming"} +{"pid": "60754491e4510cd7c836f8d7", "context": "The advent of the \u201cInternet + Education\u201d era has allowed for smarter and more effective learning methodologies. English education, a key aspect of school education, can particularly benefit from this development.", "key_idea": "This paper suggests the creation of a personalized, experiential teaching ecosystem in English education, utilizing the advantages of the 'Internet + Education' approach. To improve learning efficiency, it proposes using multiple coexistent learning modes, facilitating two-way online-offline interaction, and building a teacher-student communication platform.", "method": "The author establishes experimental and control classes and carries out teaching separately in each. Examinations are conducted before and after the teaching sessions, and a questionnaire is sent at the end of the period to assess satisfaction.", "outcome": "The proposed Internet + Education methodology was superior to traditional teaching in terms of learning effect and students\u2019 satisfaction with learning.", "future_impact": "The study suggests that the proposed methodology reduces the workload on teachers and allows them more time and energy to educate students, indicating potential improvements in the quality of education through efficient resource allocation.", "venue": "ICML", "year": 2020, "title": "Construction Strategy of Smart English Teaching Platform from the Perspective of \u201cInternet + Education\u201d"} +{"pid": "acc9cd6a-5b83-4e10-b8a0-e31c433f05ea", "context": "Spectral methods propose solutions to inferring weighted automata from samples of variable-length strings drawn from unknown target distributions, relying on a singular value decomposition of a Hankel matrix. Existing concentration bounds suggest that the concentration over the mean Hankel matrix gets looser with its size.", "key_idea": "The authors propose new dimension-free concentration bounds for several variants of Hankel matrices, arguing that the concentration rate of the Hankel matrix around its mean should not be an argument for limiting its size.", "method": "The authors conduct experiments to test the proposed dimension-free concentration bounds for various Hankel matrix variants.", "outcome": "The experiments demonstrate that the proposed bounds are tight and they significantly improve existing bounds.", "future_impact": "The results may encourage further research to not limit the size of a Hankel matrix due to the concentration rate around its mean.", "venue": "ICML", "year": 2014, "title": "Dimension-free Concentration Bounds on Hankel Matrices for Spectral Learning"} +{"pid": "9b159cc8-de78-4714-806f-8e0169b5062c", "context": "Kernel methods have gained popularity, thereby increasing the need to automate the task of specifying a suitable kernel. The problem of Multiple Kernel Learning (MKL), i.e., finding a combination of prespecified base kernels suitable for the task at hand, has been thoroughly studied.", "key_idea": "This paper proposes the framing of the MKL problem as a standard binary classification problem with additional constraints to ensure the positive definiteness of the learned kernel.", "method": "The study tests the proposed approach through experiments on nine different datasets from various domains.", "outcome": "Despite its simplicity, the proposed technique was shown to compare favorably with current leading MKL approaches on tested datasets.", "future_impact": "The new MKL framework has the potential to leverage extensive research in binary classification which would lead to development of better performing and scalable MKL algorithms that are simpler and more understandable for practitioners.", "venue": "ICML", "year": 2012, "title": "A Binary Classification Framework for Two-Stage Multiple Kernel Learning"} +{"pid": "5f02f0da91e011ee5e0258b1", "context": "Choosing the optimizer for deep learning tasks is often a challenging decision to make due to the presence of hundreds of optimization methods in the literature. Currently, it is done mostly according to personal anecdotes due to the lack of clear theoretical guidance and definite empirical evidence.", "key_idea": "The authors propose a standardized benchmark method for more than a dozen popular deep learning optimization algorithms to replace personal anecdotes with heuristics.", "method": "The authors conduct an extensive benchmark of optimization techniques through almost 35,000 individual ran tests across different tasks.", "outcome": "No single optimization method was found to clearly dominate across all tested tasks. However, a reduced subset of specific algorithms and parameter choices generally provided competitive results in their experiments.", "future_impact": "The open-sourced results from the authors' experiments can serve as well-tuned baselines for evaluating novel optimization methods in future studies, potentially reducing the computational efforts required in such tasks.", "venue": "ICML", "year": 2021, "title": "Descending through a Crowded Valley -- Benchmarking Deep Learning Optimizers"} +{"pid": "5ede0553e06a4c1b26a83f03", "context": "Computational equilibrium finding in large zero-sum extensive-form imperfect-information games has been significantly advanced by new forms of counterfactual regret minimization. However, solving the equilibrium-finding problem as a linear program (LP) has not been scalable due to the memory requirements of LP solvers and complexity.", "key_idea": "The authors propose a different approach to the equilibrium-finding problem. They give an efficient practical algorithm that factors a large payoff matrix into a product of two matrices that are typically dramatically sparser, reformulating the problem as a linear program with size only a logarithmic factor worse than counterfactual regret minimization.", "method": "The authors test the algorithm with experiments on poker endgames, comparing it to modern variants of counterfactual regret minimization in solving large extensive-form games.", "outcome": "The experiments demonstrated that modern linear program solvers are competitive against even game-specific modern variants of counterfactual regret minimization in solving large extensive-form games, and can be used to compute exact solutions unlike iterative algorithms like counterfactual regret minimization.", "future_impact": "The results suggest a potential increase in the use of LP solvers in computational equilibrium finding in large zero-sum extensive-form imperfect-information games, bringing in the possibility of computing exact solutions, which is not possible with iterative methods like CFR.", "venue": "ICML", "year": 2020, "title": "Sparsified Linear Programming for Zero-Sum Equilibrium Finding"} +{"pid": "5d8dded23a55acd1b549693c", "context": "Deep neural networks have achieved success in many fields but transferring these architectures to irregularly-sampled and asynchronous time series, like those found in many real-world datasets such as healthcare applications, remains a challenge, mainly due to unaligned measurements.", "key_idea": "The authors propose SEFT (Set Functions for Time Series), a novel framework based on differentiable set function learning for classifying irregularly sampled time series that focuses on high scalability and data efficiency.", "method": "The authors benchmarked SEFT against other methods for classifying time series data on multiple healthcare datasets.", "outcome": "SEFT performed competitively with existing methods while significantly reducing the runtime.", "future_impact": "SEFT has potential to handle large datasets and could be applied to online monitoring scenarios, which could bring improvements in healthcare and many other fields dealing with large, irregularly-sampled time series.", "venue": "ICML", "year": 2020, "title": "Set Functions for Time Series"} +{"pid": "62aa9fb55aee126c0fa5ca00", "context": "Advancements in cooperative multi-agent reinforcement learning (MARL) are largely based on common design principles, namely, value decomposition and parameter sharing. These principles typically allow for centralized training and decentralized execution (CTDE), which are efficient in practice.", "key_idea": "The authors re-examine these conventional approaches, arguing that in certain scenarios such as environments with highly multi-modal reward landscapes, value decomposition and parameter sharing may lead to undesired results.", "method": "The authors present a theoretical analysis of policy gradient (PG) methods with individual policies, comparing their performance against traditional MARL approaches. They validate their findings through empirical testing in a variety of domains, from simple matrix and grid-world games to complex benchmarks like StarCraft Multi-Agent Challenge and Google Research Football.", "outcome": "Their analysis reveals that in circumstances where standard MARL principles fail, policy gradient methods with individual policies can reliably converge to an optimal solution. This finding provides theoretical support for recent empirical observations of the effectiveness of PG in various MARL testbeds.", "future_impact": "The authors anticipate that their insights and practical suggestions for implementing multi-agent PG algorithms will guide future work in the development of more versatile and powerful MARL algorithms.", "venue": "ICML", "year": 2022, "title": "Revisiting Some Common Practices in Cooperative Multi-Agent Reinforcement Learning."} +{"pid": "5ede0553e06a4c1b26a83f1f", "context": "In systems and computational neuroscience, it is important to understand how neural circuits accumulate evidence towards a decision. Current approaches limit the number of decision-making models that can be fitted to neural data.", "key_idea": "The authors propose a general framework that includes the canonical drift-diffusion model for modeling neural activity during decision-making. This framework allows for extensions such as multi-dimensional accumulators, variable and collapsing boundaries, and discrete jumps.", "method": "The framework is based on constraining the parameters of recurrent state space models, for which a scalable variational Laplace EM inference algorithm is introduced. The authors apply this modeling approach to spiking responses recorded from monkey parietal cortex during two decision-making tasks.", "outcome": "The proposed model found that a two-dimensional accumulator better captured the responses of a set of parietal neurons than a single accumulator model. It also identified a variable lower boundary in the responses of a parietal neuron during a random dot motion task.", "future_impact": "The authors expect this framework to be useful for modeling neural dynamics in a variety of decision-making settings.", "venue": "ICML", "year": 2020, "title": "A general recurrent state space framework for modeling neural dynamics during decision-making"} +{"pid": "62abf1355aee126c0f475857", "context": "The problem of online learning with graph feedback has been extensively studied, with existing works mainly studying the adversarial and stochastic feedback separately. If the knowledge of the feedback mechanism is unavailable or incorrect, the resulting algorithms could suffer great loss. Some previous works aimed to optimize for both environments, but these made assumptions that compromised the generality of the framework and may not be applicable in real-world scenarios.", "key_idea": "The authors propose a new trade-off mechanism with a carefully-designed exploration and exploitation ratio that can work for both adversarial and stochastic feedback with a general feedback graph, without requiring prior knowledge of the feedback mechanism.", "method": "The authors test their proposed algorithm within stochastic and adversarial settings, where the regret is evaluated.", "outcome": "The proposed algorithm achieves polylogarithmic regret in a stochastic setting and the minimax-optimal regret of approximately O(T^2/3) in an adversarial setting, where T is the horizon.", "future_impact": "The authors propose a best-of-both-worlds result for general feedback graphs, which has the potential to improve online learning with general graph feedback.", "venue": "ICML", "year": 2022, "title": "Simultaneously Learning Stochastic and Adversarial Bandits with General Graph Feedback."} +{"pid": "6269f8245aee126c0f049e79", "context": "In traditional goal-conditioned reinforcement learning (RL), an agent is provided with the exact goal they intend to reach, which isn't realistic as the configuration of the goal is often unknown before performing a task.", "key_idea": "The authors propose a new form of state abstraction, a goal-conditioned bisimulation, which captures functional equivariance, enabling the reuse of skills to achieve new goals.", "method": "A metric form of this new form of state abstraction, goal-conditioned bisimulation, was learned and tested to generalize to new goals in real-world manipulation tasks.", "outcome": "The authors demonstrated that their proposed state abstraction, goal-conditioned bisimulation, can generalize to new goals in real-world manipulation tasks.", "future_impact": "The authors claim that the learned representation is not solely for goal-conditioned tasks, but can be used for any downstream task described by a state-only reward function, leading to broader possible applications in the field of reinforcement learning.", "venue": "ICML", "year": 2022, "title": "Bisimulation Makes Analogies in Goal-Conditioned Reinforcement Learning."} +{"pid": "1e22ffaa-b9fb-44d4-83e0-4cc465dffc49", "context": "The controllability of complex networks, which model many real-life systems, has been studied mainly to judge whether the network is controllable or not and how to select the appropriate driver nodes. However, these networks are often vulnerable, leading to a decline in controllability.", "key_idea": "The authors propose a method based on link prediction to make complex networks more robust to different modes of attacks, improving their controllability.", "method": "The authors validate the effectiveness of their proposed method through experiments.", "outcome": "Experiments validated the effectiveness of the proposed method, though the specific results are not mentioned in the abstract.", "future_impact": "The proposed method using link prediction to boost network robustness and controllability may enable easier control of complex network-based systems in real-world applications.", "venue": "ICML", "year": 2019, "title": "Improving Complex Network Controllability via Link Prediction"} +{"pid": "0df52370-86fc-4646-9f9b-5465c8a73934", "context": "Batch Reinforcement Learning (Batch RL) uses trajectories collected with a behaviour policy for training a new policy. Safe policy improvement (SPI) ensures with high probability that the trained policy performs better than the behavioural policy.", "key_idea": "An improvement to the Safe policy improvement with Baseline Bootstrapping algorithm (SPIBB), allows policy search over a wider set of policies by adopting a softer strategy that controls the error in the value estimates according to the local model uncertainty.", "method": "The authors propose two algorithms (one optimal and one approximate) to solve this constrained optimization problem and evaluate their approach on finite and infinite MDPs using a neural network function approximation.", "outcome": "The new approach takes more risks on uncertain actions while remaining provably-safe, showing a significant improvement over existing SPI algorithms on both finite and infinite MDPs.", "future_impact": "The improved SPIBB approach enables less conservative policy improvement in Reinforcement Learning, thus may influence future advancements in SPI algorithms.", "venue": "ICML", "year": 2019, "title": "Safe Policy Improvement with Baseline Bootstrapping"} +{"pid": "60f16ace91e011963c8d3e8c", "context": "The success of modern machine learning models on large datasets depends on extensive computational resources which have high financial and environmental costs. Extracting subsets that generalize as well as the full data is one way to approach this issue.", "key_idea": "The authors propose a general framework, GRAD-MATCH, that finds subsets which closely match the gradient of the training or validation set. The subsets are found using an orthogonal matching pursuit algorithm.", "method": "The orthogonal matching pursuit algorithm is rigorously tested through theoretical convergence guarantees and extensive experiments on real-world datasets.", "outcome": "GRAD-MATCH significantly and consistently outperforms several recent data-selection algorithms and achieves the best accuracy-efficiency trade-off.", "future_impact": "GRAD-MATCH, which optimizes the accuracy-efficiency trade-off, can potentially reduce the financial and environmental costs of training deep learning models. It is freely available as a part of the CORDS toolkit on GitHub, suggesting its potential for widespread adoption and impact on efficient model training.", "venue": "ICML", "year": 2021, "title": "Grad-Match: Gradient Matching Based Data Subset Selection For Efficient Deep Model Training"} +{"pid": "61f8a4c35aee126c0fee01d6", "context": "Self-supervised Contrastive Learning (CL) has been recently shown to be very effective in preventing deep networks from overfitting noisy labels, but the theoretical understanding of this effect is limited.", "key_idea": "The study proposes a theoretical argument that the representation matrix learned by contrastive learning boosts robustness through two properties: one prominent singular value for each sub-class in the data and a large alignment between the prominent singular vectors and the clean labels of each sub-class.", "method": "The authors use mathematical reasoning to prove the theoretical arguments and then experiment with deep networks with noisy labels, comparing the performance of those pretrained with contrastive learning and those without.", "outcome": "Fine-tuned deep networks pretrained with contrastive learning perform better initially, and their initial robustness enables robust training methods to achieve state-of-the-art performance under extreme noise levels (average increases in accuracy of 27.18% and 15.58% on CIFAR-10 and CIFAR-100 with 80% symmetric noisy labels, and 4.11% increase on WebVision).", "future_impact": "This improved understanding of why Contrastive Learning benefits robustness against label noise could inform future methods and approaches in the field, particularly regarding training models under high noise levels.", "venue": "ICML", "year": 2022, "title": "Investigating Why Contrastive Learning Benefits Robustness against Label Noise."} +{"pid": "30ad7fc5-532c-4de3-a6a8-de436b4dd381", "context": "Predicting the probability distribution of a variable rather than its most likely value can be more useful in many applications, such as meteorology and finance. However, prediction becomes complex when the distribution of the phenomenon is significantly different from a normal distribution, as in the case of surf height data.", "key_idea": "The study proposes an ensemble of mixture density networks to predict the probability density function of surf height, with the objective of predicting if it will fall within a 'surfable' range.", "method": "The authors employed an ensemble of mixture density networks to examine surf data and predict the probability density function.", "outcome": "The evaluation proved the effectiveness of using an ensemble of mixture density networks for predicting the probability density function of surf height data.", "future_impact": "Although not directly stated, the successful implementation of the proposed method implies that it can be potentially used to improve predictions in areas where predicting the probability distribution of a variable is important, beyond just predicting most likely values.", "venue": "ICML", "year": 2005, "title": "Predicting probability distributions for surf height using an ensemble of mixture density networks"} +{"pid": "6201df4a5aee126c0f64dcd3", "context": "Data poisoning attacks aim to manipulate model behaviors by distorting the training data. A previously proposed defense to this is Deep Partition Aggregation (DPA), which mitigates this threat by predicting through an aggregation of base classifiers trained on disjoint subsets of data.", "key_idea": "The authors propose an improved certified defense against general poisoning attacks called Finite Aggregation. Unlike DPA which directly splits the training set into disjoint subsets, Finite Aggregation first splits the training set into smaller disjoint subsets and then combines duplicates to build larger (but not disjoint) subsets for training base classifiers.", "method": "The proposed method is ordinarily tested on MNIST, CIFAR-10, and GTSRB datasets to measure its effectiveness.", "outcome": "Empirical results showed that the proposed Finite Aggregation consistently improves certificates on MNIST, CIFAR-10, and GTSRB, boosting certified fractions by up to 3.05%, 3.87% and 4.77%, respectively, while keeping the same clean accuracies as DPA\u2019s.", "future_impact": "The approach of Finite Aggregation effectively establishes a new state of the art in pointwise certified robustness against data poisoning, potentially providing an improved defense mechanism in applications susceptible to data poisoning attacks.", "venue": "ICML", "year": 2022, "title": "Improved Certified Defenses against Data Poisoning with (Deterministic) Finite Aggregation."} +{"pid": "51678644-eb4b-48b7-99a0-73a5c140b00d", "context": "In the current Multiple Input Multiple Output (MIMO) systems, multiple-symbol differential detection algorithms require large storage space and have high complexity.", "key_idea": "The authors suggest a Probabilistic Sorting Memory Constrained Tree Search Algorithm (PSMCTS) that uses the performance advantage of a sorting algorithm and the storage advantage of a Memory Constrained Tree Search (MCTS). A pruning PSMCTS (PPSMCTS) which further reduces computational complexity is also proposed.", "method": "The study uses simulations to test the proposed PSMCTS and PPSMCTS algorithms under fixed memory situations, small storage capacity conditions, and low signal noise ratio (SNR) regions.", "outcome": "Simulation results show that the performance of PSMCTS is comparable to that of the ML algorithm, with a lower computational complexity than the MCTS algorithm in small storage capacity conditions. The PPSMCTS outperformed the PSMCTS in reducing computational complexity.", "future_impact": "Theoretical analysis and simulation prove that the two proposed algorithms can effectively retain the beneficial feature of the MCTS algorithm, suggesting their suitability for hardware implementation.", "venue": "ICML", "year": 2018, "title": "Probabilistic Sorting Memory Constrained Tree Search Algorithm for MIMO System"} +{"pid": "486f2f8b-3854-4858-a9e7-733948437f44", "context": "The huge traffic generated by a growing number of network applications is increasing the burden of network management. Current traffic classification methods have unstable recognition rates and high computational complexity, which affects network traffic management and application category monitoring.", "key_idea": "The authors propose a new method of using deep learning technology, specifically a model based on Deep Belief Network (DBN), to classify network applications, and construct a model suitable for this task using the Tensorflow framework.", "method": "The authors compare the classification performances of the DBN-based model and the Backpropagation (BP)-based model on real data sets.", "outcome": "The experimental results show that the applications classification model based on DBN has higher classification accuracy for Peer-to-Peer (P2P) applications.", "future_impact": "The ability to accurately identify and categorize network applications based on deep learning technology can improve understanding of network traffic conditions, people\u2019s online behavior, and accelerate the development of the Internet.", "venue": "ICML", "year": 2018, "title": "A Research of Network Applications Classification Based on Deep Learning"} +{"pid": "5ec3ae5291e0112b16089fe2", "context": "Causal feature selection in time series with latent variables is a complex task, especially in the identification of direct and indirect causes.", "key_idea": "The authors propose a constrained-based causal feature selection method for time series with latent variables, which takes into consideration graph constraints and only requires two conditional independence tests for each observed candidate time series.", "method": "They prove the soundness and completeness of their method using theoretical proofs and conduct practical experiments using a wide range of simulated and real data.", "outcome": "Their method proves sound and complete under specific graph constraints, and in experiments, it showed very low false positive rates and relatively low false negative rates.", "future_impact": "This method can provide a more accurate causal feature selection in time series with latent variables, possibly leading to more accurate predictions and data analysis.", "venue": "ICML", "year": 2021, "title": "Necessary and Sufficient Conditions for Causal Feature Selection in Time Series with Latent Common Causes"} +{"pid": "200dc485-0608-45ba-ac40-86a5d7ab13d4", "context": "In reinforcement learning, model-free approaches address the task by directly mapping external and internal states to actions while model-based methods attempt to construct a model of the environment and then select optimal actions based on that model.", "key_idea": "The authors propose a novel algorithm that integrates both model-free and model-based approaches into a single algorithm, switching between both modes depending on the current environmental state and the status of the agent's knowledge.", "method": "The authors prove the efficiency of the approach theoretically. They also conduct numerical simulations to demonstrate the effectiveness of the algorithm.", "outcome": "The proposed approach improves performance whenever environmental knowledge is available, without compromising performance when such knowledge is absent. Numerical simulations show the effectiveness of the approach.", "future_impact": "The approach may be efficient in enhancing policy gradient learning, suggesting possible future applications.", "venue": "ICML", "year": 2011, "title": "Integrating Partial Model Knowledge in Model Free RL Algorithms"} +{"pid": "7c623c10-4960-4fec-9054-03c7e1b42ec5", "context": "The relationship between the motivation of using Internet language and the intercultural communication ability of college students needed examination.", "key_idea": "This study investigates whether and how the intercultural communication ability of college students can be improved through their motivations in using internet language.", "method": "The authors employed variance analysis, Pearson correlation analysis, and regression analysis to study and analyze the data collected from 215 college students in Jilin Province.", "outcome": "The study identified practical motivation as the most important factor in college students' use of Internet language. There is a significant positive correlation between intercultural communication ability and the motivation of using Internet language with the latter explaining 75.2% of the former. Students had strong cultural awareness but lack real-life practice of intercultural communication.", "future_impact": "The motivation of using Internet language can have a substantial impact on intercultural communication ability, suggesting that intercultural communication skills could potentially be improved by stimulating practical motivation.", "venue": "ICML", "year": 2019, "title": "An Empirical Study on the Relationship between Using Motivation of Internet Language and Intercultural Communication Competence of College Students"} +{"pid": "61f360c55aee126c0f7d5fdc", "context": "Despite reliable evaluation benchmarks driving progress in machine learning, there is a lack of a multilingual benchmark in vision-and-language research which has thus mostly focused on English language tasks.", "key_idea": "The authors propose the Image-Grounded Language Understanding Evaluation (IGLUE) benchmark which, by aggregating pre-existing datasets and creating new ones, offers a variety of tasks in 20 diverse languages to evaluate multilingual multimodal models for transfer learning.", "method": "The benchmarks are used to evaluate multilingual multimodal models for transfer learning, both in a zero-shot setting and newly defined few-shot learning setups. Various state-of-the-art models were evaluated.", "outcome": "The results show that translate-test transfer is superior to zero-shot transfer and that few-shot learning is difficult to implement for many tasks. It is also revealed that downstream performance is partially explained by the amount of available unlabelled textual data for pretraining, and only weakly by the typological distance of target\u2013source languages.", "future_impact": "The new IGLUE benchmark is hoped to encourage future research in the field of multilingual and multimodal machine learning.", "venue": "ICML", "year": 2022, "title": "IGLUE: A Benchmark for Transfer Learning across Modalities, Tasks, and Languages."} +{"pid": "6269f8235aee126c0f049b0d", "context": "Aquatic locomotion is a classic fluid-structure interaction (FSI) problem of interest to biologists and engineers that is computationally expensive to solve. Optimizing robotic swimmer design generally involves cumbersome and expensive gradient-free procedures on top of costly simulations.", "key_idea": "The authors present a novel, fully differentiable hybrid approach to solving the FSI problem that combines a 2D direct numerical simulation for the deformable solid structure of the swimmer and a physics-constrained neural network surrogate model to capture the hydrodynamic effects of the fluid.", "method": "The deformable solid simulation of the swimmer\u2019s body uses state-of-the-art techniques from computer graphics to speed up the finite-element method (FEM). The fluid simulation uses a U-Net architecture trained with a physics-based loss function to predict the flow field at each time step, and the outputs from the neural network are sampled using an immersed boundary method (IBM) to compute the swimming motion accurately and efficiently. The authors demonstrate the efficiency and differentiability of the hybrid simulator on a 2D carangiform swimmer.", "outcome": "The authors demonstrate the computational efficiency and differentiability of their proposed hybrid simulator on a 2D carangiform swimmer.", "future_impact": "Due to its differentiability, the simulator can potentially be used for computational design of controls for soft bodies immersed in fluids via direct gradient-based optimization, which should provide a more efficient and effective means of robotic swimmer design.", "venue": "ICML", "year": 2022, "title": "Fast Aquatic Swimmer Optimization with Differentiable Projective Dynamics and Neural Network Hydrodynamic Models."} +{"pid": "bee9374e-891b-4858-9e2e-722221699aa3", "context": "The specific non-linear relations among the dataset in the context of Support Vector Machines (SVMs) can prevent the proper definition of the margin and affect the model\u2019s accuracy.", "key_idea": "The authors propose a novel method called 'hyperplane folding' that manipulates the dataset by splitting, rotating, and merging to increase the margin in SVMs based on the location of the support vectors.", "method": "The authors perform a small experiment with three folding iterations on 3-dimensional data points with non-linear relations to validate the method, using a standard SVM implementation and some basic manipulation of the data points.", "outcome": "The hyperplane folding method resulted in increased margin and improved accuracy of the model in treating non-linear relations in the SVMs.", "future_impact": "The authors anticipate that the hyperplane folding method could help improve the interpretability of the data in SVMs.", "venue": "ICML", "year": 2019, "title": "Handling Non-linear Relations in Support Vector Machines through Hyperplane Folding"} +{"pid": "5e2ac03c3a55ac8999c1ada2", "context": "Deep Learning representations of data produced by generative adversarial nets (GANs) are complex to understand and analyze.", "key_idea": "The paper introduces the idea that deep learning representations of GAN-data are random vectors which fall into the class of concentrated random vectors, and they can be described by their first two statistical moments for a wide range of standard classifiers.", "method": "The authors validate their theory by generating images with the BigGAN model and testing across different popular deep representation networks.", "outcome": "The results validate the theoretical findings that the deep learning representations of GAN-data behave almost as if they were drawn from a Gaussian mixture.", "future_impact": "The results can lead to better understanding of how GAN generated deep learning representations behave and can be characterized.", "venue": "ICML", "year": 2020, "title": "Random Matrix Theory Proves that Deep Learning Representations of GAN-data Behave as Gaussian Mixtures"} +{"pid": "5485799d-6e2c-46d5-aa91-2bec5d08e7dd", "context": "Semi-supervised learning is a popular framework for controlling labeling cost while enhancing the accuracy of models, but quantifying the asymptotic accuracy of generative semi-supervised learning remains a challenge.", "key_idea": "The authors extend the concept of stochastic composite likelihood to quantify the asymptotic accuracy of generative semi-supervised learning.", "method": "The authors' methodology is demonstrated through simulation studies and real world experiments using naive Bayes for text classification and Markov random fields (MRF) and Conditional random fields (CRF) for structured prediction in Natural Language Processing (NLP).", "outcome": "The authors succeed in quantifying the asymptotic accuracy of generative semi-supervised learning while successfully measuring the value associated with different labeling policies.", "future_impact": "The framework proposed in this paper has the potential to help answer the critical question of how much data to label and in what manner, which could significantly influence future semi-supervised learning practices.", "venue": "ICML", "year": 2010, "title": "Asymptotic Analysis of Generative Semi-Supervised Learning"} +{"pid": "5ede0553e06a4c1b26a84159", "context": "The trace regression model, a direct extension of the well-studied linear regression model, allows one to map matrices to real-valued outputs. However, it could not handle a more complex scenario of mapping matrices to a matrix output.", "key_idea": "The paper introduces the partial-trace regression model, a family of linear mappings from matrix-valued inputs to matrix-valued outputs, subsuming the trace regression model and the linear regression model, supported by a learning framework using low-rank Kraus representation of completely positive maps to learn partial trace regression models from data.", "method": "The model's performance is validated through synthetic and real-world experiments conducted for both matrix-to-matrix regression and positive semidefinite matrix completion tasks, which are both formulated as partial trace regression problems.", "outcome": "The paper demonstrates relevance and validates the framework through experiments, but no specific quantitative results or metrics are given in the abstract.", "future_impact": "The ability to perform matrix-to-matrix regression and positive semidefinite matrix completion through this model may open up new opportunities for handling more complex data forms in scientific data processing.", "venue": "ICML", "year": 2020, "title": "Partial Trace Regression and Low-Rank Kraus Decomposition"} +{"pid": "dc1543ba-13a6-4599-b500-53130cb3f4d7", "context": "Label propagation is a popular semi-supervised learning framework. However, current label propagation algorithms require computation of an inverse matrix, resulting in high computational cost.", "key_idea": "The paper proposes an efficient label propagation algorithm which iteratively computes lower and upper bounds of labeling scores to prune unnecessary score computations, providing exactly the same labeling results as those yielded by optimal labeling scores.", "method": "The algorithm is validated through comparison with existing label propagation methods in terms of computational efficiency.", "outcome": "The results demonstrate that the proposed algorithm significantly outperforms existing label propagation methods by reducing computation costs.", "future_impact": "The proposed efficient label propagation algorithm can provide avenues for improving the performance of semi-supervised learning tasks in terms of computation time and cost.", "venue": "ICML", "year": 2014, "title": "Efficient Label Propagation"} +{"pid": "f2eed3f6-cc6b-4973-b769-f660f2c1da5d", "context": "The lasso procedure has been the target of substantial research over the last fifteen years, with many known results about its behavior for a fixed or optimally chosen smoothing parameter. However, less is known about the lasso's behavior when the smoothing parameter is chosen in a data dependent manner.", "key_idea": "This paper aims to uncover the behavior of the lasso procedure when the smoothing parameter is chosen via cross-validation, in a high-dimensional setting.", "method": "The study considers a high-dimensional setting wherein the number of predictors p = n\u03b1, \u03b1 > 0, grows with the number of observations, to assess the risk consistency of the lasso when the smoothing parameter is determined through cross-validation.", "outcome": "The paper provides the first result about the risk consistency of lasso when the smoothing parameter is chosen via cross-validation in a high-dimensional setting.", "future_impact": "This discovery could lead to improvements in the application of the lasso procedure, particularly in situations where the smoothing parameter needs to be determined in a data-dependent manner.", "venue": "ICML", "year": 2013, "title": "The lasso, persistence, and cross-validation"} +{"pid": "5ede0553e06a4c1b26a83e61", "context": "The problem of attributing a model\u2019s prediction to its base features is well-studied, and the Shapley value is commonly used for such attribution. However, extending this to feature interactions still remains a challenge.", "key_idea": "The authors propose a generalization of the Shapley value, known as Shapley-Taylor index, which attributes the model's prediction to interactions of subsets of features up to some size 'k', analogous to the truncated Taylor Series decomposition approach.", "method": "The Shapley Taylor index is axiomatized using standard Shapley axioms - linearity, dummy, symmetry and efficiency - and an additional interaction distribution axiom. The proposed method is contrasted against the Shapley Interaction index from cooperative game theory literature and applied to three models.", "outcome": "Through the application of the Shapley-Taylor index to three models, the authors are able to identify interesting qualitative insights.", "future_impact": "The Shapley-Taylor index's ability to attribute a model's prediction to interactions of subsets of features could provide more robust interpretation of complex models and impact future studies on model interpretation.", "venue": "ICML", "year": 2020, "title": "The Shapley Taylor Interaction Index"} +{"pid": "37d84937-e72d-4c25-a657-4e4c3619b53c", "context": "Pharmacological magnetic resonance imaging (phMRI) data, where the blood oxygen level-dependent (BOLD) response to an acute drug challenge is measured, has typically been analysed using general linear models applied to each voxel individually. This requires a pre-defined model of the expected response to the pharmacological stimulus which is often defined using pharmacokinetic profiles, phMRI data from pilot studies, cognitive or physiological variables, or a simple pre-post boxcar profile.", "key_idea": "The authors propose a data-driven approach to modeling the temporal profile of phMRI data, fitting a basis function to the data in a Bayesian framework using Gaussian processes.", "method": "The proposed data-driven modeling of BOLD response curves is compared with two previous multivariate approaches to fMRI analysis.", "outcome": "The Bayesian Gaussian process modeling method was found to outperform the two previous multivariate approaches while also providing information about the shape of the BOLD response, which enhances model interpretability.", "future_impact": "The data-driven approach presented in the paper may lead to enhanced modeling and interpretation of fMRI data in future neuroimaging studies.", "venue": "ICML", "year": 2011, "title": "Data-driven modeling of BOLD drug response curves using Gaussian process learning"} +{"pid": "627332775aee126c0f18d52d", "context": "Recent investigations in noise contrastive estimation suggest that, both empirically and theoretically, increasing the number of 'negative samples' in the contrastive loss initially improves downstream classification performance. However, beyond a certain threshold, it impairs performance due to a 'collision-coverage' trade-off.", "key_idea": "Despite common understanding, the authors theorize that in a contrastive learning setting, the downstream performance of the representation optimizing the (population) contrastive loss does not deteriorate with the number of negative samples.", "method": "The authors theoretically explain their idea in a simple setting where positive pairs are generated by sampling from the latent class. They also provide empirical support for their theoretical results on CIFAR-10 and CIFAR-100 datasets.", "outcome": "Their theoretical results establish that the downstream performance of the representation optimizing contrastive loss does not degrade with the number of negative samples, which challenges conventional wisdom. Results are further supported by empirical evidence from experiments on CIFAR-10 and CIFAR-100 datasets.", "future_impact": "Results questioning the previously held belief about the adverse effect of 'negative samples' in contrastive learning could lead to reconsideration and adjustments of current learning methodologies.", "venue": "ICML", "year": 2022, "title": "Do More Negative Samples Necessarily Hurt In Contrastive Learning?"} +{"pid": "e442388c-9c39-462a-a75e-9bcbac626958", "context": "The rapid development of recommendation systems has highlighted the importance of predicting user behavior accurately.", "key_idea": "The authors propose the use of a random forest as a method for recommending insurance products.", "method": "A comparative study is conducted where the authors apply the Random Forest method to recommendation of insurance products and compared it with ID3, C4.5, Naive-Bayes, and Nearest-neighbor algorithms.", "outcome": "The experimental results show that the prediction error of the Random Forest method is lower than ID3 (2.02%), C4.5 (1.09%), Naive-Bayes (1.67%), and Nearest-neighbor (5.97%).", "future_impact": "It has been shown to be highly feasible to recommend insurance products with the Random Forest method.", "venue": "ICML", "year": 2019, "title": "Research on Recommendation of Insurance Products Based on Random Forest"} +{"pid": "85221277-e0f5-4d72-a8df-7aac303b34f2", "context": "Distance-based methods in machine learning and pattern recognition rely on a metric distance between points in the input space, and this metric typically is specified a priori.", "key_idea": "The paper proposes learning metrics from data instead of specifying them in advance, utilizing kernel methods and multidimensional scaling (MDS) techniques. The authors present a specific family of discriminant kernels that induce Euclidean and Fisher separable metrics.", "method": "The framework employs multidimensional scaling techniques to find the low-dimensional Euclidean representations of the induced metrics.", "outcome": "The resulting feature vectors, which include information from input data and corresponding labels, have Fisher separability, which prepares them well to be utilized in distance-based classifiers.", "future_impact": "Learning metrics from data with this method could change the way distance-based classifiers are developed and used, potentially improving their efficiency and effectiveness in pattern recognition applications.", "venue": "ICML", "year": 2003, "title": "Learning metrics via discriminant kernels and multidimensional scaling: toward expected Euclidean representation"} +{"pid": "810da6bd-9173-4b7d-8a2e-1cac0ba5f53a", "context": "Current methods for learning operator effects from training examples may struggle with noisy and nondeterministic domains, and may not integrate efficiently with case-based approaches.", "key_idea": "The authors propose an algorithm called Continuous Conceptual Set Covering (CCSC) which uses engineering knowledge to learn operator effects in noisy and non-deterministic scenarios.", "method": "An empirical evaluation of the CCSC algorithm was undertaken using a tray-tilting domain, and its performance was compared against a case-based approach.", "outcome": "In the study, the CCSC was found to learn faster than the case-based alternative. However, the best results were obtained by integrating the CCSC with the case-based approach.", "future_impact": "Integration of CCSC with case-based approaches may lead to improved outcome in learning processes within non-deterministic domains.", "venue": "ICML", "year": 1991, "title": "Continous conceptual set covering: learning robot operators from examples"} +{"pid": "5ede0553e06a4c1b26a8412b", "context": "Previous works like the p-rescaled gradient flow (p-RGF) proposed by Wibisono et al. (2016) and the normalized gradient flow (NGF) by Cortes (2006) focus on the long-time behavior of continuous-time optimization algorithms.", "key_idea": "This study investigates a Lyapunov-like differential inequality, which allows for the analysis of finite-time stability of a continuous-time state-space dynamical system. Based on this, they synthesize first and second-order dynamical systems for finite-time convergence to the minima of a cost function.", "method": "Strategies are discussed for discretizing the proposed flows, and numerical experiments are conducted to test these strategies.", "outcome": "The study proves that the p-rescaled gradient flow is finite-time convergent, provided the cost function is gradient dominated of order q in (1,p). This effectively bridges a gap between the p-RGF and the NGF. The numerical experiments illustrate these results.", "future_impact": "This study's finite-time convergence analysis could provide valuable insights for the development and analysis of novel optimization algorithms in continuous-time systems.", "venue": "ICML", "year": 2020, "title": "Finite-Time Convergence in Continuous-Time Optimization"} +{"pid": "a80b732c-1ba5-41bf-944d-26d26a603c18", "context": "Learning invariant representations is a key challenge in machine learning and pattern recognition, and existing approaches don't sufficiently incorporate transformations into feature learning algorithms.", "key_idea": "The authors present a novel framework of transformation-invariant feature learning by integrating linear transformations into the feature learning algorithms. They demonstrate this via the transformation-invariant restricted Boltzmann machine, which provides invariance in its feature representation via probabilistic max pooling.", "method": "The authors evaluate their approach on several image classification benchmark datasets, including MNIST variations, CIFAR-10, and STL-10, as well as on a phone classification task using the TIMIT dataset.", "outcome": "The proposed method shows competitive or superior classification performance when compared to the state-of-the-art on benchmarked datasets. Additionally, the method achieves state-of-the-art performance on phone classification tasks with the TIMIT dataset.", "future_impact": "The framework may be extended into other unsupervised learning methods such as autoencoders or sparse encoding, suggesting wider applicability and widespread influence in domains beyond those tested.", "venue": "ICML", "year": 2012, "title": "Learning Invariant Representations with Local Transformations"} +{"pid": "319c26e1-bbab-4f8b-86a1-20442fa7302f", "context": "Outliers detection is significant in commercial data mining but there is no fixed mathematical or statistical definition for what constitutes an outlier. This causes problems as less extreme outliers or anomalies might go undetected due to the presence of the most extreme outliers, as standard deviation, a main factor in outlier detection, can be distorted by such extreme values.", "key_idea": "This study presents a novel outlier identification model that uses an improved normal distribution method which can explore and interpret various types of outliers.", "method": "The proposed model's ability to fully or partially eliminate outliers is explored using a retail point-of-sale (POS) dataset, comparing the results to those obtained by the standard normal distribution method.", "outcome": "The experiments found that the enhanced normal distribution method was more accurate than the standard form. The client subjectively examined the results and judged most of the detected outliers to be true outliers, with some representing potentially fraudulent transactions.", "future_impact": "The new model has the potential to bring forth new knowledge in outliers detection, such as the identification of fraudulent transactions, providing a more comprehensive and accurate method for detecting outliers in retail point of sale (POS) data.", "venue": "ICML", "year": 2019, "title": "Outliers Identification Model in Point-of-Sales Data Using Enhanced Normal Distribution Method"} +{"pid": "6215a5cc5aee126c0f337b1f", "context": "Current audio-visual navigation tasks need the agent to locate the sound source using audio-visual observations. However, existing works assume a clean environment that solely includes the target sound, which is not ideal in realistic conditions due to unexpected sound noise or intentional interference.", "key_idea": "The authors propose an acoustically complex environment where a sound attacker plays a zero-sum game with the agent, moving and changing the volume and category of sound to distract the agent while the agent tries to navigate to the goal under interference.", "method": "The authors design an environment where the attacker has certain constraints, then employ a centralized critic with decentralized actors to develop a joint training mechanism. The experiments are conducted on two real-world 3D scan datasets - Replica and Matterport3D.", "outcome": "The agent trained in the designed environment was found to be effective and robust when transferred to a clean environment or an environment harboring sound attackers with a random policy.", "future_impact": "This work could lead to improved robustness of agents towards unexpected sound attacks in audio-visual navigation, making them more suitable for real-world applications.", "venue": "ICLR", "year": 2022, "title": "Sound Adversarial Audio-Visual Navigation."} +{"pid": "615e657b5244ab9dcbf21ff3", "context": "Biological spiking neural networks (SNNs) can temporally encode information in their outputs in the rank order in which neurons fire, making them potentially more rapid and efficient than artificial neural networks (ANNs) when dealing with temporal input. However, ANNs are simpler to train and usually achieve superior performance.", "key_idea": "The authors propose the application of temporal coding such as rank coding (RC), inspired by SNNs, to conventional ANNs such as LSTMs. In this new approach, backpropagation applies through time using standard real-valued activations, but only from a strategically early time step of each sequential input example decided by a threshold-crossing event.", "method": "The authors demonstrate this idea in two toy problems of sequence classification, and in a temporally-encoded MNIST dataset, as well as spoken-word classification of Google Speech Commands.", "outcome": "The rank-coding method leads to computational savings and speedups, reduces time-to-insight during inference with a minimal decrease in accuracy. The rank-coded model achieves 99.19% accuracy after the first input time-step in the temporally-encoded MNIST dataset and outperforms non-RC-trained early inference with LSTM models in spoken-word classification.", "future_impact": "The speed-accuracy trade-off can be adjusted by varying the threshold or a regularization parameter that rewards output entropy, implying its potential use in a wider range of applications.", "venue": "ICLR", "year": 2022, "title": "Spike-inspired rank coding for fast and accurate recurrent neural networks"} +{"pid": "95f42870-7984-4fb5-a944-1c55d0a72aa4", "context": "Large transformer-based language models trained on vast text corpora have impressive generation capabilities, but controlling attributes of the generated language, such as topic or sentiment, is challenging without modifying the model architecture or fine-tuning on attribute-specific data.", "key_idea": "The authors propose the Plug and Play Language Model (PPLM) for controllable language generation, which combines a pretrained language model with one or more simple attribute classifiers that guide text generation without any further training of the language model.", "method": "The attribute models are simple classifiers that can be a user-specified bag of words or a learned single layer. The model samples demonstrate control over a range of topics and sentiment styles using both automated and human-annotated evaluations.", "outcome": "The Plug and Play Language Model showed control over a range of topics and sentiment styles, and extensive automated and human annotated evaluations showed attribute alignment and fluency.", "future_impact": "PPLMs are flexible in that any combination of differentiable attribute models may be used to steer text generation, suggesting diverse and creative applications beyond the examples given in this paper.", "venue": "ICLR", "year": 2019, "title": "Plug and Play Language Models: A Simple Approach to Controlled Text Generation"} +{"pid": "6168f1a35244ab9dcbe2ff24", "context": "Most previous studies on linguistic information encoding in hidden word representations have not focused on understanding how and in which individual neurons this information is encoded. The common approach to investigate this is to use an external probe to rank neurons based on their relevance to a linguistic attribute, and the obtained ranking is evaluated using the same probe that produced it.", "key_idea": "The authors identified two pitfalls in the common method for analyzing individual neurons in language models: the methodology confounds probe quality and ranking quality, and it focuses on encoded information without considering the information used by the model.", "method": "The authors separated probe quality and ranking quality to draw conclusions on each. They also compared two existing neuron ranking methods with a new simple method they introduced, evaluating all methods with regard to both identified pitfalls.", "outcome": "The authors demonstrated that probe quality and ranking quality are distinct factors in analyzing individual neurons in language models. They also showed that the encoded information and the information used by the model are not the same.", "future_impact": "This study, by highlighting potential pitfalls in current methods, could steer the future research towards developing more accurate and comprehensive methodologies for analyzing individual neurons in language models.", "venue": "ICLR", "year": 2022, "title": "On the Pitfalls of Analyzing Individual Neurons in Language Models"} +{"pid": "81a0866d-a999-4b03-ba25-29d8bee1cee5", "context": "Prior research lacked efficient algorithms that could learn domain-invariant image representations and address domain mismatch issues in machine learning context.", "key_idea": "The paper presents a new algorithm that learns representations which compensate for domain mismatch and can be efficiently implemented as linear classifiers. The authors introduce a linear transformation which maps the test domain features to the training domain as part of training the classifier.", "method": "The transformation and classifier parameters are optimized collectively and the process includes an efficient cost function based on misclassification loss.", "outcome": "The algorithm was tested on several image datasets, demonstrating improved accuracy and computational benefits compared to previous solutions.", "future_impact": "The paper introduces an algorithm that has several features which haven't been available in a single algorithm before: multi-class adaptation through representation learning, ability to map between different feature spaces, and scalability to larger datasets.", "venue": "ICLR", "year": 2013, "title": "Efficient Learning of Domain-invariant Image Representations"} +{"pid": "b857c700-6c2e-49a6-809b-ed5e95f97eab", "context": "Existing environments for training and evaluating the generalization ability of intelligent systems fail to offer visual realism or rich content.", "key_idea": "The authors propose House3D, an efficient and extensible environment featuring 45,622 human-designed, fully labeled 3D house scenes equipped with various objects, textures, and layouts.", "method": "The authors investigate the task of concept-driven navigation (RoomNav), in which an agent learns to navigate towards a semantic concept target while obeying physical laws. Reinforcement learning (RL) agents with both continuous and discrete action spaces are trained to test their generalization in unseen environments.", "outcome": "Through this method, it was observed that training on larger house sets improved generalization, using semantic signals boosted generalization performance, and gated networks on semantic input signals further improved both training and generalization.", "future_impact": "The authors anticipate House3D to serve as a crucial building block in designing practical intelligent systems and hope it will be widely adopted by the research community.", "venue": "ICLR", "year": 2018, "title": "Building Generalizable Agents with a Realistic and Rich 3D Environment"} +{"pid": "83dda395-6338-4807-8cc6-d00a7c97b7e4", "context": "Existing probabilistic programming languages do not treat inference as a first class citizen and are not as flexible and computationally efficient as traditional deep learning methodologies.", "key_idea": "The authors propose Edward, a Turing-complete probabilistic programming language, defining two compositional representations - random variables and inference - to provide flexibility similar to traditional deep learning and ensuring inference is treated as a first class citizen.", "method": "The authors benchmarked Edward by fitting a model using a variety of composable inference methods in a logistic regression task. They compared its efficiency with Stan and PyMC3.", "outcome": "The proposed language, Edward, demonstrated significant speedups over existing probabilistic systems. It proved to be at least 35 times faster than Stan and 6 times faster than PyMC3, with no runtime overhead compared to handwritten TensorFlow.", "future_impact": "Edward can facilitate design of rich variational models and generative adversarial networks, promoting increased flexibility and computational efficiency in probabilistic programming.", "venue": "ICLR", "year": 2017, "title": "Deep Probabilistic Programming"} +{"pid": "798d36d5-e6aa-4051-b97c-73c145c5101c", "context": "The detection of misclassified or out-of-distribution examples in neural networks is a crucial task in machine learning, but existing methods often struggle to consistently distinguish between correct, incorrect, and out-of-scope examples.", "key_idea": "The authors propose a simple baseline for detection that leverages probabilities from softmax distributions, hypothesizing that correctly classified examples will have greater maximum softmax probabilities than misclassified and out-of-distribution examples.", "method": "The proposed baseline is evaluated across a variety of tasks in computer vision, natural language processing, and automatic speech recognition to verify its robustness and versatility.", "outcome": "The assessment illustrates that the proposed baseline effectively detects misclassified and out-of-distribution examples across various tasks and domains.", "future_impact": "The study demonstrates some cases where the proposed baseline can be outperformed, indicating opportunities for future investigations and improvements in these underexplored detection tasks.", "venue": "ICLR", "year": 2017, "title": "A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks"} +{"pid": "6232a74d5aee126c0fe13e61", "context": "As reinforcement learning (RL) has achieved near human-level performance in various tasks, its robustness against training-time (poisoning) attacks, where a subset of training trajectories could be manipulated arbitrarily, remains largely unanswered.", "key_idea": "This work proposes the first certification framework, COPA, to certify the number of poisoning trajectories that an offline RL system can tolerate based on different certification criteria.", "method": "The researchers propose two certification criteria: per-state action stability and cumulative reward bound. This is coupled with partition and aggregation protocols to train robust policies. The COPA framework is then applied to three RL environments trained with different algorithms to evaluate the robustness against poisoning attacks.", "outcome": "The experiments conclude that robust aggregation protocols, such as temporal aggregation, can significantly improve certifications. The efficiency and tightness of the proposed certification for both per-state action stability and cumulative reward bound were confirmed. It was also found that the certification for different training algorithms and environments varies, indicating their intrinsic robustness properties.", "future_impact": "Having a certification method in place for the robustness of offline RL systems against poisoning attacks could drive an increased adoption of more secure and robust RL systems in practice.", "venue": "ICLR", "year": 2022, "title": "COPA: Certifying Robust Policies for Offline Reinforcement Learning against Poisoning Attacks"} +{"pid": "3fc1a0ab-b116-4ba0-b94c-15e6945ce097", "context": "Recurrent Neural Networks (RNNs) are the prevalent model for processing sequential data. Advances have mostly been focussed on optimization or modelling of RNNs, motivated by addressing issues of vanishing and exploding gradients. However, the problem of overfitting has received considerably less attention.", "key_idea": "The authors analyze fast dropout, a recent regularization method for generalized linear models and neural networks, from a back-propagation inspired perspective. The fast dropout implements a quadratic form of an adaptive, per-parameter regularizer, which rewards large weights when underfitting, penalizes them for overconfident predictions, and vanishes at minima of an unregularized training loss.", "method": "The authors test the hypothesis that using a fast dropout improves the performance of RNNs on four musical data sets.", "outcome": "The application of fast dropout demonstrates improvement in the performance of Recurrent Neural Networks on four musical datasets.", "future_impact": "The elimination of a global weight attractor, particularly beneficial for RNNs as the dynamics are not biased towards a certain regime, may inspire the development of more efficient RNN models in future.", "venue": "ICLR", "year": 2014, "title": "On Fast Dropout and its Applicability to Recurrent Networks"} +{"pid": "5ee7495191e01198a507f7e2", "context": "Designing task-oriented dialogue systems is challenging as it needs to generate utterances fulfilling user requests while also ensuring comprehensibility. Previous approaches faced issues such as the bias in annotated system utterances and the reduction in the comprehensibility of generated system utterances when improving the performance on fulfilling user requests.", "key_idea": "The authors propose a model called HDNO (Hierarchical Dialogue system with Natural language Options), which models the hierarchical structure between dialogue policy and natural language generator (NLG). The model utilizes the option framework and applies a latent dialogue act to avoid designing specific dialogue act representations.", "method": "HDNO is trained via hierarchical reinforcement learning (HRL) with asynchronous updates. A discriminator modelled with language models is used as an additional reward to improve comprehensibility. The model is tested on MultiWoz 2.0 and MultiWoz 2.1 datasets for multi-domain dialogues.", "outcome": "The proposed HDNO model showed improvements on the performance evaluated by automatic evaluation metrics and human evaluation when compared with a word-level E2E model, LaRL, and HDSA.", "future_impact": "This work has not explicitly stated its future impact. However, the proposed approach offers a possible direction for designing highly communicative and comprehensible dialogue systems.", "venue": "ICLR", "year": 2021, "title": "Modelling Hierarchical Structure between Dialogue Policy and Natural Language Generator with Option Framework for Task-oriented Dialogue System"} +{"pid": "5361cb41-95ef-4b1a-9d1f-175e2e329e49", "context": "Large-scale distributed training requires significant communication bandwidth for gradient exchange, which limits the scalability of multi-node training and requires expensive high-bandwidth network infrastructure. The situation gets worse with distributed training on mobile devices or federated learning.", "key_idea": "The paper introduces Deep Gradient Compression (DGC), a method that identifies redundant gradient exchanges in distributed stochastic gradient descent (SGD) and aims to reduce communication bandwidth while preserving accuracy.", "method": "The DGC method employs techniques like momentum correction, local gradient clipping, momentum factor masking, and warm-up training to preserve accuracy during compression. The method is applied to tasks like image classification, speech recognition, and language modeling using multiple datasets including Cifar10, ImageNet, Penn Treebank, and Librispeech Corpus.", "outcome": "Deep Gradient Compression was able to achieve a gradient compression ratio of 270x to 600x without losing accuracy, reducing the gradient size of models like ResNet-50 and DeepSpeech significantly. This facilitates large-scale distributed training on less expensive hardware and on mobile devices.", "future_impact": "The reduction in communication bandwidth for large-scale distributed training can enable usage of more inexpensive infrastructure options in the future and might facilitate improved distributed training, especially in mobile environments.", "venue": "ICLR", "year": 2018, "title": "Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training"} +{"pid": "5f3665c791e011372ac2d992", "context": "Bayesian neural networks often require the use of a 'tempered' or 'cold' posterior to achieve performance similar to standard neural networks. This practice goes against Bayesian principles that, given an accurate prior, suggest any artificial changes to the posterior should harm performance.", "key_idea": "Rather than inaccuracies in the prior, the authors suggest that the root of the problem lies in the use of the wrong likelihood for Bayesian Neural Networks in image classification. The authors argue that the standard image benchmark datasets, like CIFAR-10, are carefully curated which affects their assumption of randomness.", "method": "A generative model describing the curation process is developed, which gives a Bayesian explanation for cold posteriors.", "outcome": "The likelihood produced under the new generative model closely matches the tempered likelihoods previously used, hence supporting the authors' argument.", "future_impact": "This study's findings could provoke a shift in how Bayesian neural networks handle likelihoods and break from the dependency on tempered or cold posteriors.", "venue": "ICLR", "year": 2021, "title": "A statistical theory of cold posteriors in deep neural networks"} +{"pid": "b7ba202c-be35-4aa3-92ef-32cbe5cd7669", "context": "Learning tasks on source code have been considered recently, but most work has tried to transfer natural language methods and does not capitalize on the unique opportunities offered by code's known syntax. For example, long-range dependencies induced by using the same variable or function in distant locations are often not considered.", "key_idea": "This paper proposes to use graphs to represent the syntactic and semantic structure of the code and further apply graph-based deep learning methods to reason over program structures.", "method": "The researchers constructed graphs from source code and scale Gated Graph Neural Networks training to large graphs. The method is evaluated on two tasks: VarNaming, where a network attempts to predict the name of a variable given its usage, and VarMisuse, where the network learns to reason about selecting the correct variable that should be used at a given program location.", "outcome": "Their comparison to methods that use less structured program representations shows advantages of modeling known structure. Tests showed that VarMisuse identifies a number of bugs in mature open-source projects.", "future_impact": "The method is implied to be beneficial for more complex tasks on code comprehension as it helps infer meaningful names and solve the VarMisuse task in many cases.", "venue": "ICLR", "year": 2018, "title": "Learning to Represent Programs with Graphs"} +{"pid": "5fd8847b91e0119b22c1f0be", "context": "The human brain processes language input within a distributed and hierarchical architecture, encoding contextual information over varying timescales. In contrast, the functional organisation and processing timescales in recurrent neural networks oerforming natural language processing tasks are not fully understood.", "key_idea": "The authors tackle this knowledge gap by applying neuroscience techniques to map the 'processing timescales' of individual units within a word-level LSTM language model. They identified two classes of long-timescale units: 'Controller' units which composed a densely interconnected subnetwork, and 'Integrator' units which showed the longest timescales in the network.", "method": "Long timescales were assigned to units found to track long-range syntactic dependencies, revealing a new cluster of previously unreported long-timescale units. The functional role of units was further examined by studying the relationship between their processing timescales and network connectivity. Tests were also extended to a character-level LSTM model.", "outcome": "The application of the timescale-mapping method led to the identification of new clusters of long-timescale units. Ablating integrator and controller units affected model performance at different position of a sentence, implying distinctive functions for these sets of units.", "future_impact": "This model-free technique for mapping the timescale organization in neural networks provides potential for further understanding and improving the functional organization of LSTM language models and other neural network models.", "venue": "ICLR", "year": 2021, "title": "Mapping the Timescale Organization of Neural Language Models"} +{"pid": "600831a49e795ed227f530c2", "context": "Combinatorial generalisation, an essential component of human intelligence, is a concept that current AI systems struggle with. Recently, it has been suggested that learning disentangled representations might help solve this issue.", "key_idea": "The authors begin with the idea that disentangled representations could capture the compositional structure of the world and hence enable combinatorial generalisation, and then set out to systematically test how the degree of disentanglement affects various forms of generalisation.", "method": "Three classes of variational autoencoders (VAEs) are trained on two datasets on an unsupervised task, and combinations of generative factors are excluded during training. At test time, the models are asked to reconstruct the missing combinations to measure generalisation performance.", "outcome": "Irrespective of the degree of disentanglement, the models only support weak combinatorial generalisation. Even when tested with perfectly disentangled representations as the latents or with a model on a more complex task requiring independent generative factors to be controlled, the results remained the same.", "future_impact": "The study suggests that although learning disentangled representations can improve interpretability and sample efficiency in some downstream tasks, they are not sufficient for supporting more difficult forms of generalisation, indicating scope for further research in this area.", "venue": "ICLR", "year": 2021, "title": "The role of Disentanglement in Generalisation"} +{"pid": "009ddfcd-eb72-4989-ace5-ddd83a91eb58", "context": "Methods that learn representations of graph nodes play a critical role in network analysis enabling many downstream learning tasks. However, most existing approaches represent nodes as point vectors in a low-dimensional continuous space, without considering the uncertainty about the representation.", "key_idea": "The authors propose Graph2Gauss, a novel approach that represents each node as a Gaussian distribution to capture uncertainty in embeddings, and adopts a personalized ranking formulation regarding the node distances leveraging both network structure and node attributes.", "method": "The proposed method, which is capable of handling inductive learning scenarios and different types of graphs (plain/attributed, directed/undirected), is evaluated through experiments on real world networks, focusing on tasks such as link prediction, node classification and also for estimating neighborhood diversity and detecting the intrinsic latent dimensionality of a graph.", "outcome": "Experiment results indicate that Graph2Gauss outperforms state-of-the-art network embedding methods in several different tasks demonstrating the high performance of the approach.", "future_impact": "The proposed method can give rise to improved graph embedding techniques, especially where capturing uncertainty and exploiting node attributes are key considerations. The capability of the method for estimating neighborhood diversity and detecting the intrinsic latent dimensionality of a graph can also be beneficial to future graph analysis tasks.", "venue": "ICLR", "year": 2018, "title": "Deep Gaussian Embedding of Graphs: Unsupervised Inductive Learning via Ranking"} +{"pid": "5f7c543a91e0117ac2a78be2", "context": "Graph-structured data provide diverse types of information which play different roles in data representation, but many popular Graph Neural Network (GNN) models are incapable of effectively utilizing all the relational information stored in the graphs.", "key_idea": "The authors propose the Copula Graph Neural Network (CopulaGNN) which effectively utilises the representational and correlational roles of graphs. Providing guidance for the model to construct better node features and indicating the correlation between node outcomes conditional on node features.", "method": "The methodology of the study involves the use of a simulation. The authors also performed experiments on two types of regression tasks to verify the effectiveness of the proposed method.", "outcome": "The experimental results validate the effectiveness of the proposed CopulaGNN method in utilizing both representational and correlational information stored in the graphs.", "future_impact": "The proposed CopulaGNN model has potential to broadly impact the future use of Graph Neural Networks by offering an efficient way to leverage all the representational and correlational information stored in graphs for node-level prediction tasks.", "venue": "ICLR", "year": 2021, "title": "CopulaGNN: Towards Integrating Representational and Correlational Roles of Graphs in Graph Neural Networks"} +{"pid": "5e5e189793d709897ce1d3cc", "context": "Nesterov SGD is commonly used for training modern neural networks and other machine learning models, yet its advantages over standard SGD are not theoretically clarified and it might even diverge for certain step sizes that ensures convergence of SGD.", "key_idea": "The paper introduces a compensation term to Nesterov SGD to create a new algorithm called MaSS, which converges for the same step sizes as SGD and is theoretically identified to obtain accelerated convergence rates over SGD.", "method": "Authors provide theoretical proofs for the performance of MaSS algorithm and its convergence properties. The paper also conducts experimental evaluation of MaSS for several standard architectures of deep networks, including ResNet and convolutional networks, comparing it with SGD, Nesterov SGD and Adam.", "outcome": "Theoretical proof shows accelerated convergence rates of MaSS over SGD for any mini-batch size in the linear setting and its convergence rate matches the well-known accelerated rate of the Nesterov's method for full batch. Experimental results show improved performance of MaSS over SGD, Nesterov SGD and Adam.", "future_impact": "Analysis on the relation of the convergence rate and optimal hyperparameters on mini-batch size, demonstrating three distinct regimes - linear scaling, diminishing returns and saturation, may guide further improvements and optimization in training deep learning models.", "venue": "ICLR", "year": 2020, "title": "Accelerating SGD with momentum for over-parameterized learning"} +{"pid": "600832ef9e795ed227f5314b", "context": "There has been a rise in model stealing attacks due to increasing value of machine learning models as intellectual property. Current defenses involve inserting unique watermarks in the model's decision surface but it proves inadequate because they may not be preserved during model stealing.", "key_idea": "The paper proposes the concept of dataset inference as a defense against model stealing. The idea is based on the observation that the knowledge contained in the stolen model's training set is common to all stolen copies.", "method": "The authors develop an approach for dataset inference that combines statistical testing with the ability to estimate the distance of multiple data points to the decision boundary. Their experiments were performed on CIFAR10 and CIFAR100.", "outcome": "These experiments showed that model owners can claim with confidence greater than 99% that their model or dataset was stolen, despite only exposing 50 of the stolen model's training points. Dataset inference defends against state-of-the-art attacks, even when the adversary is adaptive.", "future_impact": "Unlike prior work, dataset inference does not require retraining or overfitting the defended model, making it a potential robust defense mechanism against model stealing attacks in the future.", "venue": "ICLR", "year": 2021, "title": "Dataset Inference: Ownership Resolution in Machine Learning"} +{"pid": "5eede0b791e0116a23aafd5b", "context": "Architecture search methods suffer from challenges like inefficient generalization ability, lack of stochasticity for exploration in the search space, and large memory consumption which inhibits their application on large-scale tasks.", "key_idea": "The paper presents DrNAS, a novel differentiable architecture search method that treats architecture mixing weight as random variables modeled by Dirichlet distribution and optimizes Dirichlet parameters in an end-to-end manner. It also utilizes a progressive learning scheme to tackle memory consumption issues.", "method": "The authors evaluate their proposed method through experiments on the CIFAR-10 and ImageNet datasets under the mobile setting, as well as NAS-Bench-201.", "outcome": "The proposed method demonstrates promising results, with a test error of 2.46% for CIFAR-10, 23.7% for ImageNet under the mobile setting, and state-of-the-art results on NAS-Bench-201 on all three datasets.", "future_impact": "The paper's insights could potentially aid in the effective design of future neural architecture search algorithms.", "venue": "ICLR", "year": 2021, "title": "DrNAS: Dirichlet Neural Architecture Search"} +{"pid": "600833a39e795ed227f531a6", "context": "State-of-the-art implementations of reinforcement learning (RL) algorithms make many design decisions that affect performance, and there is a discrepancy between published descriptions of algorithms and their implementations. This makes it difficult to attribute progress in RL.", "key_idea": "The authors implement over 50 design choices within a unified on-policy deep actor-critic framework, which is used to investigate the impact of these choices on algorithm performance.", "method": "Through large-scale empirical study, over 250,000 agents are trained in five different continuous control environments of varying complexity.", "outcome": "The authors provide insights and practical recommendations for the training of on-policy deep actor-critic RL agents, based on their large-scale study.", "future_impact": "These findings can contribute to better understanding and implementation of reinforcement learning algorithms by providing practical guidelines based on empirical observations.", "venue": "ICLR", "year": 2021, "title": "What Matters for On-Policy Deep Actor-Critic Methods? A Large-Scale Study"} +{"pid": "61664e615244ab9dcb455761", "context": "CNNs are powerful generalization systems but understanding how their inductive bias influences model decisions and what causes one model to decide differently from another demands further investigation.", "key_idea": "The authors aim to understand the factors that influence a model's decisions by conducting a detailed analysis on ImageNet validation set. They identify a condition of 'dichotomous data difficulty' (DDD) in which a substantial proportion of images are either 'trivial' or 'impossible' for model's discrimination.", "method": "The authors perform a meticulous, epoch-to-epoch, image-by-image study of model decisions on the ImageNet validation set, across different architectures and objectives, thereby investigating the role of images and their difficulties.", "outcome": "The study found that all models have a similar decision boundary and that 46.0% of images in the validation sets are 'trivial' and 11.5% are 'impossible' for models applied. The differences in decision boundaries between two models can only be attributable to 42.5% of the images. Removing 'trivial' and 'impossible' images revealed more pronounced differences between models.", "future_impact": "The findings imply the importance of investigating the decisive role of images and their difficulty distribution in future comparisons of brains, machines and behaviour. Humans are demonstrated to be highly accurate at predicting the perceived difficulty level for CNNs of images.", "venue": "ICLR", "year": 2022, "title": "Trivial or Impossible --- dichotomous data difficulty masks model differences (on ImageNet and beyond)"} +{"pid": "600832769e795ed227f5310b", "context": "Hopfield networks and Restricted Boltzmann Machines are two important models at the interface of statistical physics, machine learning, and neuroscience. Though a relationship between HNs and RBMs has been previously noted in the case of orthogonal encoded patterns, the investigation hasn't been extended to correlated pattern HNs.", "key_idea": "The authors establish an exact mapping between correlated pattern Hopfield networks and Restricted Boltzmann Machines, demonstrating that a HN with N binary variables and potentially correlated binary patterns can transform into an RBM with N binary visible variables and Gaussian-hidden variables.", "method": "The authors outline the conditions for reverse mapping and conduct experiments on the MNIST dataset to test the utility of this mapping for initializing RBM weights.", "outcome": "The experiments conducted using the MNIST dataset suggest that the mapping provides a beneficial initialization to the Restricted Boltzmann Machine weights.", "future_impact": "The understanding of this correspondence could have significance for the training of RBMs, and for enhancing our understanding of feature extraction methods which utilize RBMs.", "venue": "ICLR", "year": 2021, "title": "On the mapping between Hopfield networks and Restricted Boltzmann Machines"} +{"pid": "63180bf390e50fcafded7135", "context": "Many recent approaches to natural language tasks are built on the abilities of large language models, particularly their ability to perform in-context learning, where they learn a new task from a few task demonstrations, without any parameter updates.", "key_idea": "This work proposes an annotation-efficient, two-step framework involving selective annotation that chooses a pool of examples to annotate from unlabeled data in advance, and prompt retrieval that retrieves task examples from the annotated pool at test time. A graph-based selective annotation method, voke-k, is introduced to select diverse, representative examples.", "method": "The proposed framework and methods are tested in extensive experiments on 10 different datasets, covering classification, commonsense reasoning, dialogue, and text/code generation, and under various scenarios including different language model sizes, alternate selective annotation methods, and cases involving test data domain shift.", "outcome": "The selective annotation method improves task performance significantly, achieving average relative gains of 12.9% and 11.4% under annotation budgets of 18 and 100 respectively, compared to random selection. Additionally, it yields similar performance to state-of-the-art supervised finetuning approaches, with 10-100 times less annotation cost across 10 tasks.", "future_impact": "The authors anticipate that their research studies will serve as a basis for future data annotations as large language models are increasingly applied to new tasks.", "venue": "ICLR", "year": 2022, "title": "Selective Annotation Makes Language Models Better Few-Shot Learners"} +{"pid": "5f117bc991e011264d447776", "context": "Despite the progress in robust deep learning, there is still a significant gap in robustness between real-world perturbations and those defined in adversarial defenses.", "key_idea": "The authors propose to bridge this gap by learning perturbation sets from data to characterize real-world effects for robust training and evaluation. A conditional generator is used to define the perturbation set over a constrained region of the latent space.", "method": "The authors use a conditional variational autoencoder to learn perturbation sets, with desirable properties and a theoretical proof. They measure the quality of these sets both quantitatively and qualitatively, and use them to train models to become robust to adversarial image corruptions and lighting variations.", "outcome": "The method successfully produces a variety of meaningful perturbations of different complexities and scales. Models trained using these learned perturbation sets prove to be robust to adversarial image corruptions and lighting variations, and improved generalization on non-adversarial data is observed.", "future_impact": "The disclosed method of learning perturbation sets from data to improve model robustness could inspire new approaches for cutting-edge adversarial defenses in machine learning systems.", "venue": "ICLR", "year": 2021, "title": "Learning perturbation sets for robust machine learning"} +{"pid": "61fb47e15aee126c0f873a4a", "context": "Previous work has reported that adversarially robust networks seem more human perceptually-aligned than non-robust networks, but it remains unclear how closely the features and constraints in these representations mimic those found in human vision.", "key_idea": "The paper investigates how human perceptual behavior compares to the behavior of images synthesized to match adversarially robust representations, non-robust representations, and a model of peripheral vision (Texforms).", "method": "The authors conducted a psychophysics experiment with a set of metameric discrimination tasks, measuring how well human observers could distinguish between different types of synthesized images when presented at various distances from their visual center.", "outcome": "The authors found that the discriminability of images matching robust representation and the texture model decreased to near chance levels when stimuli were presented farther in the periphery, and performance on robust and texture-model images followed similar trends within participants, suggesting adversaries robust representations capture peripheral computation better than non-robust representations.", "future_impact": "The findings may encourage further research into incorporating localized texture summary statistic representations into Deep Neural Networks (DNNs) in order to bolster properties like adversarial robustness.", "venue": "ICLR", "year": 2022, "title": "Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks"} +{"pid": "5d9b0cbe3a55acb039198fb5", "context": "Current generative models for text-to-speech translation are primarily unsupervised models and hence struggle to have consistent and interpretable variables. Moreover, control over aspects such as affect and speaking rate is limited in existing models.", "key_idea": "The authors propose a novel generative model that leverages semi-supervised learning with probabilistic latent variables. This approach enables partial supervision on some latent variables leading to forced consistent and interpretable purposes.", "method": "The model's ability to discover and control important but rarely labelled attributes of speech like affect and speaking rate, was tested with minimal supervision levels (1% or 30 minutes).", "outcome": "The proposed semi-supervised generative model was able to reliably discover and control speech attributes without degrading synthesis quality compared to a state-of-the-art baseline, even with minimal supervision.", "future_impact": "The authors plan to release audio samples demonstrating the effectiveness of their new model. This example of using semi-supervised learning models for controllable speech synthesis could inspire and facilitate further research in this field.", "venue": "ICLR", "year": 2020, "title": "Semi-Supervised Generative Modeling for Controllable Speech Synthesis"} +{"pid": "98ee7f4c-4cfe-4ced-8fa4-71ef942385fd", "context": "While distributed vector representations of words have driven much success in NLP when trained on large amounts of unsupervised text, extending this success to learning representations of sequences of words or sentences is an open problem.", "key_idea": "The authors propose a multi-task learning framework for sentence representations amalgamating diverse training objectives into a single model. This model is trained on multiple data sources across different training objectives.", "method": "The researchers train their multi-task learning model on several data sources with multiple training objectives and over 100 million sentences and test it across various tasks, including transfer learning and low-resource settings.", "outcome": "The experiments show that sharing a single recurrent sentence encoder across weakly related tasks significantly improves performance over previous methods.", "future_impact": "The significant improvement in transfer learning and low-resource settings using the general-purpose representations learned in this study could potentially make these representations highly beneficial in various NLP applications.", "venue": "ICLR", "year": 2018, "title": "Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning"} +{"pid": "617771c35244ab9dcbe79d5d", "context": "High-performing models trained with supervision generally make similar predictions, implying shared biases irrespective of the training methodology. This could limit the ensembling benefits and render low-accuracy models as of little practical use.", "key_idea": "The authors challenge the assumption that models necessarily learn similar functions. They instead propose that models trained with diverging methodologies can exhibit different generalization behaviors and specialize in different subdomains of the data.", "method": "A large-scale empirical study is conducted across various hyper-parameters, architectures, frameworks, and datasets. The authors compare performance of model pairs trained with diverging methodologies and examine their errors and generalization behaviors.", "outcome": "The study finds that model pairs that diverge more in training methodology produce increasingly uncorrelated errors and specialize in different subdomains of the data, leading to higher ensemble performance. Low-accuracy models surprisingly can be used to improve high-accuracy models. Diverging training methodology yields representations that capture overlapping feature sets which boost downstream performance when combined.", "future_impact": "This study could impact how we view the role of training methodologies and model biases in machine learning, possibly leading to a reevaluation of low-accuracy models' utility and altering how ensembles are created for improved performance.", "venue": "ICLR", "year": 2022, "title": "No One Representation to Rule Them All: Overlapping Features of Training Methods"} +{"pid": "5e5e18ed93d709897ce3dc5c", "context": "Transformer-based models such as Transformer-XL and BERT have achieved success on various natural language processing tasks. However, they suffer from a degeneration problem where contextualized embeddings at the output layer tend to occupy an anisotropic cone in the vector space.", "key_idea": "The authors propose a novel spectrum control approach to address the representation degeneration problem. The method guides the spectra training of the output embedding matrix with a slow-decaying singular value prior distribution through a reparameterization framework.", "method": "The authors use their proposed spectrum control method on Transformer-XL for language modeling and various Transformer-based models for machine translation, using common benchmark datasets for these tasks.", "outcome": "The experiments show that the proposed method encourages isotropy of the learned word representations while maintaining the modelling power of these contextual neural models. It outperforms the state-of-the-art Transformer-XL modeling for language model, and various Transformer-based models for machine translation.", "future_impact": "The core idea of directly guiding the spectra training of output embeddings could potentially improve other types of models in the future, however this is not explicitly mentioned in the abstract.", "venue": "ICLR", "year": 2020, "title": "Improving Neural Language Generation with Spectrum Control"} +{"pid": "61baae695244ab9dcb643f1b", "context": "Deep generative models often struggle to simultaneously deliver high sample quality, mode coverage, and fast sampling, a challenge known as the generative learning trilemma. In particular, denoising diffusion models have shown impressive sample quality and diversity, but their slow sampling restricts their application in many real-world settings.", "key_idea": "The authors suggest that the slow sampling inherent to denoising diffusion models is due to the Gaussian assumption in the denoising step. To mitigate this, they propose to model the denoising distribution using a complex multimodal distribution.", "method": "The authors introduce denoising diffusion generative adversarial networks (denoising diffusion GANs) that model each denoising step using a multimodal conditional GAN. Evaluations of this method involved the CIFAR-10 dataset.", "outcome": "Denoising diffusion GANs are shown to deliver sample quality and diversity comparable to original diffusion models but perform 2000x faster on the CIFAR-10 dataset. Compared to traditional GANs, the new model has better mode coverage and sample diversity.", "future_impact": "The introduction of denoising diffusion GANs represents the first model to substantially reduce the sampling cost in diffusion models, thus providing the potential for application in real-world settings due to lower costs.", "venue": "ICLR", "year": 2022, "title": "Tackling the Generative Learning Trilemma with Denoising Diffusion GANs"} +{"pid": "60055a9b91e0118e0cce2262", "context": "Prior works reveal that neural networks for image classification tend to learn associations with low-level texture or background instead of recognizing object shape, relying on simple correlations and ignoring complex ones that could potentially offer better generalization.", "key_idea": "The authors develop a more robust and interpretable classifier by decomposing the image generation process into independent causal mechanisms that can separately disentangle object shape, object texture, and background, allowing for counterfactual image generation.", "method": "The authors validate their approach by applying it to the MNIST and ImageNet datasets, demonstrating the model's ability to generate counterfactual images and measuring its effects on out-of-distribution robustness and performance on the original classification task.", "outcome": "The proposed model is shown to efficiently generate counterfactual images and improve out-of-distribution robustness with a nominal drop in performance on the original classification task. The model is also trainable on a single GPU, exploiting common pre-trained models as inductive biases.", "future_impact": "The proposed generative model points towards the development of more robust and interpretable classifiers that explicitly expose a task's causal structure, potentially improving the robustness and interpretability of AI models in image classification tasks.", "venue": "ICLR", "year": 2021, "title": "Counterfactual Generative Networks"} +{"pid": "ce03d8a6-455d-473e-aa92-006a41602d1e", "context": "Many generative learning algorithms involve complex probability distribution over random variables which are often intractable due to normalization constants or latent variable normalization. Measuring the quality of these models once they have been trained is difficult due to the lack of analytic expression for the unnormalized probability function and non-tractable approximations.", "key_idea": "The authors propose a more efficient estimator that constructs a non-parametric density estimate of the model's probability function from samples generated by the model.", "method": "An efficient estimator is created by revisiting the idea of constructing a non-parametric density estimator from samples generated by the model. A biased variant of the estimator is also proposed for use with a finite number of samples for the purpose of model comparison.", "outcome": "The authors have successfully developed an estimator that provides a lower bound on the true test log-likelihood and an unbiased estimator as the number of generated samples goes to infinity, incorporating the effect of poor mixing.", "future_impact": "The enhanced estimator could be used for estimating the quality of generative models, possibly aiding in tasks such as early stopping during training and model comparison using a finite number of samples.", "venue": "ICLR", "year": 2014, "title": "Bounding the Test Log-Likelihood of Generative Models"} +{"pid": "6008330b9e795ed227f5315e", "context": "Multi-task learning (MTL) is a widely used technique in representation learning. However, training all tasks simultaneously often leads to a particular issue where certain tasks are trained more adequately than others\u2014known as the 'partial training' issue.", "key_idea": "A method is proposed to learn multiple tasks impartially. The method optimizes scaling factors for task-shared parameters and weighs task-specific losses to ensure they are kept at a comparable scale. A hybrid balance method is also proposed which combines gradient balance and loss balance to improve performance. The approach, called Impartial Multi-task Learning (IMTL), can be trained end-to-end, applies to all kinds of losses, and is scale-invariant.", "method": "IMTL is evaluated on standard MTL benchmarks, including Cityscapes, NYUv2, and CelebA.", "outcome": "IMTL achieves new state-of-the-art results among loss weighting methods under the same experimental settings on the standard MTL benchmarks.", "future_impact": "The end-to-end training capability, general applicability to different types of losses, and scale-invariance of the IMTL approach potentially provide a more versatile and robust approach in multi-task learning, influencing future model designs in this area.", "venue": "ICLR", "year": 2021, "title": "Towards Impartial Multi-task Learning"} +{"pid": "617771bd5244ab9dcbe795ce", "context": "Humans are known for their efficient exploration of new environments. However, the computational cognitive mechanisms supporting this efficiency have not been fully understood, posing a challenge to advancing the study of the human mind and developing more efficient exploration algorithms.", "key_idea": "The authors hypothesize that humans explore new environments efficiently by inferring the structure of unobserved spaces using spatial information collected from previously explored spaces. They propose a model based on program induction in a Hierarchical Bayesian framework, which explicitly reasons about uncertainty with strong spatial priors.", "method": "The authors use a new behavioral Map Induction Task to test their computational framework and compare it with non-inductive models and state-of-the-art planning algorithms applied to a realistic spatial navigation domain.", "outcome": "The results show that their computational framework explains human exploration behavior better than non-inductive models and outperforms state-of-the-art planning algorithms in a realistic spatial navigation domain.", "future_impact": "This study paves the way for advancing the understanding of human cognitive mechanism and enhancing algorithmic exploration efficiency, which could benefit various fields such as robotics and AI navigation.", "venue": "ICLR", "year": 2022, "title": "Map Induction: Compositional spatial submap learning for efficient exploration in novel environments"} +{"pid": "5fa14ed291e011f3c66576e2", "context": "Large-scale language models deliver impressive performance but are hindered by their large size, power consumption, and slower inference tally, which are impractical for low-resource platforms. Knowledge distillation (KD) can be used for model compression, but there are issues with large-scale neural networks memorizing training instances and making inconsistent predictions with slight data distribution changes. Also, when there is limited task-specific data available, the student model cannot request much information from the teacher model.", "key_idea": "The authors propose MixKD, a data-agnostic distillation framework that utilizes Mixup, a data augmentation technique. MixKD enables the student model to mimic the behavior of the teacher model not only on original training examples but also on linear interpolations of example pairs.", "method": "To confirm MixKD's effectiveness, the authors conduct experiments on the GLUE benchmark, including trials in a limited-data setting, and ablation studies.", "outcome": "The experiments showed that MixKD consistently outperforms standard KD training and several competitive baselines on the GLUE benchmark. They further revealed the advantages of the proposed approach in limited-data settings.", "future_impact": "Although the abstract does not specify future impact, the proposed MixKD could potentially be beneficial for compressing large models to be compatible on lower resource platforms, thereby improving the wider applicability of large-scale language models.", "venue": "ICLR", "year": 2021, "title": "MixKD: Towards Efficient Distillation of Large-scale Language Models"} +{"pid": "600831309e795ed227f530a1", "context": "Recent work has shown that various forms of convolutional kernel methods can compete with standard supervised deep convolutional networks on datasets like CIFAR-10, obtaining accuracies in the range of 87-90% while being more amenable to theoretical analysis.", "key_idea": "The authors highlight the importance of a data-dependent feature extraction step, typically a whitened dictionary of patches, which gives rise to data-driven convolutional kernel methods. This is said to be the key ingredient for the high performance of these methods.", "method": "The authors extensively study the effect of data-driven feature extraction in convolutional kernel methods. They perform experiments using one of the simplest instances of such kernel methods which is based on a single layer of image patches followed by a linear classifier. They also scale this method to the challenging ImageNet dataset and conduct experiments to analyze the dictionary used.", "outcome": "The simple approach based on a single layer of image patches followed by a linear classifier achieves classification accuracies on CIFAR-10 in the same range as previous more sophisticated convolutional kernel methods. On the ImageNet dataset, this approach exceeds all existing non-learned representation methods and set a new baseline for object recognition without representation learning methods.", "future_impact": "This paper initiates the investigation of convolutional kernel models on ImageNet which could inspire further research in the field.", "venue": "ICLR", "year": 2021, "title": "The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods"} +{"pid": "600831319e795ed227f530a2", "context": "Building interpretable parameterizations of real-world decision-making on the basis of demonstrated behavior is essential for introspecting and auditing policies in different institutions. Real-world decision-making often cannot accommodate active experimentation, like in healthcare.", "key_idea": "The authors propose learning explanations of expert decisions by modeling their reward function in terms of preferences with respect to 'what-if' outcomes. They integrate counterfactual reasoning into batch inverse reinforcement learning to learn cost-benefit tradeoffs associated with expert actions.", "method": "They perform illustrative experiments in both real and simulated medical environments using batch, counterfactual inverse reinforcement learning approach.", "outcome": "The authors find that their approach effectively recovers accurate and interpretable descriptions of behavior.", "future_impact": "The counterfactual inverse reinforcement learning approach can be adapted in scenarios where expert policies depend on histories of observations rather than just current states, impacting future policy evaluation and decision making in various fields.", "venue": "ICLR", "year": 2021, "title": "Learning \"What-if\" Explanations for Sequential Decision-Making"} +{"pid": "5f6dc7fe91e01153370055f3", "context": "Recent network pruning methods focus on pruning models early-on in training using importance measures that were originally designed to prune trained models. Despite lacking justification for their use early-on in training, these measures result in surprisingly low accuracy loss.", "key_idea": "The authors develop a general framework that uses gradient flow to unify state-of-the-art importance measures through the norm of model parameters, aiming to better understand and optimize the network pruning in early training stage.", "method": "The authors establish the relationships between pruning measures and evolution of model parameters in their framework. These relationships are then validated on VGG-13, MobileNet-V1, and ResNet-56 models trained on CIFAR-10/CIFAR-100.", "outcome": "The study shows magnitude-based pruning, loss-preservation based pruning, and gradient-norm based pruning each have different impacts to the model training processes and final performances.", "future_impact": "The results and insights from the novel gradient flow framework could provide guidance for developing more effective network pruning methods in controlling the model training and evolution in the early stage.", "venue": "ICLR", "year": 2021, "title": "A Gradient Flow Framework For Analyzing Network Pruning"} +{"pid": "615e657b5244ab9dcbf21f3f", "context": "Unsupervised skill discovery algorithms are reinforcement learning techniques that learn a set of policies without access to a reward function, resembling representation learning algorithms in supervised learning. Despite advances, there is still little understanding about the optimality of these skill learning algorithms.", "key_idea": "The paper introduces the notion that unsupervised skill discovery algorithms based on mutual information maximization are not optimal for every possible reward function. Instead, they propose that the distribution over skills offers an optimal initialization that minimizes regret against adversarially-chosen reward functions, conditioned on a specific adaptation procedure.", "method": "The authors carry out an analytical discourse on unsupervised skill discovery algorithms based on mutual information maximization. Their work involves mathematical and geometric analysis of these systems.", "outcome": "This paper reveals that while the skills learned via unsupervised skill discovery algorithms based on mutual information maximization are not optimal for every possible reward function, the distribution over skills does provide an optimal initialization, reducing regret against adversarially-chosen reward functions with a specific adaptation process.", "future_impact": "The geometric perspective provided on these skill learning methods could influence the understanding and subsequent developments in reinforcement learning and unsupervised learning.", "venue": "ICLR", "year": 2022, "title": "The Information Geometry of Unsupervised Reinforcement Learning"} +{"pid": "60c2d40691e0117e30ca27d2", "context": "In real-time public health forecasting, data collection is non-trivial and demanding task. Often, data undergoes multiple revisions before reaching a stable value, due to human or technical constraints. This 'backfill' phenomenon and its effect on model performance has barely been studied in the prior literature.", "key_idea": "The paper introduces the multi-variate backfill problem, using COVID-19 as the example, and proposes Back2Future, a novel neural framework that aims to refine a given model's predictions in real-time leveraging the patterns in backfill dynamics.", "method": "A detailed dataset is constructed composed of relevant signals over the past year of the COVID-19 pandemic. The backfill dynamics were systematically characterized and leveraged to formulate a novel problem and neural framework. Extensive experiments are performed to compare the performance of Back2Future with existing models and non-trivial baselines.", "outcome": "The experiments demonstrate that the method refines the performance of top models for COVID-19 forecasting, yielding 18% improvement over baselines, enabling new state-of-the-art performance. It also improves model evaluation, allowing policy-makers to understand the true accuracy of forecasting models in real-time.", "future_impact": "This study lays the groundwork for future study into the impact and management of backfill dynamics in real-time public health data, and the improvement of predictive model accuracy.", "venue": "ICLR", "year": 2022, "title": "Back2Future: Leveraging Backfill Dynamics for Improving Real-time Predictions in Future"} +{"pid": "61b95d115244ab9dcbf12e3b", "context": "Score-based generative models (SGMs) rely on a diffusion process that gradually perturbs the data towards a tractable distribution. The complexity of this denoising task is uniquely determined by the diffusion process. Current SGMs employ simplistic diffusions which limit the performance of generative modeling.", "key_idea": "A novel critically-damped Langevin diffusion (CLD) is proposed, based on connections to statistical mechanics. It allows for superior performance in score-based generative models by introducing an easier task - learning the score function of the velocity given data - instead of learning scores of the data directly.", "method": "The authors derived a novel score matching objective for CLD, introducing a new sampling scheme for efficient synthesis from CLD-based diffusion models. They compare the resulting models with existing SGMs using similar network architectures and compute budgets.", "outcome": "The CLD-based SGMs were found to outperform previous models in synthesis quality. The novel sampler for CLD also significantly outperformed other solvers such as Euler\u2013Maruyama.", "future_impact": "This framework provides new insights into score-based denoising diffusion models and may be readily used for high-resolution image synthesis.", "venue": "ICLR", "year": 2022, "title": "Score-Based Generative Modeling with Critically-Damped Langevin Diffusion"} +{"pid": "5feeeb6c91e0113b2659fdac", "context": "Knowledge distillation (KD) is crucial for training non-autoregressive translation (NAT) models, using an autoregressive teacher model to reduce data complexity. However, a side effect of this training is the propagation of lexical choice errors on low-frequency words from the teacher model to the NAT model.", "key_idea": "The study proposes exposing raw data to NAT models in order to restore the information of low-frequency words that is missed in the distilled data; an extra Kullback-Leibler divergence term is introduced to compare the lexical choice of the NAT model with that embedded in the raw data.", "method": "The authors experiment across different language pairs and model architectures to validate their proposed approach.", "outcome": "The proposed approach has been successful in reducing lexical choice errors on low-frequency words and has pushed the state-of-the-art (SOTA) NAT performance on the WMT14 English-German and WMT16 Romanian-English datasets up to 27.8 and 33.8 BLEU points, respectively.", "future_impact": "The codes and models will be released, potentially enabling other researchers to use and build upon this work in the future.", "venue": "ICLR", "year": 2021, "title": "Understanding and Improving Lexical Choice in Non-Autoregressive Translation"} +{"pid": "8b57815a-52ef-4dd8-a7ab-212f167fda38", "context": "Residual networks (Resnets) have become a prominent architecture in deep learning, but there remains a need for a comprehensive understanding of Resnets. A previous view argues that Resnets perform iterative refinement of features.", "key_idea": "The authors further study the view that Resnets perform iterative refinement of features, and formalize the notion of iterative refinement in Resnets by showing how they naturally encourage features to move along the negative gradient of the loss during the feedforward phase.", "method": "The authors evaluate Resnets both analytically and empirically, and analyze layer behavior in terms of representation learning and iterative refinement.", "outcome": "The empirical analysis suggests that Resnets perform both representation learning and iterative refinement, with representation learning behavior predominantly happening in the first few layers, and iterative refinement happening in later layers. It is also observed that na\u00efvely sharing residual layers can lead to representation explosion and harm generalization performance.", "future_impact": "The authors suggest that simple existing strategies can help alleviate the problem of representation explosion caused by sharing residual layers, which could inform future enhancements to the Resnet architecture.", "venue": "ICLR", "year": 2018, "title": "Residual Connections Encourage Iterative Inference"} +{"pid": "5f915b3b91e011126509be03", "context": "Group equivariant convolutional networks (GCNNs) have shown improved performance by incorporating symmetry priors. While the G-steerability constraint, a key feature of GCNN, has been derived, a general characterization of G-steerable kernel spaces is still missing.", "key_idea": "The authors draw an analogy between the constraints underlying steerable kernels and spherical tensor operators from quantum mechanics to provide a general characterization of G-steerable kernel spaces for compact groups.", "method": "The authors generalize the Wigner-Eckart theorem for spherical tensor operators to provide a mathematical proof and characterization of steerable kernel spaces.", "outcome": "The authors successfully characterize steerable kernel spaces in terms of 1) generalized reduced matrix elements, 2) Clebsch-Gordan coefficients, and 3) harmonic basis functions on homogeneous spaces.", "future_impact": "This general characterization for compact groups can potentially help in understanding, designing, and improving models based on group equivariant convolutional networks.", "venue": "ICLR", "year": 2021, "title": "A Wigner-Eckart Theorem for Group Equivariant Convolution Kernels"} +{"pid": "6257c5b85aee126c0f4695a2", "context": "Stochastic Variational Inference is a dominant framework for learning large-scale probabilistic latent variable models, but typical assumptions on the factorization or independence of the latent variables may restrict its capacity for inference and generative modeling. The current trend is to build more expressive variational models by designing deep hierarchies of interdependent latent variables. However, these very deep models are facing diminishing returns in performance improvement and increased inference and training time.", "key_idea": "This paper proposes deep attentive variational autoencoder, which uses attention mechanisms to build more expressive variational distributions in deep probabilistic models, by explicitly modeling both local and global interactions in the latent space.", "method": "The deep attentive variational autoencoder is tested on a variety of established datasets to measure its performance and impact on training time.", "outcome": "The deep attentive variational autoencoder achieves state-of-the-art log-likelihoods while using fewer latent layers and requiring less training time than existing models.", "future_impact": "The proposed non-local inference mechanism could potentially reduce the computational footprint of future models by relieving the need for deep hierarchies.", "venue": "ICLR", "year": 2022, "title": "Deep Attentive Variational Inference"} +{"pid": "5ef3247a91e0110c353da9fe", "context": "The key challenge in adversarial robustness is the lack of a precise mathematical characterization of human perception used to define imperceptible adversarial attacks. Current attacks and defenses rely on restrictive threat models. Models robust against one type of threat remain vulnerable to others.", "key_idea": "The authors propose adversarial training against all imperceptible adversarial examples using an approximation by deep neural network, called neural perceptual threat model (NPTM). It considers adversarial examples with a bounded neural perceptual distance to natural images.", "method": "The authors perform an extensive perceptual study to validate the Neural Perceptual Threat Model (NPTM) and develop novel perceptual adversarial attacks and defenses. They test Perceptual Adversarial Training (PAT) on CIFAR-10 and ImageNet-100 against five diverse adversarial attacks.", "outcome": "Perceptual Adversarial Training (PAT) has been found to correlate well with human judgement of perceptibility and achieves state-of-the-art robustness against a union of five diverse adversarial attacks, resulting in more than double the accuracy over the next best model, without training against any of them.", "future_impact": "PAT demonstrates the potential for adversarial training defences to generalize well to unforeseen perturbation types, which can be vital in sensitive applications where a specific threat model cannot be assumed.", "venue": "ICLR", "year": 2020, "title": "Perceptual Adversarial Robustness: Defense Against Unseen Threat Models"} +{"pid": "d861e520-0885-48f4-b8ba-893592e4c767", "context": "Recommender systems are widely used for suggesting products and services to customers based on their historical behavior or preferences. However, handling large dimensionality of input and output data and capturing customer's behavior changes over time are challenges in building effective recommender systems.", "key_idea": "The authors propose a recommender system that uses a two-layer neural network for personalized recommendations, which encodes historical customer behavior to predict future behavior. It includes a convolutional layer for learning the importance of purchase history based on purchase date.", "method": "The authors validate their proposition through offline experiments comparing seasonality changes, online A/B tests, and by implementing their model across all digital categories. They also describe their production pipeline and how model enhancements have been handled.", "outcome": "The two-layer neural network model outperforms other modeling techniques including the existing production recommender. The model demonstrated scalability across all digital categories and yielded significant improvements as confirmed in an online A/B test.", "future_impact": "The authors have open-sourced their deep learning library which supports multi-gpu model parallel training. This feature is identified as crucial for building neural network based recommenders handling large dimensionality of input and output data.", "venue": "ICLR", "year": 2018, "title": "THE EFFECTIVENESS OF A TWO-LAYER NEURAL NETWORK FOR RECOMMENDATIONS"} +{"pid": "6296d9145aee126c0f7312c6", "context": "Large-scale pre-trained transformers have been successful in text and text-to-image generation, however, their application to video generation faces challenges due to the high computational cost, scarcity and low relevance of text-video datasets.", "key_idea": "The authors present CogVideo, a large-scale transformer model for text-to-video generation trained by inheriting a pre-trained text-to-image model called CogView2.", "method": "The authors propose the use of a multi-frame-rate hierarchical training strategy to better align text and video clips.", "outcome": "CogVideo is able to outperform all publicly available models in machine and human evaluations.", "future_impact": "As probably the first open-source large-scale pre-trained text-to-video model, CogVideo may pave the way for further advances in the field of text-to-video generation.", "venue": "ICLR", "year": 2022, "title": "CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers"} +{"pid": "5f9bdd3391e011dcf482d864", "context": "EEG-based neural measures have been shown to function as reliable objective correlates of depression. However, their clinical utility has been limited due to the lack of automated ways to deal with the inherent noise associated with EEG data at scale and ignorance about the aspects of the EEG signal that may be markers of a clinical disorder.", "key_idea": "The authors adapted an unsupervised pipeline from the deep representation learning literature to learn a disentangled representation using \u03b2-VAE to denoise the EEG signal and to extract interpretable features associated with a sparse set of clinical labels using a Symbol-Concept Association Network (SCAN).", "method": "The authors validate their method by comparing it with the canonical hand-engineered baseline classification method on a number of factors, including participant age and depression diagnosis. The method also includes a way to automatically extract denoised Event Related Potentials (ERPs) from novel, single EEG trajectories.", "outcome": "The adapted pipeline outperforms the canonical hand-engineered baseline classification method on several factors including participant age and depression diagnosis. It can be used to extract denoised ERPs from novel, single EEG trajectories, and supports fast supervised re-mapping to various clinical labels.", "future_impact": "The learned disentangled representations can serve as meaningful markers for clinical factors as automatically detected by SCAN, potentially making it easier for human interpretability and post-hoc expert analysis of the recommendations made by the model.", "venue": "ICLR", "year": 2021, "title": "Representation learning for improved interpretability and classification accuracy of clinical factors from EEG"} +{"pid": "60fe28ab5244ab9dcb2f1660", "context": "Deep neural networks (DNNs) are vulnerable to adversarial examples which aim to mislead DNNs. Defense mechanisms including empirical and theoretical defense approaches have been explored widely for a single model, but less understanding on the robustness of ensemble ML models.", "key_idea": "This paper provides an analysis and certified robustness for ensemble ML models, and identify the suffered conditions of robustness for different ensemble protocols. A new training technique, Diversity Regularized Training (DRT), is introduced to improve certifiably robust ensemble ML models.", "method": "The authors conduct extensive experiments using DRT enhanced ensembles, and evaluates the performance using L2-robustness on MNIST, CIFAR-10, and ImageNet datasets.", "outcome": "Empirical results indicate that the DRT enhanced ensembles can consistently achieve higher certified robustness in comparison to existing single and ensemble ML models, demonstrating the state-of-the-art certified L2-robustness on the tested datasets.", "future_impact": "The theoretical findings and the proposed Diversity Regularized Training (DRT) provides a path for further improvements to the robustness of ensemble ML models against adversarial examples.", "venue": "ICLR", "year": 2022, "title": "On the Certified Robustness for Ensemble Models and Beyond"} +{"pid": "6385788490e50fcafdf4998e", "context": "Traditional Transformer-based models for multivariate time series forecasting have limitations in terms of efficiency and long-term forecasting accuracy.", "key_idea": "The authors propose an efficient design of Transformer-based models for time series forecasting, featuring subseries-level patch segmentation of input time series tokens and a channel-independent structure where each univariate time series within the channels shares the same embedding and Transformer weights.", "method": "The author utilizes their proposed model, channel-independent patch time series Transformer (PatchTST), for multivariate time series forecasting and self-supervised representation learning tasks. They compare its long-term forecasting accuracy against state-of-the-art Transformer-based models.", "outcome": "The proposed PatchTST shows improvements in long-term forecasting accuracy compared to state-of-the-art Transformer-based models. It also performs excellently in self-supervised pre-training tasks, outperforming supervised training on large datasets.", "future_impact": "It is implied that the introduced PatchTST model could significantly improve the efficiency and accuracy of time series forecasting, with potential to offer excellent performance in self-supervised learning tasks. The strategy of incorporating pre-trained representations could enhance the forecasting accuracy of different datasets.", "venue": "ICLR", "year": 2022, "title": "A Time Series is Worth 64 Words: Long-term Forecasting with Transformers"} +{"pid": "620b19c85aee126c0f7e6daf", "context": "Advanced representation learning techniques depend on reliable and general evaluation methods. Several algorithms based on geometric and topological analysis of a manifold approximated from the learned data representations have been proposed, but they often fail when it comes to challenging geometric arrangements of representations such as clusters with varying shape and density and outliers.", "key_idea": "The authors introduce Delaunay Component Analysis (DCA), an evaluation algorithm that approximates the data manifold using a more suitable neighbourhood graph called the Delaunay graph, to provide reliable manifold estimation even for challenging geometric arrangements.", "method": "They validate the proposed DCA method experimentally on representations obtained from neural networks trained with contrastive objective, supervised and generative models.", "outcome": "The DCA method is successfully validated on various models, demonstrating its usefulness in evaluating different data representations, specifically dealing with challenging geometric arrangements.", "future_impact": "The framework introduced in this study opens the avenue for more precise assessment of the quality of individual novel data representations, which may contribute to more reliable and generalizable representation learning techniques.", "venue": "ICLR", "year": 2022, "title": "Delaunay Component Analysis for Evaluation of Data Representations"} +{"pid": "6257c5ad5aee126c0f468261", "context": "Learning in games, particularly using standard online learning algorithms, is a relevant field in AI and ML, but has yet to have the behavior of a probability distribution or measure over a set of initial conditions figured out. Existing research has not explored in depth how uncertainty about the initial conditions effects the outcomes of these systems.", "key_idea": "The paper investigates the evolution of uncertainty in learning-in-game systems, using differential entropy as a measure. Importantly, the study focuses its analysis to the uncertainty of initial conditions rather than their single instances.", "method": "The authors applied volume analysis and differential entropy to study the evolution of uncertainty in learning-in-game systems, particularly in zero-sum games using MWU and FTRL algorithms. Further, the framework was also applied to study related systems and different standard online optimization algorithms in numerous games.", "outcome": "The authors found that the differential entropy of these learning-in-game systems increases linearly with time, indicating an increase in their unpredictability over time.", "future_impact": "The developed framework and findings could be applied to more in-depth studies of uncertainties in other learning systems and optimization algorithms, shedding light on the dynamic nature of game-theoretic and ML systems.", "venue": "ICLR", "year": 2022, "title": "The Evolution of Uncertainty of Learning in Games"} +{"pid": "6257c5ba5aee126c0f4698c2", "context": "Model-Agnostic Meta-Learning (MAML) aims to find initial weights that allow fast adaptation to new tasks. However, how to choose the adaptation (inner loop) learning rate in MAML and its effect on the adaptation error is less explored.", "key_idea": "The authors aim to study the effect of the adaptation learning rate in MAML, provide a way to estimate the optimal rates minimizing the population risk, interpret the underlying dependence between the optimal adaptation learning rate and the input data, and compare MAML with empirical risk minimization (ERM).", "method": "The study is conducted using mixed linear regression to analyze the effect of adaptation learning rates. Analysis includes estimating optimal adaptation learning rates, interpreting the dependence between learning rate and input data, and comparing MAML with ERM.", "outcome": "The study shows that MAML produces an initialization with a smaller average distance to the task optima as compared to ERM, consistent with previous practical findings. The findings are supported by numerical experiments.", "future_impact": "This research could guide the selection of the adaptation learning rate in MAML, thus improving the algorithm's efficiency and performance.", "venue": "ICLR", "year": 2022, "title": "Unraveling Model-Agnostic Meta-Learning via The Adaptation Learning Rate."} +{"pid": "33adc30f-fbbd-4a05-a81d-965fa97705ef", "context": "The problem in reinforcement learning is that it often fails to solve unseen tasks, generalize to larger map sizes, and navigate in dynamic environments.", "key_idea": "A new module called Value Propagation (VProp) is proposed, which is a parameter-efficient differentiable planning module based on Value Iteration that can be trained via reinforcement learning to overcome the stated problems.", "method": "The authors evaluate the proposed module on MazeBase grid-worlds, randomly generated environments of several different sizes including ones with stochastic elements.", "outcome": "The VProp showed successful results, demonstrating its capability to generalize to larger map sizes and navigate in dynamic environments including ones with stochastic elements.", "future_impact": "The proposed model provides a cost-efficient learning system to build low-level size-invariant planners for a variety of interactive navigation problems.", "venue": "ICLR", "year": 2018, "title": "Value Propagation Networks"} +{"pid": "04909651-234f-4982-b5da-ed3453562264", "context": "The optimization of deep neural networks can be highly challenging due to the non-convex nature of the loss function, which can involve problematic landscapes like saddle-surfaces that are difficult to overcome using simple gradient descent algorithms.", "key_idea": "The paper proposes a method to optimize highly non-convex neural network objectives, where the training starts with a smoothed or mollified objective function. The complexity of the function increases as the training proceeds and eventually becomes the original, difficult-to-optimize function.", "method": "The method's effectiveness is validated through its application to various challenging optimization tasks. Improvements are shown in these tasks, and a relationship is established between recent works on continuation methods for neural networks and mollifiers.", "outcome": "The paper evidenced improvements on various difficult optimization tasks by employing the proposed mollifying networks methodology.", "future_impact": "The study established a connection between continuation methods and mollifiers in neural networks, offering new insights that could guide future optimization methods in similar domain.", "venue": "ICLR", "year": 2017, "title": "Mollifying Networks"} +{"pid": "865ffdcd-237d-4d1d-af0a-87c220639c72", "context": "Convolutional Neural Networks (CNNs) have become fundamental in computer vision, yet most researchers rely on pre-trained models due to the difficulty of properly initializing these networks from scratch. Poor initialization can lead to issues such as vanishing or exploding gradients.", "key_idea": "The authors propose a fast and simple data-dependent initialization procedure that sets the weights of a network such that all units in the network train at roughly the same rate, avoiding the vanishing or exploding gradient problem.", "method": "The new procedure was tested on standard computer vision tasks, such as image classification and object detection.", "outcome": "The new initialization procedure matches the current state-of-the-art unsupervised or self-supervised pre-training methods on standard computer vision tasks, while being three orders of magnitude faster. When used with pre-training methods, the initialization procedure even outperforms prior work.", "future_impact": "The method introduced could help close the gap between supervised and unsupervised pre-training, providing a fast and simpler approach to initializing CNNs.", "venue": "ICLR", "year": 2016, "title": "Data-dependent Initializations of Convolutional Neural Networks"} +{"pid": "5e5e18b993d709897ce2adaf", "context": "The effective use of deep neural networks in Multi-Task Reinforcement Learning has traditionally faced challenges due to individual task representation.", "key_idea": "The study leverages the assumption that learning from different tasks sharing common properties can help to generalize the knowledge of them, resulting in more effective feature extraction compared to learning a single task.", "method": "The authors extend the well-known finite-time bounds of Approximate Value-Iteration to the multi-task setting and propose multi-task extensions of three Reinforcement Learning algorithms. These models are then empirically evaluated on widely used Reinforcement Learning benchmarks.", "outcome": "The paper provides theoretical guarantees highlighting conditions for which sharing representations is convenient among tasks. The proposed multi-task extensions show significant improvements over the single-task counterparts in terms of sample efficiency and performance.", "future_impact": "The study suggests that sharing representations among tasks in Multi-Task Reinforcement Learning could lead to significant performance improvements, indicating potentially important implications for the design of future Reinforcement Learning algorithms.", "venue": "ICLR", "year": 2024, "title": "Sharing Knowledge in Multi-Task Deep Reinforcement Learning"} +{"pid": "f033385b-d5fa-46aa-8164-c948b6883d90", "context": "The Oz project, which focuses on creating interactive, dramatic stories, is working on developing characters for these stories. In traditional storytelling mediums like novels and movies, characters without clear emotions are deemed 'flat' and thus not believable to audiences.", "key_idea": "The authors are developing an agent architecture, Tok, and an emotion subsystem, Em, to enable artists to build emotional characters for interactive stories. The aim is to make characters appear emotional which adds believability.", "method": "The authors design emotions in their system to be generated by a flexible set of rules. These Em-generated emotions such as joy, distress, fear, hope and more, are connected to accomplishment or failure of character goals. The authors also use Behavioral Features (BFs) to determine how emotions affect other processes in their Tok system.", "outcome": "With their system, not all characters must act in the same way when they have the same emotion; it is possible to create a wide variety of characters, each reacting differently to the same emotion, which adds individualism to the characters.", "future_impact": "The work provides a significant foundation for creating emotional characters in interactive stories, which can lead to more flexible and believable storytelling in computer-controlled drama.", "venue": "AAAI", "year": 1994, "title": "Building emotional characters for interactive drama"} +{"pid": "bd2dea6b-4b7a-43e5-a7a2-a8e622e260c2", "context": "There is a growing interest in security games to handle resource allocation problems in security applications, especially patrolling security games (PSGs) where resources are mobile. However, one of the major challenges in security games is the development of scalable algorithms to handle real-world situations and current literature primarily focuses on heuristics and decomposition techniques.", "key_idea": "In contrast to existing approaches, this study proposes the use of abstractions in security games, specifically PSGs, as a means to design scalable algorithms. The authors also introduce the issue of searching for the optimal abstraction.", "method": "The authors define some classes of abstractions and provide parametric algorithms to automatically generate these abstractions. They also develop an anytime algorithm to find the optimal abstraction within these classes.", "outcome": "The study demonstrates that using abstractions can help in relaxing the patrolling strategies' Markovianity constraint, commonly found in PSGs, and assist in solving larger game instances.", "future_impact": "The use of abstractions and the automatic generation of these abstractions in PSGs could open a new direction for designing scalable algorithms in diverse security game scenarios.", "venue": "AAAI", "year": 2011, "title": "Automated abstractions for patrolling security games"} +{"pid": "5f7c463491e0117ac2a78a4c", "context": "People increasingly use social media, like Twitter, to communicate during disasters. However, processing and classifying these disaster-related events on social media efficiently and in real-time is a challenging task due to the extremely noisy and often short nature of user-generated messages.", "key_idea": "The authors propose a weakly supervised approach for rapidly creating classifiers for fine-grained event categories from Twitter messages, leveraging preceeding context tweets and reply tweets to enrich the tweet representations. Their key novelty lies in creating high-quality labeled data in a timely manner by clustering tweets containing an event keyword and asking a domain expert to quickly disambiguate word senses and label clusters.", "method": "They evaluate their approach on two hurricane events, Harvey and Florence, with the aim of outperforming supervised classifiers that are trained with significantly more annotated tweets.", "outcome": "The evaluation showed that with only 1-2 person-hours of human supervision, the proposed weakly supervised classifiers outperformed the supervised classifiers trained using more than ten thousand annotated tweets, a process which took over 50 person-hours.", "future_impact": "This approach could help in the quick and productive use of social media data for disaster management and allow for significant time-saving in critical scenarios.", "venue": "AAAI", "year": 2020, "title": "Weakly-Supervised Fine-Grained Event Recognition On Social Media Texts For Disaster Management"} +{"pid": "07603f42-77d9-4b93-ad49-33a0402443c8", "context": "Previously, intelligent tutoring systems (ITSs) lacked centralized reasoning and were complex to improve or customize the teaching goals.", "key_idea": "The authors propose a two-agent machine learning architecture called ADVISOR for ITS. The first agent learns a model of how students behave in various contexts, while the second uses this model and reinforcement learning to derive a teaching policy that aligns with a specified educational goal.", "method": "The authors evaluate each component of ADVISOR separately and then evaluate the system as a whole with actual students to assess its ability to learn a teaching policy that meets a given educational goal.", "outcome": "Evaluation results demonstrated each individual agent functioned adequately. The complete architecture, when tested with actual students, was successful in learning and applying a teaching policy that met the educational objective.", "future_impact": "Although ADVISOR was integrated with a specific intelligent tutor, the authors suggest the technique can be broadly applied to a range of ITSs.", "venue": "AAAI", "year": 2000, "title": "Advisor: a machine-learning architecture for intelligent tutor construction"} +{"pid": "5e5e190093d709897ce44bbc", "context": "In dialog systems, dialog act recognition and sentiment classification are two correlative tasks that capture speakers' intentions. Existing systems either treat these as separate tasks or jointly model them without explicitly modeling their interaction and relation.", "key_idea": "The authors propose a Deep Co-Interactive Relation Network (DCR-Net) that explicitly considers the cross-impact and models the interaction between dialog act recognition and sentiment classification by introducing a co-interactive relation layer.", "method": "The authors evaluate the proposed model with different relation layers on two public datasets: Mastodon and Dailydialog. They also incorporate the Bidirectional Encoder Representation from Transformer (BERT) into their framework.", "outcome": "Experiments show that the proposed model outperforms the state-of-the-art joint model by 4.3% and 3.4% in terms of F1 score on dialog act recognition task, 5.7% and 12.4% on sentiment classification, respectively.", "future_impact": "The paper emphasizes the effectiveness of explicitly modeling the relation between the two tasks and the multi-step interaction mechanism, potentially informing future work in this field.", "venue": "AAAI", "year": 2020, "title": "DCR-Net: A Deep Co-Interactive Relation Network for Joint Dialog Act Recognition and Sentiment Classification."} +{"pid": "44810a55-f104-44b0-b31c-ba692d52b308", "context": "Currently arithmetic problem solving systems struggle to adapt to new types of problems without alterations to the program and lack a learning capacity for acquiring additional problem-solving rules.", "key_idea": "LAPS is proposed, as a problem-solving system with the capacity to solve many arithmetic problems written in natural languages through an existing pool of rules and to acquire new knowledge and rules when confronted with novel problems, following advice from a human user.", "method": "When LAPS cannot solve a given problem due to a shortage of knowledge, it queries the user on the problem solution method and acquires new knowledge and rules based on user advice.", "outcome": "With acquired and synthesized rules, LAPS successfully solved arithmetic problems presented to it.", "future_impact": "The ability of LAPS to continually learn and improve its problem-solving capabilities through interaction with a human user, will allow for a broader application of problem solving in natural languages beyond arithmetic problems.", "venue": "AAAI", "year": 1986, "title": "Learning arithmetic problem solver"} +{"pid": "64a1272f-fea4-430e-9ddd-cd8b1b295cce", "context": "Prior studies have examined performances in chess games by human players and computers, but there's room for deeper understanding of their preferences and how these impact outcomes.", "key_idea": "This paper examines the differences in preferences and performance between humans and computers in chess, including error scaling based on advantage, scoring expectations related to turn order, and tendency towards specific positions.", "method": "The authors conduct a distributional analysis of large data-sets of chess games played by humans and computers to identify differences in preferences and performance.", "outcome": "The study showed that humans made more errors and had lower scoring expectations when it was their turn to move. Humans also preferred to guide games into positions with fewer options and earlier resolutions, even when teamed with a computer.", "future_impact": "The findings provide new regularities of human and computer performances, which may have implications for decision-agent domains beyond chess.", "venue": "AAAI", "year": 2014, "title": "Human and Computer Preferences at Chess."} +{"pid": "6020e02a9e795e62379b0e1e", "context": "Previous work on text summarization in the scientific domain mainly focused on the content of the input document, and did not consider its citation network. However, scientific papers often contain uncommon domain-specific terms that can challenge a model's understanding.", "key_idea": "The authors redefine scientific paper summarization to utilize citation graph and propose a citation graph-based summarization model (CGSum) incorporating information from both the source paper and its references.", "method": "The authors construct a novel scientific papers summarization dataset Semantic Scholar Network (SSN) which contains 141K research papers in different domains and 661K citation relationships. They employ this dataset in conducting extensive experiments to validate their model.", "outcome": "The proposed model, even with its simple architecture, demonstrates competitive performance when compared with pretrained models. The results also highlight the role of the citation graph in understanding the content of papers and generating high-quality summaries.", "future_impact": "The research suggests that the citation graph is crucial for better understanding the content of papers and generating high-quality summaries and paves the way for future research in enhancing scientific paper summarization using citation networks.", "venue": "AAAI", "year": 2021, "title": "Enhancing Scientific Papers Summarization With Citation Graph"} +{"pid": "5f32769991e011bc1612b01a", "context": "In the Multi-Armed Bandit (MAB) problem, most algorithms focus on minimizing regret, which can lead to over-exploration in larger problems or brief interaction environments. These algorithms often struggle to find the optimal arm and therefore may produce undesirable results.", "key_idea": "The authors propose a novel, more lenient regret criterion for the MAB problem that disregards suboptimality gaps smaller than a certain threshold. Furthermore, they suggest a variant of the Thompson Sampling (TS) algorithm, referred to as (sic)-TS, that operates according to this new lenient regret criterion.", "method": "The authors demonstrate the performance and efficiency of the proposed (sic)-TS algorithm theoretically, proving its asymptotic optimality in terms of the lenient regret.", "outcome": "It is shown that when the mean of the optimal arm is high enough, the lenient regret of (sic)-TS is bounded by a constant, indicating efficiency under certain conditions.", "future_impact": "(sic)-TS can potentially improve performance in scenarios where the agent knows a lower bound of the suboptimality gaps, paving the way for better action decisions in MAB problem settings.", "venue": "AAAI", "year": 2021, "title": "Lenient Regret For Multi-Armed Bandits"} +{"pid": "619ba6861c45e57ce99e8f84", "context": "Assignments form the core of student learning experience in AI education and the need for innovative assignment designs is recognized by the community.", "key_idea": "The Abstract introduces the 'Model AI Assignments' session which is a platform that gathers and disseminates superior AI assignment designs from the educational community.", "method": "The session presents selected abstracts of six AI assignments from the 2021 session that meet specific criteria including adaptability, playful engagement, and flexibility.", "outcome": "The session successfully collected and presented six AI assignments which are adoptable, engaging and flexible for different instructor needs.", "future_impact": "The assignments presented can be used by instructors in the AI education community to enhance the student learning experience. The supporting resources and specifications for these assignments are available online for further adoption and adaptation.", "venue": "AAAI", "year": 2021, "title": "Model Ai Assignments 2021"} +{"pid": "ea855a02-65aa-44a2-9077-6f3389cba09b", "context": "The ongoing research has been focusing on the facility location problem under a model where monetary transfers between the mechanism and the agents are not possible (no-money constraint) and facilities serve different purposes (heterogeneity). Previous studies mostly use utilitarian optimization methods.", "key_idea": "This study proposes to use non-utilitarian optimization functions for such problems, specifically focusing on minimizing the maximum connection cost of the agents.", "method": "The authors investigate both deterministic and randomized algorithms and derive both lower and upper bounds for the approximability of strategyproof mechanisms.", "outcome": "The study derived lower and upper bounds regarding the approximability of strategyproof mechanisms when non-utilitarian optimization functions are applied in facility location problem.", "future_impact": "This study sets a precedent for using non-utilitarian optimization functions in heterogeneous facility location problems which could inspire future development of algorithms and mechanisms in this area.", "venue": "AAAI", "year": 2015, "title": "Truthful mechanisms without money for non-utilitarian heterogeneous facility location"} +{"pid": "5fdc802991e01104c9181085", "context": "Speech enhancement is challenging due to the diversity of background noise types. Most of the current methods are focused on modelling the speech rather than the noise.", "key_idea": "The authors propose a novel model called SN-Net that simultaneously models speech and noise using a two-branch convolutional neural network. One branch predicts speech and the other predicts noise. The model introduces interaction modules at several intermediate feature domains to benefit each other.", "method": "The authors evaluated their model using various evaluation metrics on public datasets.", "outcome": "The authors discovered that the interaction module played a key role in simultaneous modeling, and their proposed SN-Net outperformed the state-of-the-art methods by a large margin on various evaluation metrics.", "future_impact": "The proposed SN-Net also showed superior performance for speaker separation, suggesting potential for future use in speaker separation tasks.", "venue": "AAAI", "year": 2021, "title": "Interactive Speech And Noise Modeling For Speech Enhancement"} +{"pid": "39086091-d792-44f5-96fc-38b0061b20bc", "context": "Code clone detection has attracted significant attention from researchers, and many methods have been proposed. However, these existings methods lack robustness against adversarial revisions of the code, as they do not consider the dynamic 'contest' between plagiarists (the code clone creators) and detectors.", "key_idea": "This paper proposes a novel clone detection approach, Adversarial Clone Detection (ACD), which mimics the adversarial process between plagiarists and detectors. It is designed to build a robust clone detectector and also to model the behavior of plagiarists.", "method": "The authors model the adversarial process and use the learned policy of the plagiarist to construct the clone detector.", "outcome": "Experiments show that the plagiarist's learned policy can help in building a stronger clone detector, substantially outperforming existing clone detection methods.", "future_impact": "The understanding and modeling of plagiarist behavior may help identify the vulnerabilities of current software clone detection tools, potentially leading to improved clone detection methods.", "venue": "AAAI", "year": 2019, "title": "Find Me If You Can: Deep Software Clone Detection by Exploiting the Contest between the Plagiarist and the Detector"} +{"pid": "6020e0639e795e62379b0e44", "context": "Existing methods of extracting values from key fields in form-like documents often ignore the semantics of the keys, considering them only as class labels. This leaves them incapable to handle zero-shot keys. Although these models often leverage attention mechanisms, the learned features may not accurately reflect why humans would recognise the value for the key, and thus may not generalise well to new documents.", "key_idea": "The authors propose a Key-Aware and Trigger-Aware (KATA) extraction model which uses the input key to explicitly learn two mappings: from key representations to trigger representations, and then from trigger representations to values. These mappings could be intrinsic and invariant across different keys and documents.", "method": "The model uses a large training set constructed based on Wikipedia data to pre-train these two mappings. It is then fine-tuned to two applications.", "outcome": "The proposed KATA model achieves more than 70% accuracy for the extraction of zero-shot keys, at which all previous methods fail.", "future_impact": "This approach can improve the performance of many downstream applications such as knowledge base construction, question answering, document comprehension and could potentially generalize to new documents.", "venue": "AAAI", "year": 2021, "title": "Extracting Zero-Shot Structured Information From Form-Like Documents: Pretraining With Keys And Triggers"} +{"pid": "62bac4fc-ac83-401b-944a-ed8c31d13f2c", "context": "Propositional Power Default Reasoning (PDR) is a form of non-monotonic reasoning where the underlying logic is Kleene's 3-valued propositional logic, known for its complex skeptical entailment problem.", "key_idea": "PDR provides a more concise meaning of the problem of skeptical entailment, which in turn has better complexity characteristics than the usual formalisms.", "method": "An implementation called 'powdef' is created to encode and solve hard graph problems and explore instances of skeptical entailment generated randomly.", "outcome": "The implementation shows that PDR has better complexity characteristics than usual formalisms, with complexity co-NP(3)-Complete instead of ?2P-Complete.", "future_impact": "The improved efficiency of PDR in terms of complexity could be beneficial in solving hard graph problems and other applications requiring non-monotonic reasoning.", "venue": "AAAI", "year": 1998, "title": "Experimenting with power default reasoning"} +{"pid": "6215a4242c356815940386f9", "context": "Learning effective embeddings for irregularly sampled time-series that operate on different time scales is crucial for tasks like classification and clustering. Existing time-series similarity measures aren't valid distance metrics, making it challenging to accurately capture the relationships in multivariate time-series data and learn useful embeddings. Traditional Monte-Carlo-based expectation evaluations for measuring distances face difficulties due to high data dimensionality and the inability to sample from encapsulating distributions.", "key_idea": "The authors introduce a novel distance metric, I-SEA, that leverages Importance Sampling and Non-parametric Density Estimation. Unlike earlier measures, I-SEA focuses on adaptively estimating distances and enables implicit estimation from high-dimensional distributions, leading to increased accuracy and reduced variance.", "method": "The authors delineate the properties of I-SEA theoretically and verify its capabilities through experimental evaluations on real-world healthcare datasets.", "outcome": "I-SEA, the proposed distance metric demonstrated superior performance in tasks such as classification and representation learning and showed improved accuracy and reduced variance when evaluated on real-world healthcare datasets.", "future_impact": "The development of I-SEA opens new possibilities for learning effective embeddings for time-series data, potentially improving the performance of machine learning tasks such as classification and representation learning particularly in complex fields like healthcare.", "venue": "AAAI", "year": 2022, "title": "I-SEA: Importance Sampling and Expected Alignment-Based Deep Distance Metric Learning for Time Series Analysis and Embedding."} +{"pid": "5e71f4b491e0115656f5d254", "context": "Previous approaches to generating complex scenes with multiple objects from layouts face some problems. These include the generation of spurious objects not corresponding to bounding boxes in the layout and the merging of objects in the images due to overlapping bounding boxes in the layout.", "key_idea": "The main idea is the Object-Centric Generative Adversarial Network (OC-GAN), a new image generation method based on Scene-Graph Similarity Module (SGSM) that understands individual objects and their relationships. In addition, changes are proposed to the generator's conditioning mechanism to improve object instance-awareness.", "method": "Durability and performance of the proposed OC-GAN is evaluated through quantitative evaluation and ablation studies, using COCO-Stuff and Visual Genome datasets.", "outcome": "OC-GAN improves image quality, and alleviates previously identified failure modes. It outperforms past state-of-the-art techniques on both the COCO-Stuff and Visual Genome datasets.", "future_impact": "The paper introduces an object-centric adaptation of the Frechet Inception Distance metric, called Scene-FID. This new metric, being more suited for multi-object images, could be influential in the evaluation of future layout-to-image generation methods.", "venue": "AAAI", "year": 2021, "title": "Object-Centric Image Generation From Layouts"} +{"pid": "eed27a42-b4ce-43e1-9efd-f8498b1d790b", "context": "The task of inferring causal relationships from patterns of dependencies in the context of constructing probabilistic networks from human judgments is less understood, particularly determining the directionality of interactions from non-temporal probabilistic information.", "key_idea": "This paper develops efficient algorithm for inferring the directionality of interactions from non-temporal probabilistic information, a crucial step in attributing a causal interpretation to these interactions.", "method": "An efficient algorithm is developed that recovers the structure of the underlying causal polytree given data generated by an undisclosed causal polytree and also determines the directionality of all its identifiable links.", "outcome": "The authors created an efficient algorithm capable of determining the directionality of identifiable links and the underlying structure of causal polytrees.", "future_impact": "The development of this algorithm could pave the way for improved and more accurate inference of causal relationships from patterns of dependency, which would have implications in the construction of probabilistic networks from human judgments.", "venue": "AAAI", "year": 1990, "title": "Learning causal trees from dependence information"} +{"pid": "94045a47-8fc7-4582-a620-424f81ae2578", "context": "While rescaling the classes according to their misclassification costs is a popular approach to cost-sensitive learning in binary-class problems, recent studies show that it is often not effective when applied to multi-class problems directly.", "key_idea": "This paper proposes a new approach to rescaling for multi-class cost-sensitive learning, which should be chosen if the user wishes to utilize rescaling for such problems. Before applying rescaling, however, the consistency of the costs must be examined.", "method": "The authors analyze why the traditional rescaling approach often proves ineffective on multi-class problems, and through this analysis, introduce a new approach for using rescaling in multi-class cost-sensitive learning.", "outcome": "The authors demonstrate that the proposed method is effective when unequal misclassification costs and class imbalances occur simultaneously and that it can also handle pure class-imbalance learning.", "future_impact": "This new approach provides a unified framework for addressing multi-class cost-sensitive learning as well as multi-class class-imbalance learning, paving the way for further research in these areas.", "venue": "AAAI", "year": 2006, "title": "On multi-class cost-sensitive learning"} +{"pid": "604098ae-18e4-4c29-81da-4749941c66b3", "context": "Relational learning has enabled new machine learning applications, but efficient knowledge extraction from such networks is a challenge. Specifically, the CILP++ system that depends on the BCP propositionalization algorithm struggles with the issue that the features generated by BCP lack independent relational description which hinders sound knowledge extraction.", "key_idea": "The paper presents a methodology for generating independent propositional features for BCP using semi-propositionalization of bottom clauses, aiming to allow proper relational knowledge representation and extraction from CILP++ networks.", "method": "The proposed semi-propositionalization method is tested through empirical results, comparing accuracy and runtimes with the original version of BCP.", "outcome": "The new semi-propositionalization approach exhibits comparable accuracy and runtime to the original BCP while enabling proper relational knowledge representation of features.", "future_impact": "The proposed approach will improve knowledge extraction in neural-symbolic systems like CILP++, which will enhance the application of these systems in various areas such as link analysis in social networks and graph mining.", "venue": "AAAI", "year": 2015, "title": "Neural Relational Learning Through Semi-Propositionalization of Bottom Clauses"} +{"pid": "1c5a8cff-9ae4-4e3c-b36e-7419ee717aee", "context": "Prior to this study, the design of algorithms has been a manual and expertise-intensive process, necessitating the need for an automated system to design algorithms. Previous works have attempted to design such a system with mixed results.", "key_idea": "The author proposes a novel system for designing algorithms, which operates in two problem spaces -- one focusing on the algorithm itself and the other on the task-domain. The partially worked out algorithms are represented as configurations of dataflow components that construct and modify the representations.", "method": "The author provides a detailed description of the automatic algorithm-design system, highlighting its use of general-purpose operators and how they are adapted to different situations using instantiation and means-ends analysis rules. The system also utilizes symbolic and test-case execution rules that drive the component-refinement process by exposing problems and opportunities.", "outcome": "Though the author describes the implementation of the system and provides a specification for it, the system's performance is not evaluated in the paper, and therefore there are no reported outcomes.", "future_impact": "This work could contribute to the development of more advanced and efficient algorithm-design systems, potentially automating the entire process of algorithm design and testing.", "venue": "AAAI", "year": 1983, "title": "An automatic algorithm designer: an initial implementation"} +{"pid": "6020e22f9e795e62379b5a9f", "context": "The teacher-student framework in reinforcement learning deploys an advising mechanism where a teacher advises the student about the optimal action to take in a given state. This mechanism, while effective, doesn't allow the teacher to give more informative and nuanced advice about the state of the environment", "key_idea": "The authors extend the current advising framework to allow the teacher to provide not only the optimal action, but also a qualitative assessment of the state. A novel architecture, Advice Replay Memory (ARM), is introduced to effectively reuse the advice provided by the teacher.", "method": "The authors tested their approach on multiple Atari 2600 games using a fixed set of hyper-parameters, looking at performance of students when provided with different advising budget and teacher quality.", "outcome": "Experiments demonstrate that even a student advised by a sub-optimal teacher can achieve significant performance increases and ultimately outperform the teacher. The approach outperforms baselines even with comparatively suboptimal teachers and a small advising budget.", "future_impact": "The authors' contributions, including supplementing a student's knowledge, introduction of ARM, performance boost with coarse state categorization, and enabling a student to outperform teacher, could guide future enhancements to the teacher-student framework.", "venue": "AAAI", "year": 2021, "title": "An Enhanced Advising Model In Teacher-Student Framework Using State Categorization"} +{"pid": "4f7efc69-f01a-4043-b458-58b7323e01ec", "context": "The stable model semantics was recently generalized for the full first-order language, a syntax translation approach akin to McCarthy's circumscription. However, the decidability and undecidability of first-order language fragments under both stable model semantics and circumscription remain unclear.", "key_idea": "The paper investigates the decidability and undecidability of various fragments of first-order language under both semantics of stable models and circumscription, aiming to identify maximally decidable and undecidable classes.", "method": "The researchers explored different fragments of first-order language under the semantics of stable models and circumscription to identify decidable and undecidable classes.", "outcome": "The paper reveals that the boundaries between decidability and undecidability for the two semantics are distinct despite the similarity in definitions. Furthermore, it is found that for all fragments considered in this study, decidability under the semantics of circumscription aligns with that in classical first-order logic.", "future_impact": "The discrepancies observed in the boundaries between decidability and undecidability for these two semantics could influence future explorations in this domain.", "venue": "AAAI", "year": 2010, "title": "Decidable fragments of first-order language under stable model semantics and circumscription"} +{"pid": "a2440b36-703e-428a-abc8-53cf84062a9c", "context": "The exact computation of probabilities of missing deadlines in series-parallel schedules is NP-hard. Existing methods for approximation of discrete random variables are not efficient and optimal.", "key_idea": "This paper presents an efficient algorithm for creating an approximation of a discrete random variable with a support size limited to a specific number (m). This approximation minimizes the Kolmogorov distance from the original random variable.", "method": "The author presented several variants of the main algorithm and conducted a detailed empirical evaluation of their performance and complexity.", "outcome": "The author showed that the proposed algorithms perform efficiently and effectively in practice, as demonstrated through empirical evaluations.", "future_impact": "The algorithms proposed in this paper could be used for estimating the probability of missing timelines in series-parallel schedules, potentially providing more efficient and accurate approximations for these complex calculations.", "venue": "AAAI", "year": 2019, "title": "Efficient Optimal Approximation of Discrete Random Variables for Estimation of Probabilities of Missing Deadlines."} +{"pid": "7738d793-ae51-4cbb-a8c4-fa60ba34c0f0", "context": "Interpreting temporal relations between events mentioned in narratives (such as which event occurs before another) has been a challenge. Previous works lacked rigorous descriptions for time relations of tenses and did not provide deterministic interpreting algorithms.", "key_idea": "This paper presents a three-level language concept-based algorithm for interpreting temporal relations in narrative text, and a novel, more rigorous description for time relations of tences is introduced, which can express all 16 tenses in English and incorporate the interval properties of events into the relations.", "method": "The algorithm decides the relations through aspectual information for verbs, time relations for tenses, and time relations between clauses and/or sentences. In interpreting time relations between clauses, the use of anaphoric references to events and the concept of situational description for an event is emphasized. To make the algorithm deterministic, a fixed rather than parallel order for applying interpreting rules was proposed.", "outcome": "The authors suggest a tree-like structure for the representation of temporal relations between events, which allows us to include vaguely specified relations, and facilitates the interpretation of subsequent temporal relations.", "future_impact": "Since the interpretation of temporal relations between events may be clarified later, this work can improve subsequent interpretations of temporal relations in narrative texts.", "venue": "AAAI", "year": 1988, "title": "The interpretation of temporal relations in narrative"} +{"pid": "6020e3cd9e795e62379b5ba5", "context": "Predictive uncertainty in humans is highly interpretable but poorly accessible. The predictive uncertainty of machine learning models, while poorly interpretable, is relatively easily accessible. Estimating human predictive uncertainty is crucial to improve sampling efficiency and prediction reliability of human-in-the-loop systems used in smart healthcare, smart education, and human-computer interactions.", "key_idea": "The authors propose a model for inferring human predictive uncertainty, called proxy ensemble network (PEN), which exploits simple and universally accessible deterministic neural networks.", "method": "Simulations with a few benchmark datasets were performed. To show the applicability in real-world problems, behavioral experiments were conducted, in which 64 physicians classified medical images and reported their level of confidence.", "outcome": "The model was able to efficiently learn human uncertainty from a small amount of data. The PEN could predict both the uncertainty range and diagnoses given by subjects with high accuracy.", "future_impact": "This way of accessing human predictive uncertainty via deterministic neural networks has the potential to guide human decision making and can assist individuals in learning more efficiently and accurately.", "venue": "AAAI", "year": 2021, "title": "Human Uncertainty Inference Via Deterministic Ensemble Neural Networks"} +{"pid": "84d638d6-c8ec-4f64-97e6-b007b8bc6c0d", "context": "Factorization Machines (FMs) are widely used for collaborative filtering and recommendation tasks due to their ability to express and model pairwise feature interactions.", "key_idea": "This paper proposes Holographic Factorization Machines (HFM), a method that enhances the representation capability of FMs by replacing the inner product in FMs with holographic reduced representations, without increasing parameter size.", "method": "The authors conduct tests on nine publicly available datasets for collaborative filtering with explicit feedback to evaluate HFM's performance, and compare it against Attentional Factorization Machines and Neural Matrix Factorization.", "outcome": "HFM consistently improves over vanilla FMs by up to 4% in terms of mean squared error, and achieves state-of-the-art performance on all nine datasets.", "future_impact": "The authors propose a neural adaptation of HFM which could enhance its capability to handle nonlinear structures, suggesting a potential direction for future improvements.", "venue": "AAAI", "year": 2019, "title": "Holographic Factorization Machines for Recommendation"} +{"pid": "61b2c57b5244ab9dcb1d7356", "context": "Convolutional Neural Networks (CNNs) are efficient for classification tasks but are often viewed as a 'blackbox', making understanding predictions and debugging challenging. Current debugging and explanation methods diagnose potential causes for model prediction, but subsequent model optimization is done manually.", "key_idea": "A tool named Model Doctor, which automates model diagnosing and treating, is proposed. This tool is based on two findings that: (1) each category is connected with sparse and specific convolution kernels and (2) adversarial samples and normal samples behave differently in feature space.", "method": "A simple aggregate gradient constraint is developed to diagnose and optimize CNN classifiers. The efficacy of Model Doctor is assessed through experiments on existing CNN classifiers.", "outcome": "The proposed Model Doctor can be applied to all existing CNN classifiers, and it improves the accuracy of 16 mainstream CNN classifiers by 1%~5%.", "future_impact": "Model Doctor, as a versatile module, could potentially provide a standardized tool for diagnosing and optimizing different CNN classifiers, enhancing their accuracy and operational efficiency.", "venue": "AAAI", "year": 2022, "title": "Model Doctor: A Simple Gradient Aggregation Strategy for Diagnosing and Treating CNN Classifiers."} +{"pid": "bd2ebb29-56e2-4e7d-a25d-927280252497", "context": "Explanation-based learning (EBL) techniques have been widely researched and applied to problem solving. However, they are known to suffer from the utility problem, where the cost of utilizing the learned knowledge outweighs its benefits.", "key_idea": "The authors propose viewing the learning process as a sequence of transformations from a problem solving episode to a learned rule to analyze the increase in cost of a learned rule in an EBL system, specifically focusing on the Soar problem solving system which uses a chunking variant of EBL.", "method": "The paper analyses how the cost changes through the transformations by decomposing the chunking process into a sequence of transformations from problem solving to a chunk, then identifies sources which can make the output chunk expensive.", "outcome": "By analyzing these transformations of the Soar problem solving system, the authors identified a set of sources which can make the output chunk expensive.", "future_impact": "The resulting analysis provides a useful tool for revealing the sources of cost increase in the learning system and can guide future work in mitigating these sources, impacting how EBL systems are designed and evaluated.", "venue": "AAAI", "year": 1996, "title": "A transformational analysis of the EBL utility problem"} +{"pid": "b4229aa5-ae6a-463a-8ca7-0073ffaf922c", "context": "Current vision models classify objects based on features, but they haven't yet utilized taxonomical organization of class labels, making them less interpretable, especially compared to human cognition which often hierarchizes categories while identifying objects.", "key_idea": "The authors present a novel model that uses hierarchically organized prototypes to classify objects at multiple levels in a predefined taxonomy, providing distinct explanations for each level of predictions for an image.", "method": "The authors test their model on a subset of ImageNet for two tasks: 1) classification of data from familiar classes, and 2) classification of data from previously unseen classes at the appropriate level in the taxonomy.", "outcome": "The new model performed nearly as well as its black-box counterpart in both tasks while providing interpretable classifications.", "future_impact": "The hierarchical prototypes model could be used for future work in making machine vision systems more interpretable and in processing images from previously unseen classes at the level of the taxonomy to which they correctly relate.", "venue": "AAAI", "year": 2019, "title": "Interpretable Image Recognition with Hierarchical Prototypes"} +{"pid": "2a195328-5d57-4b8c-b32e-968a6a44aec8", "context": "Many computer science departments have seen a decline in enrollments. To attract students, especially non-majors, to computer science, subjects like Artificial Intelligence can be introduced early in the course curriculum.", "key_idea": "The authors propose two courses - an introductory course on AI and robotics for non-majors without prerequisites, and an elective course on machine learning taught in a tutorial format, with minimal prerequisites making it accessible to more students.", "method": "Courses are designed with ties to Cognitive Science program, enabling an extra means of advertising the courses to students who might not consider computer science. The structure and content of the courses are aimed at attracting non-majors to computer science.", "outcome": "The courses designed to attract non-computer science majors to AI topics resulted in students going on to enroll in more computer science courses and to consider becoming Computer Science majors.", "future_impact": "The proposed courses may set a precedent for the creation of future computer science coursework, potentially aiding in higher student enrollment by providing introductions to intriguing subjects like AI and robotics.", "venue": "AAAI", "year": 2008, "title": "Artificial Intelligence for Non-Majors at Multiple Levels."} +{"pid": "18824ea5-87ea-4e7a-a267-a6076b4631c3", "context": "In decision tree learning, determining a multiple-split test based on a tree-structured attribute that maximizes Quinlan's gain-ratio measure is a challenging task due to the exponentially growing number of possible tests.", "key_idea": "The authors propose an efficient algorithm that can find the optimal multiple-split test, without the need to enumerate and evaluate all possibilities.", "method": "The efficiency and effectiveness of the algorithm are assessed by its ability to solve the problem in a time proportional to dm for a training set of m examples and an attribute hierarchy of height d.", "outcome": "The proposed algorithm is able to run in time proportional to dm, demonstrating its computational efficiency.", "future_impact": "The algorithm, due to its computational efficiency, is practical for real-world use in decision tree learning, potentially contributing to more effective learning solutions.", "venue": "AAAI", "year": 1996, "title": "An efficient algorithm for finding optimal gain-ratio multiple-split tests on hierarchical attributes in decision tree learning"} +{"pid": "a9120c28-7462-43b3-ae35-d0689525e93b", "context": "Existing works focusing on image description mainly emphasize on generating expressive descriptions. The few that concentrate on generating stylish (e.g., romantic, lyric, etc.) descriptions suffer from limited style variation and content digression.", "key_idea": "The authors propose a controllable stylish image description generation model that can learn to generate stylish descriptions more related to the image content and can be trained with any monolingual corpus without collecting new paired image and stylish descriptions. They introduce the Domain Layer Norm (DLN) to enable users to generate various stylish descriptions.", "method": "The authors conducted extensive experimental validation and user study on various stylish image description generation tasks to verify the effectiveness of the proposed model.", "outcome": "The extensive experimental validations and user studies showcase the competitive advantages of the proposed model in stylish image description generation tasks.", "future_impact": "The proposed model, with its ability to generate various stylish descriptions by including new style-specific parameters, can open up new possibilities for producing image descriptions with various styles, enhancing the flexibility and overall applicability of image description generation models.", "venue": "AAAI", "year": 2019, "title": "Unsupervised Stylish Image Description Generation via Domain Layer Norm"} +{"pid": "6c546d82-e426-468c-886c-4bde3fd9215c", "context": "In the domain of construction robots, inspired by Harvard University's TERMES project, robots are required to gather blocks from a reservoir and build structures much larger than themselves. The problem of minimizing the number of pickup and drop-off operations performed on blocks in order to build user-specified structures presents a key challenge.", "key_idea": "The authors propose a tree-based algorithm that heuristically solves this problem using the idea of performing dynamic programming on a spanning tree in the inner loop and searching for a good tree in the outer loop.", "method": "The authors test their algorithm through simulations, considering large problem instances and scalability.", "outcome": "It is found that their algorithm performs well in simulations and easily scales to large problem instances.", "future_impact": "The authors believe their algorithm offers valuable lessons for the construction domain by comparing the success of their algorithm with the failures of off-the-shelf planning technologies.", "venue": "AAAI", "year": 2014, "title": "A Tree-Based Algorithm for Construction Robots"} +{"pid": "78c98506-3a8a-402f-bf5e-33388de88ab1", "context": "End-user interactive machine learning is a rising tool for enhancing human productivity with unstructured data. Existing work has demonstrated the creation of such systems for specific applications, but a generalized understanding of how to effectively design end-user interaction with interactive machine learning systems is lacking.", "key_idea": "The authors strive to understand how to design effective end-user interactions with interactive machine learning systems, with the focus on balancing the needs of end-users and machine learning algorithms.", "method": "The authors conduct three explorations in designing for effective end-user interaction with machine learning in CueFlik, a system developed to support Web image search.", "outcome": "The explorations demonstrate that interactions designed considering both the needs of end-users and machine learning algorithms can significantly improve the effectiveness of end-user interactive machine learning.", "future_impact": "This study may provide insights for future developments in end-user interactive machine learning systems and influence designs to better balance the needs of end-users and machine learning algorithms.", "venue": "AAAI", "year": 2011, "title": "Effective end-user interaction with machine learning"} +{"pid": "bd9408aa-465c-4a55-bbe0-6459aa6f44b1", "context": "Previous attempts at representing defeasible entailment and specificity in the framework of default logic may have shortcomings or room for improvement.", "key_idea": "The authors present a novel representation of defeasible entailment and specificity that bolsters the classical language with the language of conditionals possessing its own monotonic, internal logic.", "method": "The proposal's efficacy is showcased through the demonstration of how inheritance reasoning can be naturally depicted within this representative framework and then generalized to the entire classical language.", "outcome": "In this study, it was demonstrated that inheritance reasoning can be naturally integrated within the suggested framework and can be widely applicable to the entire classical language.", "future_impact": "This work has the potential to improve the understanding and application of defeasible entailment and specificity in default logic, possibly leading to more versatile and robust logical frameworks.", "venue": "AAAI", "year": 2007, "title": "On Default Representation of Defeasible Inference and Specificity."} +{"pid": "28158250-5f05-4f0a-8bda-83859c6baa91", "context": "While there is an abundance of theoretical work on Qualitative Spatial Reasoning (QSR), practical application requirements of QSR for real-world problems, such as detecting physically defective parts in manufacturing, are often ignored.", "key_idea": "This paper discusses the use of ontologies to compare different QSR formalisms with respect to defining spatial concepts, specifically focusing on defects like cracks and holes. It identifies a set of artifacts necessary for modeling these defects as a QSR problem.", "method": "The ontologies are introduced and used in a conceptual analysis of the problem of detecting physically defective parts, such as cracks and holes in manufacturing. The identified set of artifacts is compared with the capabilities of existing region-based QSR approaches.", "outcome": "It is found that most existing region-based QSR approaches fail to distinguish defects like cracks and holes effectively.", "future_impact": "The proposed set of problems can be used in future work to evaluate different QSR formalisms for their adequacy in defining and distinguishing cracks and holes.", "venue": "AAAI", "year": 2009, "title": "Detecting Physical Defects: A Practical 2D-Study of Cracks and Holes"} +{"pid": "6215a4242c356815940387ab", "context": "Activation maximization (AM) is a technique for generating input examples that maximize the activation of a target class in a classifier. GAN-based AM methods require the image set to contain some images of the target class, a limitation when dealing with unseen classes.", "key_idea": "The authors propose PROBEGAN, a GAN-based AM algorithm that can generate image classes unseen in the image set, effectively separating the target class information residing in the classifier from any target class information in the image set.", "method": "PROBEGAN is designed with a class-conditional generator, a seen-class discriminator, and an all-class unconditional discriminator. It includes AM in its training objective, and the authors conduct experiments to compare its ability to generate unseen-class images with other algorithms.", "outcome": "PROBEGAN was shown to generate images with features of unseen target classes, while retaining naturalness depicted in the image set. Also, it was found to produce higher quality unseen-class images than baseline methods.", "future_impact": "The authors suggest that PROBEGAN might be used as a model interpretation tool, increasing the accuracy and appropriateness of model representations for unseen classes.", "venue": "AAAI", "year": 2022, "title": "An Adversarial Framework for Generating Unseen Images by Activation Maximization."} +{"pid": "5ef56d78-6918-4abf-8a81-8aaa83a56976", "context": "Homeless youth are prone to HIV due to engaging in high-risk behavior, leading many agencies to conduct interventions. These interventions often involve selecting a group of youth to educate about HIV prevention, with the hope that they will spread this information through their social network. However, previous methods of strategically selecting participants do not accurately account for uncertainties in the social network\u2019s structure or its evolving state, leading to potential shortcomings in the spread of information.", "key_idea": "The authors developed PSINET, a decision support system that addresses uncertainties in network structure and evolving network state using Partially Observable Markov Decision Processes (POMDPs) in influence maximization, thus improving the spread of information about HIV prevention.", "method": "PSINET was tested through simulations, and its performance was measured by comparing the amount of information spread over current state-of-the-art influence maximization methods.", "outcome": "In simulations, PSINET showed approximately 30% more information spread over the current state-of-the-art influence maximization methods. PSINET is currently being reviewed by officials at My Friend\u2019s Place (a drop-in agency serving homeless youth in Los Angeles).", "future_impact": "Pending review and implementation, PSINET has the potential to enhance the efficiency of HIV prevention education among homeless youth by improving the spread of information through social networks, notably in real-world situations such as at My Friend's Place.", "venue": "AAAI", "year": 2015, "title": "Preventing HIV spread in homeless populations using PSINET"} +{"pid": "cecfaeeb-d43c-4a31-baa7-48e73eafbe80", "context": "Diagnosis of anomalies in ontology has been widely studied in the Semantic Web community. However, in ontology streams, diagnosis results are encapsulated within numerous, evolving ontologies rather than a single fixed ontology, leading to potential issues in identifying, managing, and understanding the evolution of diagnoses.", "key_idea": "The paper proposes a graph-based representation approach aimed at (i) organizing and linking time-evolving diagnoses effectively, and (ii) enabling scalable exploration of these diagnoses.", "method": "The authors implemented their graph-based representation approach to explore diagnoses in ontology streams and tested it with real and live data from Dublin City.", "outcome": "The authors found that their approach allowed for scalable diagnoses exploration in ontology streams, as demonstrated in the experiments using real world data from Dublin City.", "future_impact": "This approach has potential to improve understanding and identification of anomaly diagnoses in other evolving ontologies, particularly with large-scale, real-time data streams.", "venue": "AAAI", "year": 2014, "title": "Towards scalable exploration of diagnoses in an ontology stream"} +{"pid": "24eb2467-6ddf-46c1-9608-201e6bc1c8ff", "context": "The increasing problem of self-propagating code is causing intrusion attempts on the internet, with slowly propagating attacks difficult to detect because they hide within normal network traffic. Recently, anomaly-based intrusion detection systems have been able to identify quick spreading attacks, but slowly propagating attacks are still a challenge.", "key_idea": "The concept of collaborative intrusion detection systems is extended by incorporating probabilistic graphical models into end hosts and using random messaging to share state among peer detectors, which can boost the capabilities of weak anomaly detector D to detect significantly slower worms.", "method": "The authors validate their approach through theoretical analysis and evaluations, proving that their system can detect slower worms with low false positive rates by utilizing a weak anomaly detector either at the end-host or a network aggregation point.", "outcome": "A system imbued with probabilistic graphical models and gossiping among peer detectors is shown to be capable of detecting an order-of-magnitude slower worm than what could be achieved by using detector D alone. The system is also demonstrated to be scalable, distribute communication bandwidth uniformly, and leverage the enhanced computational power of distributed systems.", "future_impact": "The exploration of using probabilistic models for detection purposes indicates the potential for more robust detections compared to previous collaborative counting schemes and also allows for accounting for heterogeneous detectors in a principled manner.", "venue": "AAAI", "year": 2006, "title": "When gossip is good: distributed probabilistic inference for detection of slow network intrusions"} +{"pid": "6bb824c4-18a7-45c4-aaf3-25bde0ae83f2", "context": "Event sequences in multidimensional data pose challenges to traditional data mining techniques due to complexity in extracting interactions from patterns of damage in objects like fuel cells or earthquake occurrence patterns.", "key_idea": "The authors have proposed a novel data mining method named co-occurring cluster mining (CCM) that analyses sequences of multidimensional event data, generates cluster candidates, and tests them based on clustering in the data space and co-occurrence degree in the event sequence.", "method": "The researchers applied CCM to two practical domains: locating damage patterns in fuel cells, using sequence of acoustic emission events, and unveiling earthquake occurrence patterns, using a hypocenter catalog.", "outcome": "The co-occurring cluster mining method was able to infer mechanical interactions infuel cells and extract interactions between distant earthquakes in the given applications.", "future_impact": "The CCM method proposed may have applications in complex data structures where interactions between different elements need to be unearthed.", "venue": "AAAI", "year": 2014, "title": "Discovery of Damage Patterns in Fuel Cell and Earthquake Occurrence Patterns by Co-Occurring Cluster Mining"} +{"pid": "61f360c45aee126c0f7d5d47", "context": "Forest biomass is crucial for future climate, and scalable financing schemes like carbon offsetting certifications are needed to protect and restore forests. Existing manual methods to inventory forest carbon stock are time-consuming, labor-intensive, and subjective, often leading to overestimation and distrust in forest financing.", "key_idea": "In this study, the authors present ReforesTree, a benchmark dataset of forest carbon stock at six carbon offsetting sites in Ecuador, and propose using a deep learning-based end-to-end model for individual tree detection from low-cost RGB-only drone imagery to accurately estimate forest carbon stock.", "method": "The authors used deep learning-based end-to-end models for tree detection using low cost RGB-only drone imagery. The performance of the models was validated by comparison with the accuracy of official carbon offsetting certification standards and existing satellite-based forest biomass and carbon stock estimates.", "outcome": "The proposed model using ReforesTree dataset was found to accurately estimate forest carbon stock within the official carbon offsetting certification standards. The baseline CNN model surpassed the performance of state-of-the-art satellite-based forest biomass and carbon stock estimates for small-scale, tropical agro-forestry sites.", "future_impact": "The authors anticipate the study would encourage more machine learning research in this area, potentially enhancing the transparency and responsibility of monitoring, verification, and reporting in carbon offsetting projects, as well as global reforestation financing through accurate remote sensing.", "venue": "AAAI", "year": 2022, "title": "ReforesTree: A Dataset for Estimating Tropical Forest Carbon Stock with Deep Learning and Aerial Imagery."} +{"pid": "6020e2e79e795e62379b5b19", "context": "Existing adversarial attack methods that generate adversarial examples to confuse Deep Neural Networks (DNNs) perform well in image classification and object detection. However, these methods struggle to attack video object tracking models due to their ability to handle sequential information across video frames and the uncertainty of target categories.", "key_idea": "To solve the problem, a Unified and Effective Network (UEN) is proposed, that is competent to attack visual object tracking models.", "method": "UEN is designed with several features and tested using three different loss functions. It's evaluated with extensive experiments on state-of-the-art trackers, such as SiamRPN-based networks and DiMP, using popular tracking datasets including OTB100, UAV123, and GOT10K.", "outcome": "The UEN successfully attacks many state-of-the-art trackers, making real-time attacks possible. The attack results of UEN outperform the introduced baseline in terms of attacking ability and efficiency.", "future_impact": "The ability of UEN to launch effective real-time attacks on advanced tracking models could impact the way adversarial attacks are performed and the development of defenses against them in the future.", "venue": "AAAI", "year": 2021, "title": "A Unified Multi-Scenario Attacking Network For Visual Object Tracking"} +{"pid": "0ffdfb68-bee9-481b-a8a4-f5174fa20622", "context": "Growing demand is straining our existing electricity generation facilities and requires the active participation of the utility and the consumers to achieve energy sustainability through demand response (DR). This requires utilities to have reliable estimates of reduced consumption during DR for optimal planning, a problem that is under-studied.", "key_idea": "The authors introduce and formalize the problem of predicting reduced consumption during demand response.", "method": "The authors propose computational methods that use historical DR data as well as pre-DR conditions to make predictions of reduced consumption. The experiments are conducted in the real-world setting of a university campus microgrid.", "outcome": "The results from the experiments in a real-world setting set the foundation for more detailed modeling of reduced consumption during demand response.", "future_impact": "The problem of predicting reduced consumption during demand response and the methods proposed for tackling it could lead to more effective strategies for utility companies to manage peak demand.", "venue": "AAAI", "year": 2015, "title": "Estimating Reduced Consumption for Dynamic Demand Response"} +{"pid": "04855939-1eba-477d-bb18-13a446e95458", "context": "Reclassification risk, where health shocks persist for much longer than the usual one-year health insurance contracts, is a significant concern in the health insurance sector.", "key_idea": "The authors propose a solution of long-term insurance contracts with one-sided commitment, and aiming to determine the optimal contracts for consumers in Utah.", "method": "The authors use individual-level medical information from the Utah all-payer claims database to derive the optimal consumer contracts and assess the welfare level that full implementation of these contracts could achieve compared to key benchmarks.", "outcome": "Dynamic contracts perform extremely well for most of the population, eliminating over 94% of the welfare loss from reclassification risk for individuals who enter the market at age 25 in good health. However, dynamic contracts provide very little benefit to those with the worst health risks before the age of 25.", "future_impact": "The paper suggests that dynamic contracts can provide substantial welfare benefits under certain scenarios, but complementary policies are essential to fully unlock these benefits.", "venue": "AAAI", "year": 2019, "title": "Eliminating Latent Discrimination: Train Then Mask"} +{"pid": "61aed0d75244ab9dcb3a7050", "context": "While state-of-the-art brain-to-text systems have achieved great success decoding language directly from brain signals using neural networks, they have been limited to small, closed vocabularies, and require data from invasive devices like ECoG.", "key_idea": "The authors extend the problem to open vocabulary Electroencephalography (EEG)-To-Text Sequence-To-Sequence decoding and zero-shot sentence sentiment classification on natural reading tasks. They propose a novel framework leveraging pre-trained language models, hypothesizing that the human brain works as a special text encoder.", "method": "The proposed model is validated by performing EEG-To-Text decoding and zero-shot EEG-based ternary sentiment classification, and testing it on data from various subjects and sources.", "outcome": "The proposed model achieves a 40.1% BLEU-1 score on EEG-To-Text decoding and a 55.6% F1 score on zero-shot EEG-based ternary sentiment classification, significantly outperforming supervised baselines. The model demonstrated the ability to handle data from various subjects and sources.", "future_impact": "The proposed framework shows great potential for a high-performance open vocabulary brain-to-text system, once sufficient data is available.", "venue": "AAAI", "year": 2022, "title": "Open Vocabulary Electroencephalography-to-Text Decoding and Zero-Shot Sentiment Classification."} +{"pid": "c09edfd0-0ee4-4960-9ec4-5a231d3d4d9d", "context": "In the field of action recognition, existing deep learning methods treat sampled frames equally and all frame-level predictions are averaged at the test stage. However, within a video, discriminative actions may occur sparsely in a few key frames while many other frames are irrelevant or misleading, potentially resulting in incorrect predictions.", "key_idea": "The authors propose an attention-aware sampling method for action recognition, which selectively discards irrelevant and misleading frames and preserves the most discriminative frames for action recognition.", "method": "The process of mining key frames from videos is formulated as a Markov decision process, and the attention agent is trained through deep reinforcement learning without the need for extra labels. The attention agent uses features and predictions from a baseline model to generate importance scores for all frames and can be applied to different deep learning based action recognition models.", "outcome": "The proposed attention-aware sampling method is demonstrated to achieve competitive action recognition performance on two widely used action recognition datasets.", "future_impact": "The authors suggest that their approach is extensible and could potentially be applied to enhance the performance of other existing deep learning action recognition models, offering a new strategy for frame selection in action recognition.", "venue": "AAAI", "year": 2019, "title": "Attention-aware Sampling via Deep Reinforcement Learning for Action Recognition"} +{"pid": "6215a4242c35681594038651", "context": "AI systems are increasingly used to decide crucial matters such as loan applications. However, there are concerns they can discriminate against applicants based on sensitive attributes such as gender and ethnicity.", "key_idea": "The authors propose to evaluate the fairness of AI systems, specifically dealing with proxy and allowed variables. They describe fairness as a lack of discrimination based on sensitive attributes.", "method": "The authors formulate the decision-making process of a hypothetical regulator certifying the fairness of a bank's AI system, considering factors like the definition of fairness, dealing with proxy variables, and handling allowed variables.", "outcome": "The authors provide the theoretical result that the problem of deciding fairness as defined in the paper is co-NP-complete.", "future_impact": "The authors suggest that, despite the problem's complexity, it should be manageable in practical situations, implying potential future advancements in the method and applications for fairness testing of AI systems.", "venue": "AAAI", "year": 2022, "title": "On Testing for Discrimination Using Causal Models."} +{"pid": "21c4f1d9-d807-4c70-af1d-afa93b1506a4", "context": "The symmetric rendezvous search problem involves two robots, unaware of each other's locations, striving to meet as fast as possible. The symmetric specification requires the robots to follow the same strategy.", "key_idea": "The study presents a new symmetric rendezvous strategy for robots initially located on an open plane and then extends this to an unknown environment with polygonal obstacles.", "method": "The study involves the analysis of the competitive performance of the proposed symmetric rendezvous strategy, and quantifies its competitive complexity. The extended strategy is further analyzed for square, translating robots in terms of competitive ratio.", "outcome": "The researchers show that the competitive complexity of their strategy is O(d=R) where d is the initial robot distance and R is the communication radius. The competitive ratio of the extended strategy for square, translating robots is shown to be O(d=D), where D is the length of the robot sides.", "future_impact": "The findings of this study provide an upper limit on covering arbitrary polygonal environments, which could be valuable for independent research.", "venue": "AAAI", "year": 2012, "title": "Symmetric rendezvous in planar environments with and without obstacles"} +{"pid": "15b6c0d0-11b8-4600-96da-96b93b35e32c", "context": "Protein folding plays an essential role in protein function and stability, however, the understanding of it remains limited. The current tools and methods pose challenges in accurately identifying the folding core of proteins, essential to engineer proteins with desired functions and efficiencies.", "key_idea": "In this work, rigidity analysis is used to effectively sample and model the protein's energy landscape and to identify the folding core, suggesting a new way of understanding protein folding.", "method": "The method is validated by comparing models created using rigidity analysis of protein's energy landscape against known experimental data, and against results from other simulation tools. The proteins studied include protein G and its mutants, NuG1 and NuG2.", "outcome": "The results show that rigidity analysis improves the accuracy of landscape models, capturing the subtle folding differences between protein G and its mutants, and correlates well with experimental data. It could also suggest other components of structure that have not been previously identified experimentally.", "future_impact": "The methodology proposed in this work can offer a new perspective to improve our understanding of protein folding and potentially support the engineering of proteins with desired features and efficiencies.", "venue": "AAAI", "year": 2013, "title": "Rigidity analysis for protein motion and folding core identification"} +{"pid": "555ccabd-d6e0-4e46-b53d-1c0e1d021602", "context": "Existing knowledge of causal Bayesian networks and the distributions induced by local interventions on variables they govern lacks a complete characterization.", "key_idea": "The authors propose a complete characterization of the set of distributions resulting from local interventions on variables dictated by a causal Bayesian network, asserting that these distributions must adhere to three norms of coherence.", "method": "The authors demonstrate the usage of these coherence norms as inferential tools in tasks of learning and identification. They further derive testable coherence norms for networks containing unmeasured variables.", "outcome": "The authors successfully demonstrate the use of the three coherence norms as inferential tools in learning and identification tasks and also derive testable coherence norms for networks with unmeasured variables.", "future_impact": "The characterization of the distributions and the norms of coherence presented can potentially impact the learning and identification tasks in causal Bayesian networks, especially those containing unmeasured variables.", "venue": "AAAI", "year": 2002, "title": "A new characterization of the experimental implications of causal Bayesian networks"} +{"pid": "74a493b9-0341-46c8-8235-d275a93f832c", "context": "Reinforcement learning (RL) algorithms explore a domain's state space to learn optimal policies for control problems. However, the vast size of the state space in most problems inhibits RL technologies from fully exploring it to find good policies. State abstraction is one strategy used to reduce the size and complexity of a domain's state space.", "key_idea": "The authors propose a new approach for automatically deriving state abstractions: Evolutionary Tile Coding. This leverages a genetic algorithm for deriving effective tile codings.", "method": "The authors conducted an empirical analysis of the new algorithm, comparing it to another adaptive tile coding method and fixed tile coding. Testing was done on two RL benchmark problems.", "outcome": "Results indicated that the proposed approach can automatically derive effective state abstractions for two RL benchmark problems. It was revealed that the state space of the classical mountain car problem can be reduced to just two states, while still preserving the discovery of an optimal policy.", "future_impact": "This new approach for automatically deriving state abstractions may enable more efficient exploration and policy learning in reinforcement learning, notably in domains with large state spaces.", "venue": "AAAI", "year": 2010, "title": "Evolutionary tile coding: an automated state abstraction algorithm for reinforcement learning"} +{"pid": "d7db6b53-e48f-4a4a-9ab9-c9496edb1d98", "context": "The allocation of indivisible chores, items with negative utilities, using Max-Min Share (MmS) fair allocations has been challenging due to its intractability and the non-existence of MmS allocations in some cases.", "key_idea": "The authors propose a new fairness concept, optimal MmS, that represents the best possible allocation. They present a polynomial-time 2-approximation algorithm for MmS fairness for chore allocation.", "method": "Connections to parallel machine scheduling are utilized to provide a polynomial-time approximation scheme for computing an optimal MmS allocation when the number of agents is fixed. Additionally, an effective and efficient heuristic is developed with an ex-post worst-case analysis.", "outcome": "A polynomial-time 2-approximation algorithm for MmS fairness for chores is developed. Also, a polynomial-time approximation scheme for computing an optimal MmS allocation when the number of agents is fixed is generated, along with an effective and efficient heuristic.", "future_impact": "The suggested new concept of optimal MmS and its related algorithm could offer improved fairness in the allocation of indivisible chores and provide a path for future research in fair allocation problems.", "venue": "AAAI", "year": 2017, "title": "Algorithms for Max-Min Share Fair Allocation of Indivisible Chores"} +{"pid": "f395d0ba-5720-4cf5-806d-054b7416862e", "context": "The challenge of enhancing consumers' health literacy through virtual health environments, particularly for those affected by chronic low back pain, has not been effectively addressed. Existing approaches do not adequately consider the complexity of health literacy components.", "key_idea": "The authors propose a conceptual model of health literacy that considers its complexity and apply it in the design of a website named ONESELF, using a bottom-up approach that enables users to request the information they need.", "method": "The authors used the proposed conceptualization of health literacy to design and develop the ONESELF website that empowers users to request all necessary information.", "outcome": "The outcome of this study is a working website called ONESELF for consumers dealing with chronic low back pain. The website operationalizes vital dimensions of health literacy.", "future_impact": "The authors anticipate that the concept of health literacy and the ONESELF website would have future implications in the design of AI systems that are capable of delivering the same level of tailored information from a human-computer interaction perspective.", "venue": "AAAI", "year": 2009, "title": "Health Literacy and the Tailoring of Health Information. A Dialogue between Communication and (AI)Technology"} +{"pid": "621454435aee126c0f1d7065", "context": "Automated fact-checking is necessary to curb online misinformation. Current solutions verify claims by retrieving supporting or refuting evidence from related textual sources. However, these systems will need to verify claims against evidence sources that could be affected by the same misinformation.", "key_idea": "The authors explore the sensitivity of automated fact-checkers to synthetic adversarial evidence in two simulated settings of fabricating or altering documents in the evidence repository available to the fact-checking system.", "method": "They conduct a study across multiple models on three benchmarks (not detailed in the abstract) to examine the robustness of these systems against adversarial attacks.", "outcome": "The study demonstrates that automated fact-checkers suffer significant performance drops when faced with synthetic adversarial evidence.", "future_impact": "The authors raise important concerns about the increasing threat of modern Natural Language Generation systems serving as sources of disinformation and the challenges they pose to automated fact-checkers.", "venue": "AAAI", "year": 2022, "title": "Synthetic Disinformation Attacks on Automated Fact Verification Systems."} +{"pid": "860a5940-76d2-42a8-bfdb-8274fa50fc1b", "context": "The effectiveness of Serious Games (SG) hinges on their ability to blend gaming experiences and education effectively, which poses challenges in terms of intelligence and personalization. Fragmentation in research also complicates these challenges.", "key_idea": "The paper presents a summary of important contributions from Artificial Intelligence and Personalization to overcome research fragmentation and introduce key issues in Serious Games.", "method": "The paper amalgamates research insights from nine different institutions within a network of twenty-eight research labs and six game developers, aiming to identify hot topics for AI and Personalization within SG research.", "outcome": "The paper successfully summarizes approaches to user and learning goals modeling, user engagement, game adaptation levels, context identification through sensors and mobile technology, and content adaptation and reusability.", "future_impact": "This work promotes a focused research direction to establish benchmarks and metrics for the effective use of AI and Personalization in serious games design. It anticipates an impact area where research may have real benefits for SG products and the market as it matures.", "venue": "AAAI", "year": 2012, "title": "Artificial Intelligence and Personalization Opportunities for Serious Games"} +{"pid": "3efed9d0-faf0-44bf-8253-84413455a3af", "context": "Planning and execution systems have been used in a wide variety of systems, from scientific research to video games and military planning. However, the authors believe that most of these systems lack some key components that are present in other systems.", "key_idea": "The authors aim to enable future researchers to build more complete planning and execution systems by identifying the major technical problems that any implementer would have to face and citing recent solutions to each of these problems.", "method": "The authors review previous work on various planning and execution systems and identify the major technical problems any implementer would need to tackle.", "outcome": "The authors successfully identify the major technical problems faced in building a planning and execution system and propose recent solutions to these problems.", "future_impact": "The authors anticipate that by identifying the main technical challenges and solutions in planning and execution systems, they will assist future researchers in building more complete and efficient systems, potentially preventing serious system failure.", "venue": "AAAI", "year": 2006, "title": "Building robust planning and execution systems for virtual worlds"} +{"pid": "38fa4084-4612-4253-a10c-62af4c4d467e", "context": "Introducing robots as teammates in various domains raises human factors issues that do not necessarily arise in multi-robot coordination, especially when considering seamless integration with human group dynamics and the ability to robustly anticipate and adapt to other team members and the environment.", "key_idea": "This study proposes dynamic plan execution as a framework for explicitly modeling and efficiently reasoning on temporal information for human-robot interaction, and introduces an executive named Chaski that enables a robot to coordinate with a human to execute a shared plan under different models of teamwork.", "method": "The authors apply the Chaski executive to two Barrett Whole Arm Manipulators to demonstrate teamwork and describe ongoing work to demonstrate temporally fluid human-robot teaming using the Mobile-Dexterous-Social (MDS) robot.", "outcome": "The authors have successfully applied Chaski to two Barrett Whole Arm Manipulators and currently applying the methodology to the Mobile-Dexterous-Social robot.", "future_impact": "The advances in dynamic plan execution provide a potentially powerful framework for improving human-robot interaction, offering opportunities for further research. The application of these methods to robotics can open possibilities for seamless and efficient human-robot teamwork.", "venue": "AAAI", "year": 2010, "title": "Dynamic Execution of Temporal Plans for Temporally Fluid Human-Robot Teaming."} +{"pid": "3456e55d-e090-45bc-ab5e-54fec0b59cd2", "context": "In societies, agents operate under several, often contradictory, obligations and interdictions, which pose a challenge in achieving harmonious functioning.", "key_idea": "The authors propose an innovative solution in the form of a representation of obligations and interdictions within an organizational framework. This also includes an inference method that determines which obligations to break in situations that present contradictions.", "method": "The proposed solution is implemented within an operational, practically useful agent development language. The system is illustrated with a negotiated solution to a feature interaction problem in the telecommunications industry.", "outcome": "The system successfully supports negotiation by allowing agents to exchange deontic constraints, as evidenced by effectively solving a feature interaction problem in telecommunications.", "future_impact": "The approach could proactively manage the constraints imposed on agents in different societal systems, thus fostering better functionality and coordination in multi-agent applications.", "venue": "AAAI", "year": 1998, "title": "Agents that work in harmony by knowing and fulfilling their obligations"} +{"pid": "1a10605f-cc8e-4833-997b-f2f95f25e17b", "context": "Online map services like Google Maps and Open Street Maps have traditionally used pinch-to-zoom and drag-to-move gestures for map navigation on smartphones and tablets.", "key_idea": "This study introduces two new gestures for adjusting the zoom level of maps that allow users to identify a target area to zoom and enlarge it to cover the whole map container.", "method": "The proposed gestures, added to the traditional ones without overlap, were tested in comparison with the traditional pinch-to-zoom gesture in terms of time efficiency on tablets and smartphones.", "outcome": "Test results showed that the proposed gestures saved 30% of the time on tablets with the Two-Finger-Tap gesture and 14% on smartphones with the Tap&Tap gesture compared to the traditional pinch-to-zoom gesture.", "future_impact": "The key advantage of the proposed gestures, especially the Tap&Tap gesture for smartphones, is the ability to control the zoom level with just one hand, offering users more options for map navigation without altering their regular practices.", "venue": "WWW", "year": 2015, "title": "Two New Gestures to Zoom: Enhancing Online Maps Services"} +{"pid": "90347de9-c3c6-4f8b-8467-6e3604611ede", "context": "Millions of images on the web have been assigned latitude and longitude information but it remains challenging to effectively search and browse them due to the lack of methods to decide their geographical relevance. Collecting geographical relevance data has been attempted through game-based approaches, but the quality and utility of these data in improving applications have not been adequately demonstrated.", "key_idea": "This study focuses on the geographical relevance of images, defined as to what extent the main objects in an image match landmarks at the location where the image was taken. The authors propose a web-based, multiplayer game to collect geographical relevance data while providing an enjoyable experience for players.", "method": "The authors design and implement a multiplayer game to collect user-generated data about the geographical relevance of web images. They analyze the logs of game activity during a three week study with 147 participants and devise methods to determine image geographical relevance. A comparison experiment is conducted between the proposed methods and a commercial search engine.", "outcome": "The results of the study demonstrate that the proposed gaming approach and methods markedly improve image search relevance. It also enables the derivation of geographically relevant objects and their salient regions in images.", "future_impact": "The findings are valuable as they can be applied to other applications like image location recognition, potentially improving their performance.", "venue": "WWW", "year": 2009, "title": "A game based approach to assign geographical relevance to web images"} +{"pid": "621c3d205aee126c0fe7df48", "context": "Malicious accounts spreading misinformation in social media, particularly during the COVID-19 pandemic, has become a critical problem. Adapting fact-checking to new domains is slow and difficult to scale.", "key_idea": "The authors propose a framework that uses news-source credibility labels as weak labels for social media posts. The framework incorporates uncertainty sampling with self-supervision or relabeling to construct large-scale, diverse misinformation labeled datasets in new domains.", "method": "The proposed framework uses a detection model self-trained on the initial weak labels. Potentially inaccurate labels are identified and corrected using self-supervision or relabeling. Social context of the post in terms of the community of its associated user is considered for detecting inaccurate labels. Lastly, the framework uses a few labeled samples as class prototypes to separate high confidence samples into categories of information accuracy.", "outcome": "The approach has been demonstrated for providing a large-scale misinformation dataset on COVID-19 vaccines.", "future_impact": "This methodology could make it easier to build large-scale datasets with minimum human effort and streamline the process of fighting against misinformation in new domains, which is crucial for reliable dissemination of information.", "venue": "WWW", "year": 2022, "title": "Construction of Large-Scale Misinformation Labeled Datasets from Social Media Discourse using Label Refinement"} +{"pid": "092d7682-6d8f-46ce-be78-ecba7af6e409", "context": "Models of sentiment analysis in text typically require both a topic lexicon and a sentiment lexicon, with a need for mapping between these two to understand the specific emotional language used to describe certain topics.", "key_idea": "The paper introduces an 'affective stereotype lexicon', representing common sentiments and their most noticeable affective attributes associated with them. This qualitatively encodes how we express biases and cliches when talking about familiar topics.", "method": "The authors construct the affective stereotype lexicon and demonstrate its applicability via a demonstration system called MOODfinger. This system uses the lexicon for affective query expansion and summarization in a system for retrieving and organizing news content from the Web.", "outcome": "The affective stereotype lexicon developed is shown to be effective via MOODfinger. Though sentiments are bifurcated into simple positive and negative views, the methodology enables users to create their own nuanced moods on demand.", "future_impact": "While not explicitly mentioned, the successful adoption of the developed affective stereotype lexicon has the potential to enhance sentiment analysis systems, especially in terms of expressing nuanced moods on-demand.", "venue": "WWW", "year": 2012, "title": "In the mood for affective search with web stereotypes"} +{"pid": "c0105740-0761-474a-a0ba-2950ef28b449", "context": "XML-search requires a specific language to handle queries effectively, and W3C has proposed requirements for extending XQuery and XPath to manage full-text retrieval. However, problems exist in combining ranking with XML structured queries and meeting user needs.", "key_idea": "This study draws upon the authors' experience with INEX and evaluates how user needs can be translated into XML queries, influencing the requirements for an XML-search language.", "method": "The authors analyze the W3C requirements for extending XQuery and XPath with full-text search, comparing it with their experiences from INEX, with a particular focus on how to express user needs and how these needs can be translated into XML queries.", "outcome": "The analysis shows that the W3C requirements for extending XQuery and XPath with full-text search seem to address most of the needs identified, particularly in terms of implementing a scoring function for answer ranking. However, there are still issues regarding how structural requirements are combined with ranking.", "future_impact": "This paper's insights about user needs in XML queries and scoring functions could guide future enhancements in the design and implementation of XML-search language, particularly in combining structural requirements with ranking.", "venue": "WWW", "year": 2003, "title": "XML-search Query Language: Needs and Requirements"} +{"pid": "43b832f1-07e9-489b-ae49-f9c41e92752a", "context": "Relevant search terms for conceptual web search are traditionally generated by extracting frequently occurring key terms from retrieved documents. However, this method has limitations, including difficulties in term extraction and interference from irrelevant documents.", "key_idea": "We propose a semantic term suggestion function called Collective Intelligence based Term Suggestion (CITS), which provides a semantic graph of the search term without limiting to the specific query term and visualizes this graph for users.", "method": "CITS was compared to two popular commercial search engines through a user study involving 18 users and 50 search terms. In this user study, search results were ranked each time according to their relevance to the related terms in the entire query session.", "outcome": "The user study showed better user satisfaction with CITS compared to two popular commercial search engines, indicating the potential usefulness of the proposed method in real-world search applications.", "future_impact": "The success and potential usefulness of the CITS method in real-world search applications suggests potential for widespread adoption and further development.", "venue": "WWW", "year": 2009, "title": "Building term suggestion relational graphs from collective intelligence"} +{"pid": "650d0af6-bf0d-42f9-b128-f8f1545fb1d8", "context": "Protein-protein interaction and biomolecular pathway data are scattered across many disparate, large-scale, publicly-available databases and repositories. They are described using many different taxonomies and ontologies. Sophisticated integration, manipulation, processing and analysis are required to reveal unknown interactions and pathways that can lead to the discovery of new drugs.", "key_idea": "The BioMANTA project proposes the use of Semantic Web technologies together with a scale-out architecture to efficiently analyse, query, and reason about protein-protein interaction data.", "method": "The paper introduces the architecture, basic ontology and some of the implementation details of the BioMANTA system that aim to provide efficient, scalable RDF (Resource Description Framework) storage and inferencing.", "outcome": "This paper presents the initial results of the BioMANTA project and initial performance evaluations, although no specific results or performance metrics are given in the abstract.", "future_impact": "The completed BioMANTA system will enable knowledge representation and processing that are not currently available in typical scale-out or Semantic Web databases.", "venue": "WWW", "year": 2008, "title": "A scale-out RDF molecule store for distributed processing of biomedical data"} +{"pid": "1dca431f-a1fe-4fc9-90c9-eb1e964cbb71", "context": "Answer extraction from discussion boards has been extensively studied, with most existing work relying on supervised methods using similarity features and forum-specific features. However, these methods tend to struggle when applied to a different domain or forum, due to differences in vocabulary and feature availability.", "key_idea": "The authors propose a domain adaptive answer extractor which operates in two steps: an 'answer recognizer' that identifies sentences likely to be answers, and a 'domain relevance module' that determines the identified answer's domain significance. The methodology is domain-independent for easy adaptation to any domain.", "method": "The method involves a two-step process: firstly, an 'answer recognizer' identifies potential answers within a post, and secondly, a 'domain relevance' module assesses the domain relevance of the identified answers.", "outcome": "This paper presents the initial outcomes of this approach in the poster but does not detail specific results within the abstract.", "future_impact": "The study suggests that its domain-independent methodology could facilitate easy adaption to any given domain for answer extraction with minimal effort, thus potentially addressing a significant existing challenge in information retrieval from discussion forums.", "venue": "WWW", "year": 2012, "title": "Domain adaptive answer extraction for discussion boards"} +{"pid": "fd9c3294-98e4-4073-a459-6c929ec7e358", "context": "Understanding user behavior patterns of repeat consumption in domains like video re-watches, etc., is not sufficiently explored and recency of consumption is considered the strongest predictor of repeat consumption.", "key_idea": "The authors develop a model predicting that an item from 't' timesteps ago gets reconsumed with a probability proportional to a function of 't'. They also introduce item quality and develop a hybrid model combining both recency and quality to predict user choice.", "method": "They developed algorithms to learn reconsumption likelihood as a function of 't' and studied theoretical properties of this model. The authors also introduced a method for jointly estimating the parameters of the hybrid model.", "outcome": "The recency-based model shows a strong fit with the inferred function. The hybrid model combining recency and quality outperforms other alternatives in predicting user choice.", "future_impact": "This study lays a foundation for a better understanding and modeling of repeat consumption dynamics that could significantly enhance recommendation systems and personalization algorithms.", "venue": "WWW", "year": 2014, "title": "The dynamics of repeat consumption"} +{"pid": "36466cfa-efac-4b22-bacd-423e58e03a7d", "context": "Users of software applications generate vast amounts of unstructured log-trace data that contain clues to their intentions and interests. However, service providers often find it challenging to uncover and exploit these clues for personalization of software and web services.", "key_idea": "The authors propose a framework for personalizing software and web services that leverages unstructured log-trace data, using a user-modeling technique referred to as cutilization-to-vector, or util2vec.", "method": "The authors use 6 months of Photoshop usage history and 7 years of interaction records from 67K Behance users to develop and validate util2vec. They demonstrate its application in three scenarios: a user-tagging system, a two-phase recommendation model, and a real-time personalized inspiration engine.", "outcome": "The framework is shown to be effective in predicting areas of focus for millions of Photoshop users, improving recommendation quality by 21.2% over a popularity-based recommender, and providing real-time personalized inspirations to artists.", "future_impact": "The authors believe that their work demonstrates the potential impact of unstructured usage-log data for personalization, implying a future impact on the personalization of software and web services.", "venue": "WWW", "year": 2017, "title": "Personalizing Software and Web Services by Integrating Unstructured Application Usage Traces"} +{"pid": "3383bf98-78b5-4d1a-856f-41be639ccc75", "context": "The increasing volume of RDF data sets and their structure introduce concerns associated with the efficient storage and transmission of such data.", "key_idea": "This study explores the inherent compressibility of big RDF data sets, attributing this property to the structure of RDF graphs (power law), organization of URIs and RDF syntax verbosity.", "method": "The authors test basic approaches to compress RDF data using three real-world RDF data sets.", "outcome": "The paper's findings show that big RDF data sets are highly compressible.", "future_impact": "The results can inform the development of more efficient methods for storing and transmitting RDF data sets.", "venue": "WWW", "year": 2010, "title": "RDF compression: basic approaches"} +{"pid": "320c33f8-0ac5-4da8-8617-a15b9f00b0df", "context": "Currently, the Australian Government is collecting emergency incident data from web feeds for emergency management. The ERIC (Emergency Response Intelligence Capability) tool has been used for this process.", "key_idea": "This paper presents a comparative study of information obtained via the ERIC tool from web feeds and information available on Twitter utilizing the Emergency Situation Awareness (ESA) platform.", "method": "They compared the intelligence obtained using ERIC for a specific fire event with the information that was available on Twitter using the Emergency Situation Awareness (ESA) platform.", "outcome": "The findings show that Twitter data was reported faster, contained more specific event information, was updated more frequently, included information from the public, and remained available longer than the web feed information.", "future_impact": "This suggests that Twitter can be a highly beneficial supplementary source of information for emergency management, providing real-time details about emergency events including impact to the community.", "venue": "WWW", "year": 2013, "title": "Comparing web feeds and tweets for emergency management"} +{"pid": "b5b538c9-639b-42dc-a292-c3e0de843e46", "context": "Current search engines struggle to recognize both search intents and associated attributes within users' queries, requiring a mechanism for rich query interpretation.", "key_idea": "The authors propose mining structured query templates from search logs that are able to recognize both query intents and associated attributes; their framework is designed to discover templates with high precision and recall for matching queries in a particular domain.", "method": "They construct a tri-partite QueST network of queries, sites, and templates and use a probabilistic inferencing framework based on precision and recall, modeled as random walks in backward and forward directions. The system was tested on a real-world search log of 15 million queries.", "outcome": "The proposed query interpretation system achieved an accuracy of up to 90% (based on F-measure), with little seed knowledge and even with incomplete domain schema.", "future_impact": "The proposed method offers a new approach to improving search engine query interpretation, potentially enhancing the ability of such systems to more accurately and effectively meet users' information needs.", "venue": "WWW", "year": 2010, "title": "Towards rich query interpretation: walking back and forth for mining query templates"} +{"pid": "60641cba9e795e72406b65fe", "context": "Fact-checking has become the de facto solution for fighting fake news online. However, the cognitive biases of users can diminish the effectiveness of fact-checking.", "key_idea": "The authors conducted an experiment to explore the effect of cognitive biases on users\u2019 perception of fact-checking messages, especially focusing on the 'Lack of Evidence' label and how it might instigate uncertainty-aversion bias and disapproval bias.", "method": "The authors conducted an experiment involving 66,870 decisions, where they compared changes in users' stance towards unproven claims before and after presenting them with a hypothetical fact-checked situation. The experiment also involved user interviews about their likelihood to share certain types of claims.", "outcome": "Results showed that claims labeled as 'Lack of Evidence' were recognized similarly to false information, indicating an uncertainty-aversion bias. Users who initially disapproved of a claim were less likely to correct their views later than those who initially approved the claim when presented with fact-checking labels that contradicted their initial view, indicating a disapproval bias. Despite these biases, on average, fact-checking helped users correct their views and avoid extreme views.", "future_impact": "These observations have implications for multiple stakeholders, including platforms, policy makers, and online users, in order to design more effective fact-checking methods and messages.", "venue": "WWW", "year": 2021, "title": "An Experimental Study to Understand User Experience and Perception Bias Occurred by Fact-checking Messages"} +{"pid": "6274c64e5aee126c0f6fb379", "context": "In data-driven public health surveillance, it is a challenge to accurately identify posts where users are discussing health conditions due to the varied use of disease or symptom terms in user-generated content like social media. Existing computational research mainly focuses on health mentions in Twitter, with limited coverage of disease or symptom terms and little consideration of user behavior.", "key_idea": "The paper presents the Reddit Health Mention Dataset (RHMD) and a new method, HMCNET, which combines target keyword identification and user behavior hierarchically to improve health mention classification.", "method": "A new dataset called RHMD is manually created, which includes 10,015 Reddit posts mentioning 15 common disease or symptom terms. They use four labels to annotate the posts: personal health mentions, non-personal health mentions, figurative health mentions, and hyperbolic health mentions. HMCNET is then used for the health mention classification task on this dataset.", "outcome": "The experimental results show that the proposed HMCNET method outperforms state-of-the-art methods with an F1-Score of 0.75, an improvement of 11% over the existing best results.", "future_impact": "The new dataset and improved method, both presented in the paper, pose a strong challenge for existing health mention classification methods, potentially leading to advancements in the field.", "venue": "WWW", "year": 2022, "title": "Identification of Disease or Symptom terms in Reddit to Improve Health Mention Classification"} +{"pid": "238f559b-bc20-4184-a7ce-51a184e0e423", "context": "The growth of social networks in terms of users, resources, and interactions can often lead to users being lost or unable to find useful information. Social annotations (tags) are becoming increasingly popular and can help to avoid such user disorientation.", "key_idea": "In this study, the authors suggest utilizing social annotations in building a user profile, focusing especially on the approach of modeling and updating a tag-based user profile.", "method": "The authors provide a state-of-the-art review of social user characteristics and techniques for modeling and updating a tag-based profile, as well as the treatment of social annotations.", "outcome": "The authors demonstrate how social annotations can be treated and how tag-based profiles can be modeled for recommendation purposes.", "future_impact": "The techniques and methods discussed in the survey can be potentially useful for improving recommendation systems and user interactions on social networks.", "venue": "WWW", "year": 2012, "title": "A user profile modelling using social annotations: a survey"} +{"pid": "5e6614aa91e0116831824899", "context": "All-pairs set similarity, particularly for discovering pairs with small intersection sizes, is a critical task in data mining. Particularly in recommender systems, certain dimensions can be highly skewed due to popularity, making traditional approaches for finding similarities insufficient for large inputs.", "key_idea": "To overcome the limitations of traditional similarity search methods, the authors propose LSF-Join, a new distributed algorithm for approximate all-pairs set similarity that uses a randomized selection procedure based on Locality Sensitive Filtering, instead of previous methods based on Locality Sensitive Hashing.", "method": "The authors demonstrate the accuracy of the LSF-Join algorithm through experimental testing on multiple graphs, and provide theoretical results showing that their method efficiently finds most close pairs even for small similarity thresholds and skewed input sets.", "outcome": "The theoretical results showed that LSF-Join could efficiently find most close pairs for small similarity thresholds and skewed input sets. Experimental results demonstrated the accuracy of LSF-Join on multiple graphs.", "future_impact": "The authors did not anticipate any specific future impact of their study.", "venue": "WWW", "year": 2020, "title": "LSF-Join: Locality Sensitive Filtering for Distributed All-Pairs Set Similarity Under Skew"} +{"pid": "40f98ea7-d625-488f-823f-dd73bcf6b3f3", "context": "Drugs can have unintended side effects resulting from off-target interactions, which can adversely affect patients. Growing Web 2.0 usage has led to increasing discussion of drug side effects in online sources.", "key_idea": "The authors propose UDWarning, a novel early warning system for unrecognized drug side effects, which uses text information from the Internet.", "method": "UDWarning automatically builds a knowledge base for drug side effects by integrating information related to drug side effects from different online sources and also monitors the online information about drugs.", "outcome": "The authors demonstrate that UDWarning can discover possible unrecognized drug side effects and potentially expedite the discovery process.", "future_impact": "This system has the potential to improve the quality of healthcare by providing early warnings of unrecognized drug side effects.", "venue": "WWW", "year": 2012, "title": "An early warning system for unrecognized drug side effects discovery"} +{"pid": "8f0a6893-6741-45f2-bd0c-66abee703b39", "context": "Current methods for term definition search either generate a single combined definition or output all retrieved definitions, without adequately ranking them based on their quality.", "key_idea": "The authors tackle the problem of definition ranking. They propose to formalize the problem as either classification or ordinal regression, using SVM for classification and Ranking SVM for ordinal regression, in order to rank definition candidates based on their likelihood of being good definitions.", "method": "The authors define features for constructing the SVM and Ranking SVM models and conduct experiments on paragraphs and sentences across different domains, comparing the proposed method to heuristic rules and the conventional information retrieval method of Okapi.", "outcome": "Experiments show that SVM and Ranking SVM outperform baseline methods both when the answers are paragraphs and sentences. Also, models trained in one domain can be adapted to another.", "future_impact": "The study suggests the possibility of constructing generic models for definition ranking that can be adapted across different domains.", "venue": "WWW", "year": 2005, "title": "Ranking definitions with supervised learning methods"} +{"pid": "05d51dea-522a-4f64-8358-95fb0eb2c5dc", "context": "Web content credibility systems are critically designed, but subjectivity and bias in credibility ratings can harm the system's performance.", "key_idea": "The authors initiated the Reconcile1 study to explore the influence of users' socio-economic status and psychological traits on the subjectivity and bias in the evaluation of web content credibility.", "method": "The study involved 1503 participants giving independent evaluations on the credibility of 154 different web pages across various themes. The participants also completed a questionnaire covering additional traits.", "outcome": "The study findings showed that evaluations of Web content credibility are slightly subjective and exhibit a strong acquiescence bias.", "future_impact": "These findings shed light on the influence of user traits on content credibility evaluations, providing valuable insights for improving the design of future Web Credibility systems in order to mitigate potential subjectivity and bias.", "venue": "WWW", "year": 2013, "title": "On the subjectivity and bias of web content credibility evaluations"} +{"pid": "db4d49a9-392f-4d73-ae4d-a0e3d2c6d681", "context": "Web communities are often classified based on business-oriented factors, while textual communication in message boards, which provides valuable insight into the nature of these communities, has been overlooked.", "key_idea": "The authors propose a new approach for classifying web communities based on the communication-content that uses KeyGraph. This method visualizes co-occurrence relations between words and word clusters in text on the message boards.", "method": "Text on a message board is analyzed with KeyGraph to determine the structure of the content-flow. Three structure-features of relations between participants and words are computed and visualized: centralization, context coherence, and creative decisions.", "outcome": "The authors demonstrate that the structure obtained using KeyGraph can capture the essence of the content-flow of message boards, reflecting the interests of the participants in the community.", "future_impact": "The proposed method could be used to understand better the essence of a community, such as whether the community creates useful knowledge, how easy it is to join the community, and whether and why the community is good for making commercial advertisements.", "venue": "WWW", "year": 2002, "title": "Featuring web communities based on word co-occurrence structure of communications: 736"} +{"pid": "6db39816-4ea0-4711-bd8d-c46b5b29a151", "context": "While there are many algorithms proposed to obtain hierarchical structures, called folksonomies, from social tagging data, there is a lack of understanding on the extent folksonomies are pragmatically useful for navigating social tagging systems.", "key_idea": "The paper presents a pragmatic framework for evaluating folksonomies. The authors model exploratory navigation of a tagging system as decentralized search on a network of tags, and use the hierarchical structures learned by folksonomy algorithm as background knowledge for the decentralized search.", "method": "The authors conduct experiments using four state-of-the-art folksonomy algorithms on five different social tagging datasets. They use decentralized search on tag networks in combination with different folksonomies as hierarchical background knowledge to evaluate navigational tasks in social tagging systems.", "outcome": "The experiments reveal significant, previously undiscovered differences among the existing folksonomy algorithms in terms of their utility for navigation.", "future_impact": "The results of this paper can guide engineers in their efforts to improve the navigability of social tagging systems and help scientists in evaluating different folksonomy algorithms from a pragmatic perspective.", "venue": "WWW", "year": 2011, "title": "Pragmatic evaluation of folksonomies"} +{"pid": "f41ae5c7-0c15-420b-9559-5f34d86a89ba", "context": "A few years after the initiation of the semantic Web by the W3C, there has been no quantitative measurement of its status, which is crucial information for researchers, developers and administrators in the field.", "key_idea": "The authors aim to quantitatively measure and present the current status of the semantic Web by conducting a longitudinal study on the use of semantic markup languages over time.", "method": "The authors conducted a longitudinal study on semantic Web pages to track trends in the use of semantic markup languages. They presented early results of this study using two historical data sets from October 2003 and October 2004.", "outcome": "The study reveals that early adoption of the semantic Web is ongoing and its growth outpaces that of the entire Web within the study period. RDF dominates among semantic markup languages, accounting for about 98% of all semantic pages on the Web. The most popular application is RSS, which takes up over 60% of all semantic Web pages. The use of OWL increased by 900% in the period of study.", "future_impact": "The study provides insights into the future direction of semantic Web adoption, including the increasing usage of OWL and the dominance of RDF, which will be beneficial for researchers, developers and administrators.", "venue": "WWW", "year": 2005, "title": "The semantic webscape: a view of the semantic web"} +{"pid": "60226f7a91e011e70e451340", "context": "Graph denoising, which recovers graph structures by removing irrelevant edges and adding missing ones, has significant real-world applications (e.g., improving connections in co-authorship networks), but existing methods often require supervision from clean graphs.", "key_idea": "The paper introduces Mask-GVAE, a variational generative model, for the blind denoising of large discrete graphs without requiring any supervision or clean graphs.", "method": "Mask-GVAE uses robustness in low eigenvectors of graph Laplacian against random noise to break down the input graph into several stable clusters. Then, it handles the significant computations by decoding probabilistic smoothed subgraphs in a variational way.", "outcome": "On a variety of benchmarks, Mask-GVAE significantly outperformed other approaches in terms of PSNR (Peak Signal-to-Noise Ratio) and WL (Weisfeiler-Lehman) similarity.", "future_impact": "The technique could have substantial real-world applications such as improving the quality of connections in co-authorship networks.", "venue": "WWW", "year": 2021, "title": "Mask-GVAE: Blind Denoising Graphs via Partition"} +{"pid": "673b8184-84f4-40cb-84ef-0d488bfa2c98", "context": "Existing relationship analysis tools for networks assume homogeneity in both node type and edge type. However, recent research has started to focus on mining heterogeneous information networks where both nodes and edges can be of different types.", "key_idea": "This study introduces a novel approach for mining relationships across entities in heterogeneous networks while considering user preference over relationship type and interestingness metric. This is achieved by formalizing the relationship mining task as a top-k lightest paths problem.", "method": "The authors propose the PROphetic HEuristic Algorithm for Path Searching (PRO-HEAPS) that leverages a combination of novel graph preprocessing techniques, well-designed heuristics, and the A* search algorithm. The algorithm is evaluated on real-world large-scale graphs against a variety baseline approaches.", "outcome": "The proposed PRO-HEAPS algorithm significantly outperforms a wide variety of baseline approaches with speedups as large as 100X on real-world large-scale graphs.", "future_impact": "The authors suggest potential useful applications of their algorithm, indicating the relevance of their work for future innovations in relationship mining in heterogeneous networks.", "venue": "WWW", "year": 2016, "title": "What Links Alice and Bob?: Matching and Ranking Semantic Patterns in Heterogeneous Networks"} +{"pid": "60641d5a9e795e72406b6648", "context": "In today's fast-paced and competitive environment, the workforce needs to be agile and capable of adapting to learning new job skills. However, it is challenging for individuals to know which skills to develop at each stage of their career.", "key_idea": "The authors aim to develop a cost-effective recommendation system based on deep reinforcement learning that can provide personalized and interpretable job skill recommendations for each individual.", "method": "The authors design an environment to estimate the utilities of skill learning using a comprehensive approach based on massive job advertisement data. They develop a Skill Recommendation Deep Q-Network (SRDQN) with a multi-task structure to estimate the long-term skill learning utilities.", "outcome": "Results from extensive experiments on real-world datasets confirm the effectiveness and interpretability of the authors' deep reinforcement learning-based skill recommendation approach.", "future_impact": "The development of this cost-effective recommendation system could potentially aid individuals in identifying and acquiring necessary skills for career development in a personalized manner.", "venue": "WWW", "year": 2021, "title": "Cost-Effective and Interpretable Job Skill Recommendation with Deep Reinforcement Learning"} +{"pid": "1c4b8205-1f48-4497-919c-e5eb1407171f", "context": "In online advertising, ad campaign assessment often involves calculating the ratio of success rates between users exposed and not exposed to ads, representing ad effectiveness. However, many campaigns target users with high predicted success rates, potentially overestimating ad effectiveness as these users may perform the desired actions even without the ad.", "key_idea": "The authors introduce the concept of 'smart cheating', where ad campaigns may appear more effective due to targeting users who would perform the action with or without ad exposure. They propose a method to quantify this phenomenon using a smart cheating ratio (SCR) through causal inference.", "method": "The authors apply their proposed smart cheating ratio to multiple real ad campaigns to evaluate its efficacy and prevalence.", "outcome": "The application to real ad campaigns reveals that 'smart cheating' is extensive and can be severe in the current advertising industry.", "future_impact": "Correcting for the overestimation due to 'smart cheating' when assessing ad campaigns can help optimize budget allocation and maximize incremental success actions.", "venue": "WWW", "year": 2015, "title": "Rethink Targeting: Detect 'Smart Cheating' in Online Advertising through Causal Inference"} +{"pid": "ef3c7474-65d1-4239-8f96-737e5f820f6a", "context": "In the current Web scenario, many similar Web services are offered through heterogeneous interfaces and business protocols, and interoperation is typically enabled through developing adapters. There have been approaches to classify mismatches between service interfaces and business protocols to facilitate adapter development, but identifying the actual mismatches between two service specifications remains a challenging task.", "key_idea": "This paper presents techniques and a tool for semi-automated identification and resolution of mismatches between service interfaces and protocols, and for generating adapter specifications. It focuses on finding mismatches of type of signature, merge/split, and extra/missing messages, and also identifies all ordering mismatches between service protocols by generating a mismatch tree.", "method": "The authors implement their approach in a tool within IBM WebSphere Integration Developer and conduct experiments with real-world case studies to validate the feasibility and effectiveness of their proposed methodology.", "outcome": "The results from the experiments with real-world case studies demonstrate the viability of the proposed approach, indicating that it significantly simplifies the problem of adapting services to enable interoperation.", "future_impact": "Though not explicit in the abstract, the authors\u2019 methodology could potentially streamline the process of developing web service adapters, thereby enhancing interoperation in heterogeneous digital environments.", "venue": "WWW", "year": 2007, "title": "Semi-automated adaptation of service interactions"} +{"pid": "d35a1a97-317b-4b47-b1e1-eb9b32886331", "context": "Factorization Machine (FM) delivers state-of-the-art performance in batch learning tasks where all the training data is available before training, but due to re-training costs, it is impractical for applications where data arrives sequentially in a streaming manner.", "key_idea": "The authors address the limitation of FMs by developing an online learning algorithm. They propose a new convexification scheme leading to the Compact Convexified FM (CCFM), and then devise an Online Compact Convex Factorization Machine (OCCFM) algorithm that avoids expensive projection operations.", "method": "The use of the OCCFM algorithm was evaluated using 6 real-world datasets for recommendation and binary classification tasks.", "outcome": "Experimental results demonstrate that the Online Compact Convex Factorization Machine outperforms the state-of-art online learning algorithms.", "future_impact": "The proposed online learning algorithm with its sub-linear regret bound can offer theoretical reassurance of its usefulness, indicating potential for broader adoption in real-world applications where data arrives in a streaming manner.", "venue": "WWW", "year": 2018, "title": "Online Compact Convexified Factorization Machine"} +{"pid": "64635097d68f896efabaa6a4", "context": "Medical professionals and caregivers often interact with different modalities of electronic health records (EHRs), a process that is complex and could benefit from technological simplification.", "key_idea": "The authors introduce MediSage, an AI decision support assistant that utilizes a knowledge graph integrating general clinical resources and recent EHR data to simplify the interaction with EHRs and provide step-by-step reasoning support in healthcare scenarios.", "method": "The authors designed MediSage to combine the structured representation of knowledge with the predictive power of neural models trained over both EHR and knowledge graph data.", "outcome": "MediSage is capable of summarizing patient health, predicting patient outcomes, and providing comprehensive and personalized healthcare recommendations to medical professionals and caregivers.", "future_impact": "MediSage represents a stepping stone for further integration with biomedical language models and may enhance the provision of personalized healthcare through AI-driven decision support.", "venue": "WWW", "year": 2023, "title": "MediSage: An AI Assistant for Healthcare via Composition of Neural-Symbolic Reasoning Operators"} +{"pid": "cb52a910-f29d-4f65-acea-774bb09f1386", "context": "In Smart Cities research, integrating complex systems and diverse technologies in multifaceted structures has been a prominent recurring problem yet without any wide-reaching solution.", "key_idea": "This study suggests a solution for addressing the integration problem in Smart Cities involving a formalized accepted and managed technology regulated environment introduced by governance groups that will establish a standard for automated and autonomous components, integrating dynamically with software agents.", "method": "The paper describes the requirements for a new framework solution, identifies governance that can manage integration in smart cities, and explores how the proposed technology regulated environment would work, reportedly without incurring cost beyond processing time.", "outcome": "The study outlines a proposed framework and describes how the associated governance would operate but does not report direct outcomes or results of applying this solution in a real-world context.", "future_impact": "The proposed approach, if accepted and implemented, would give developers and vendors an obligation to meet safety standards and accept liability for malfeasance, addressing crucial issues of safety and responsibility in Smart Cities technology integration.", "venue": "WWW", "year": 2017, "title": "Establishing Smart City Technical Standards and Guidance: A Way Forward"} +{"pid": "281dece4-50f9-48d5-a730-587d95e9888c", "context": "Traditional techniques for querying graph data use structural and node similarities to find matches of a given query graph in a larger target graph. However, these techniques have largely ignored interactions between node and edge labels, which encode relationships in the graph.", "key_idea": "This paper proposes RAQ (Relationship-Aware Graph Querying), which takes into account the relationships in graphs. Given a query graph, RAQ identifies the k best matching subgraphs of the target graph that encode similar relationships as in the query graph.", "method": "The authors perform a user survey on the Internet Movie Database (IMDb) to assess the effectiveness and preference of RAQ over traditional graph querying. They also achieve practical querying through beam stack search, and perform experimental evaluations on multiple real-world graph datasets.", "outcome": "The user survey results show that 86% of the 170 surveyed users preferred relationship-aware match over traditional graph querying. The experiments on multiple real-world graph datasets demonstrate RAQ to be effective, efficient, and scalable.", "future_impact": "The proposal of RAQ suggests a potential shift in graph querying paradigms, prioritizing the understanding and utilization of relationships in graphs, which could impact knowledge discovery and exploration tasks.", "venue": "WWW", "year": 2019, "title": "RAQ: Relationship-Aware Graph Querying in Large Networks"} +{"pid": "c6e521e1-529c-45d3-bbbe-aa7789fca634", "context": "The enormous success of the World-Wide Web has created many problems within the Internet, most of them due to its huge bandwidth requirements. Additional applications using the Web's infrastructure for distributed computing further increase network load.", "key_idea": "The authors propose a comprehensive application-level approach to solving the bandwidth and latency problems induced by the Web. This approach is based on automated replication schemes and complementary name services.", "method": "The replication schemes and naming services are introduced and explained in this paper.", "outcome": "The authors argue that their approach brings many advantages not only to every Web user but also to various network and service providers currently struggling to keep up with rising user demands.", "future_impact": "The authors highlight the concepts of their freely available implementation as a potential solution to the issue of bandwidth and latency in the current web infrastructure.", "venue": "WWW", "year": 1996, "title": "Introducing application-level replication and naming into today's Web"} +{"pid": "f444857f-b324-4ac1-a8ce-afc388100dc7", "context": "Traditional server-based web applications allow access to server-hosted resources, but often exhibit poor responsiveness due to server load and network delays. Client-side web applications provide excellent interactivity at the expense of limited access to server resources.", "key_idea": "The WebWriter II Editor is a direct manipulation HTML editor that uses both server-side and client-side processing in order to achieve the advantages of both, creating a responsive interaction for a large Web application.", "method": "This editor downloads the document data structure to the browser and performs all operations locally, by using HTML frames and JavaScript code residing in roughly twenty HTML pages that are downloaded into these frames as needed.", "outcome": "WebWriter II Editor is developed using the Meteor Shower Application Architecture, which has efficiently used server-side and client-side processing.", "future_impact": "The authors believe that such a client-server architecture, based on frames, client-side data structures and multiple JavaScript-enhanced HTML pages, could be promising for a wide variety of applications.", "venue": "WWW", "year": 1997, "title": "Responsive interaction for a large Web application: the meteor shower architecture in the WebWriter II Editor"} +{"pid": "b0b06cae-1879-493a-b92e-9f6a7bfc8ec9", "context": "Although latent factor model (e.g., matrix factorization) achieves good accuracy in rating prediction, it suffers from problems including cold-start, non-transparency, and suboptimal recommendations for local users or items. Traditional methods don't use textual review information along with ratings, which results in limitations.", "key_idea": "The authors propose to exploit textual review information with ratings to handle these limitations. They introduce an aspect-aware topic model (ATM) to model user preferences and item features from different aspects and estimate the aspect importance of a user towards an item. This aspect importance is then incorporated into a novel aspect-aware latent factor model (ALFM) which relates user and item latent factors to aspect ratings.", "method": "The implementation and effectiveness of the proposed ATM and ALFM models are evaluated through comprehensive experimental studies conducted on 19 datasets from Amazon and Yelp 2017 Challenge dataset.", "outcome": "The proposed model is reported to significantly outperform existing methods, especially for users with only a few ratings. In addition, the model can interpret recommendation results in depth.", "future_impact": "The use of aspect-aware rating prediction models like ALFM could result in more accurate and interpretable recommendations, and could potentially alleviate the data sparsity problem in existing systems.", "venue": "WWW", "year": 2018, "title": "Aspect-Aware Latent Factor Model: Rating Prediction with Ratings and Reviews"} +{"pid": "3e950d83-35de-4612-a78f-da59af9704fe", "context": "The need for effective analysis of social media messages for disaster management, particularly in the Italian context, forms the backdrop for this study. Previous studies have not focused on out-domain and cross-event damage detection using these messages.", "key_idea": "This work proposes a novel approach to tweet-derived feature investigation for out-domain and cross-event damage detection, using a combination of lexical and grammatical structures and newly implemented task-specific features.", "method": "The authors carried out various experiments using a range of linguistic features describing the lexical and grammatical structure of a text, as well as ad-hoc features constructed specifically for this analysis.", "outcome": "The experiments led to the identification of the most effective features that yield the best results. Additionally, the study resulted in the development of the first manually annotated Italian corpus of social media messages for damage assessment.", "future_impact": "This work's finding could enhance the process of damage assessment during natural disasters by improving the analysis of social media messages. The creation of the Italian corpus could also support more focused research in this area in the future.", "venue": "WWW", "year": 2015, "title": "A Linguistically-driven Approach to Cross-Event Damage Assessment of Natural Disasters from Social Media Messages"} +{"pid": "2914e591-d662-40a6-8292-04e2e5cc37fd", "context": "Government initiatives to open data to the public have become popular, leading to both social and technical opportunities and challenges.", "key_idea": "The authors propose a new social machine-oriented architecture, Gov-SM, that utilizes open data and combines principles from existing architectural patterns to promote user, developer, and crowd participation in governmental issues and policy.", "method": "The authors develop a platform of specialized APIs based on Gov-SM to enable the creation of other social-technical systems.", "outcome": "The paper does not provide specific implementation results or measurable outcomes in the abstract.", "future_impact": "The proposed Gov-SM architecture has the potential to promote participation in governmental issues and policy and enable the creation of additional social-technical systems.", "venue": "WWW", "year": 2015, "title": "Towards Government as a Social Machine"} +{"pid": "28553181-958f-4248-af2f-ce44c3a125ac", "context": "Web images are typically classified based on their visual features, but the existing methods may not consider the functionality of these images in the context of web pages.", "key_idea": "A new function-oriented classification of web images is proposed, which leverages image-based features, text-based features and DCT coefficients in a Support Vector Machine (SVM) classification in a tree-structured way.", "method": "To attains high accuracy in classification, two features, image-based and text-based are used together or separately at different stages of the classification process. DCT coefficients are also utilized to differentiate photo images and illustrations.", "outcome": "As a result of applying the proposed classification method, a high degree of accuracy in classifying web images was achieved.", "future_impact": "The new categories of WWW images enabled by the proposed method could make page summarization feasible for the first time, opening up potential applications in this area.", "venue": "WWW", "year": 2005, "title": "Accuracy enhancement of function-oriented web image classification"} +{"pid": "93d12d0d-8709-451d-920e-44b5e5e0a1d8", "context": "Developers of connected software, including websites, applications, and devices, can now evaluate ideas quickly and accurately using controlled experiments, also known as A/B tests. However, the deployment and evaluation of online controlled experiments at scale across a variety of platforms presents many pitfalls and new research challenges.", "key_idea": "This paper focuses on highlighting the overall A/B testing methodology, detailing its use cases through real examples and discussing the practical and research challenges in scaling experimentation.", "method": "The paper introduces and discusses the A/B testing methodology, using real examples to walkthrough its use cases, and focuses on addressing challenges in scaling experimentation.", "outcome": "The tutorial shares key lessons learned from scaling experimentation at Microsoft to thousands of experiments per year.", "future_impact": "The insights and lessons from this paper may guide future efforts to address challenges and enhance the scaling of software experimentation, leading to more efficient and reliable A/B testing processes.", "venue": "WWW", "year": 2019, "title": "A/B Testing at Scale: Accelerating Software Innovation"} +{"pid": "41296c41-32e4-49c2-8f2c-dff3f7b66644", "context": "Although dozens of query-by-sketching/ query-by-example/ query-algebra systems have been proposed for searching large time series collections, they have not seen widespread adoption due to their complexity and limited expressiveness.", "key_idea": "In this work, the authors propose a natural language search mechanism for searching time series aiming to make it more intuitive, expressive, and efficient.", "method": "The authors demonstrate the utility of the proposed system through several case studies in different domains such as animal motion studies, medicine, and industry.", "outcome": "The proposed system is found to be expressive, intuitive and requires little space and time overhead. Moreover, it is shown to subsume both motif/discord discovery and most existing query-by-content systems in the literature.", "future_impact": "As the proposed system is text-based, it can leverage decades of research in text retrieval, including ideas such as relevance feedback, offering potential for improved performance and wider adoption in the future.", "venue": "WWW", "year": 2019, "title": "Putting the Human in the Time Series Analytics Loop"} +{"pid": "60c093fc91e0116059b15a2a", "context": "Rockburst disasters pose serious threats to the safety and progress of construction work. The existing prediction methods do not ensure accurate and rational predictions of rockburst tendencies, index weights, and classification and identification.", "key_idea": "In order to improve the accuracy of rockburst tendency prediction, the authors propose a SOM clustering-combined weighting VIKOR model, which takes into account three indicators: rock brittleness index, tangential stress index, and elastic strain energy index.", "method": "Samples are classified accurately through a self-organizing feature mapping network, and the weights of different indicators are calculated using a combination weighting method. The rockburst grades are then sorted using a multi-criteria compromise solution sorting method.", "outcome": "The simulation calculation of the VIKOR rockburst prediction method, based on SOM neural network clustering and combination weighting, is found to be in line with the engineering examples.", "future_impact": "This method has the potential to make the multi-information fusion of rockburst prediction more objective and practicable, leading to safer and more efficient construction practices.", "venue": "WWW", "year": 2021, "title": "A Kind of Prediction Based on SOM Neural Clustering and Combination Weighting Evaluation Method"} +{"pid": "e848a710-0307-452d-82ca-fb23996d6443", "context": "The optimization of bidding strategies in prior-dependent revenue-maximizing auctions, when the seller fixes the reserve prices based on the bid distributions has been a problem of interest.", "key_idea": "The authors introduce a relaxation of this objective functional to use gradient descent methods in optimizing bidding strategies in revenue-maximizing auctions.", "method": "The study uses a variational approach to understand the complexity of the original objective and introduces a relaxed gradient descent method.", "outcome": "The new strategies derived from their approach yield massive uplifts as compared to the traditional truthfully bidding strategy.", "future_impact": "The approach is general and can be applied to various value distributions and revenue-maximizing mechanisms.", "venue": "WWW", "year": 2019, "title": "Learning to Bid in Revenue Maximizing Auction"} +{"pid": "5c0ac846-9596-4215-a968-0865701b303a", "context": "The symbolism of dreams and their interpretation is a wide and complex area of study and has historical significance across cultures. However, despite many species dreaming, there are numerous open research questions in the field of dream studies.", "key_idea": "This study leverages network science and online data sources to understand symbolism in dreams and their associated meanings. It introduces dream interpretation networks for English, Chinese, and Arabic, representing various cultures.", "method": "The authors analyze communities within these networks to understand semantic relationships between symbols, gain insights about cultures from central nodes, identify interconnections by translating symbols into English, and investigate similarities by analysing the sentiment in symbol interpretations.", "outcome": "It was found that symbols within a community are semantically related, and interpretations within a community tend to have similar sentiment. Structural correlations across networks pointed out relationships between cultures, and the analyses highlighted cultural similarities and differences.", "future_impact": "This form of analysis could pave the way for a deeper understanding of dream symbolism across different cultures and languages, and may have implications for comparative cultural studies.", "venue": "WWW", "year": 2014, "title": "Connecting dream networks across cultures"} +{"pid": "605aa488e4510cd7c86eea0f", "context": "As digital technologies become more prevalent, there is a growing awareness of the importance of good security and privacy practices. However, these tools are usually designed with the living user in mind, with little consideration for how they should or will function after the user has died.", "key_idea": "This study focused on conflicts in the sharing, security, and privacy of post-mortem data, discussing a post-mortem privacy paradox where users recognize the value in planning for their digital legacy, but tend to avoid actively doing so due to existing security tools and practices.", "method": "The authors conducted two workshops with users of password managers to explore their views on the post-mortem sharing, security, and privacy of various common digital assets.", "outcome": "The study revealed a tension between the use of recommended security tools during life and providing appropriate post-mortem access to chosen assets. Most of the participants recognized the value in planning for their digital legacy but avoided actively doing so.", "future_impact": "The study provides design recommendations to encourage digital legacy planning while promoting good security habits during life, potentially influencing development of future tools and systems for managing post-mortem digital data.", "venue": "WWW", "year": 2021, "title": "From Personal Data to Digital Legacy: Exploring Conflicts in the Sharing, Security and Privacy of Post-mortem Data"} +{"pid": "b385edb0-13c2-4f52-868d-5fe5705c137e", "context": "Applications of semantic technologies require representation of and reasoning with structured objects; Objects composed of parts connected in complex ways. However, class descriptions and axioms in OWL cannot be used to describe arbitrarily connected structures, leading to underconstrained representations and performance issues in reasoning.", "key_idea": "The authors propose to extend OWL with description graphs, which allow the description of structured objects in a precise, simple way. They also propose to represent conditional aspects of the domain with SWRL-like rules over description graphs.", "method": "Based on an observation about structured objects, the researchers ensured the decidability of their new formalism. They also developed a hypertableau-based decision procedure and implemented it in the HermiT reasoner.", "outcome": "The authors extracted description graphs from the GALEN and FMA ontologies, classified them successfully, and even detected a modeling error in GALEN.", "future_impact": "The introduction of description graphs to extend OWL could lead to significant improvements in representing complex structures in semantic technologies, which would increase the capacity for inference and prevent performance issues in reasoning.", "venue": "WWW", "year": 2008, "title": "Structured objects in owl: representation and reasoning"} +{"pid": "ba71c005-376f-4f48-bda0-4da626a159bf", "context": "Community Question Answering has become a popular way of fulfilling varied information needs. Despite this, the vast body of knowledge generated by such services is not easily accessible due to the personal, often unsubstantiated nature of many answers and the need for a ranking that combines both relevance and quality.", "key_idea": "To address this problem, the authors propose a general ranking framework for factual information retrieval from social media, focused on retrieving well-formed, factual answers to questions.", "method": "A large scale evaluation is performed on a standard factoid QA benchmark. The system's learning framework is tuned with a minimum of manual labelling.", "outcome": "The results exhibit the effectiveness of the method at retrieving factual answers to questions. The authors also provide result analysis to understand which features are significant for social media search and retrieval.", "future_impact": "The system could be used as a crucial component for integrating results from a variety of social media content with general web search results, improving the overall effectiveness of information access from these platforms.", "venue": "WWW", "year": 2008, "title": "Finding the right facts in the crowd: factoid question answering over social media"} +{"pid": "6f8a4479-b614-4ae2-9f8e-aad11ad43e85", "context": "Modern question answering (QA) systems integrate numerous components to conduct specific tasks in a QA pipeline such as Named Entity Recognition, Disambiguation, Relation Extraction, and Query Building, creating a challenge in selecting and combining the most suitable components for a given question.", "key_idea": "The authors propose a study on the optimization problem of QA systems and train classifiers that optimize the selection of QA components based on features of a given question. They also devise a greedy algorithm to identify the most suitable pipelines.", "method": "The model is implemented within a QA framework named Frankenstein which can select the appropriate QA components and compose QA pipelines. This model is evaluated using the QALD and LC-QuAD benchmarks.", "outcome": "Frankenstein precisely solves the QA optimization problem, enables automatic composition of optimized QA pipelines, and outperforms the static Baseline QA pipeline.", "future_impact": "The automated pipeline generation process allows for a flexible inclusion of new QA components in Frankenstein, potentially improving the performance of generated pipelines.", "venue": "WWW", "year": 2018, "title": "Why Reinvent the Wheel \u2013 Let\u2019s Build Question Answering Systems Together"} +{"pid": "491517ad-0dbb-4925-ad98-9f7eacd94263", "context": "In crisis situations, information cascades occur on social media platforms like Twitter, often involving alerts and warning messages from emergency managers. However, the dynamics of such information diffusion and its consequent patterns are less understood.", "key_idea": "The authors propose a model for the diffusion of actionable information to study and understand the patterns and flow of information cascades on Twitter during a crisis situation.", "method": "The authors identify types of information requested or shared during a crisis and study how these messages spread among Twitter users. They analyze information cascades or patterns and interpret what these patterns reveal about information flow and user behavior.", "outcome": "The authors observe patterns of information cascades on Twitter during a crisis and derive insights on information flow and users' behaviors.", "future_impact": "Emergency managers will be able to use the findings and the proposed model to either facilitate the spreading of accurate information or impede the flow of inaccurate or improper messages during a crisis.", "venue": "WWW", "year": 2012, "title": "Information cascades in social media in response to a crisis: a preliminary model and a case study"} +{"pid": "35948cea-2fbc-4ffd-ad70-fd165ad7d208", "context": "Digital badges are widely used by online platforms to encourage users to take certain actions, but their causal effect on user behavior isn't well understood due to the lack of counterfactual data and complexity of factors influencing user behavior over time.", "key_idea": "The authors focus on the causal effect of first-time badges, which are awarded after a user first performs a specific action, the authors also introduce a causal inference framework for these badges.", "method": "The authors utilize the delayed introduction of several badges in a popular Q&A website, and introduce technical innovations like a survival-based hypothesis testing procedure and a bootstrap difference-in-differences method.", "outcome": "It is found that first-time badges steer user behavior if the initial benefit from the corresponding action is low enough. If this isn't the case, no significant effect was observed. For badges that did successfully steer behavior, a counterfactual analysis showed that they improved site community functionality.", "future_impact": "The new causal inference framework for analyzing the impact of digital badges could inform their design and deployment, opening up new opportunities to tailor and optimize user interaction with online platforms.", "venue": "WWW", "year": 2018, "title": "On the Causal Effect of Badges"} +{"pid": "7a35ad83-ef7a-4e6b-a1c4-978743c3c6fc", "context": "The TAO Array is a major part of global climate monitoring, consisting of around 70 moored buoys across the Equatorial Pacific Ocean. It telemeters atmospheric and oceanographic data to on-shore computers via the NOAA Argos satellite system.", "key_idea": "The real-time data from the TOGA-TAO Array of moored buoys is being made accessible through a Mosaic interface designed by the TAO Project office.", "method": "Present displays include latitude-time plots, longitude-time plots, and latitude-longitude plots of surface variables, and are being expanded to incorporate additional views of the four-dimensional data set from the buoy array, including animations of the buoy data and comparisons of the buoy data with operational analyses model data from the National Meteorological Center.", "outcome": "The TAO array allows, for the first time, observationally coherent basin-wide fields of wind and upper ocean thermal and current structure to be constructed in real-time for assimilation into numerical models and for diagnostic studies.", "future_impact": "The availability of these real-time data through a Mosaic interface provides an expanded view of the four-dimensional dataset from the array of buoys for use in future climate monitoring and diagnostic studies.", "venue": "WWW", "year": 1995, "title": "Mosaic access to real-time data from the TOGA-TAO array of moored buoys"} +{"pid": "635bcaaf90e50fcafd33c312", "context": "The authors held data literacy workshops for adults, specifically targeting socially, economically, or educationally disadvantaged groups in Ireland from April 2021 to January 2022.", "key_idea": "This paper presents an overview of the structure of the data literacy workshops, challenges encountered (mainly due to the pandemic), and the lessons learned.", "method": "They describe their approach to the workshops and present an analysis of participant evaluations.", "outcome": "The authors reported a high drop-out rate, but those who participated and filled out the evaluation form expressed satisfaction with the workshop's content and format.", "future_impact": "The authors indicate the crucial role of collaboration between scholars and education and public engagement teams (EPE) team members, with both parties perceived as equals, in successful project implementation.", "venue": "WWW", "year": 2022, "title": "One year of DALIDA Data Literacy Workshops for Adults: A Report"} +{"pid": "e812e793-131a-4359-84af-c0288ed684af", "context": "Optimal interaction and system performance in resource-limited clients like wireless PDAs necessitate customization in the view of the Web. However, there's a lack of a dynamic middleware to enable this.", "key_idea": "The paper introduces a new Web middleware architecture called Web Stream Customizers (WSC) which allows users to customize their view of the web. WSC provides two points of control between client and server for adaptive system-based and content-based customization.", "method": "The authors discuss the implementation of the WSC architecture and use it in three non-trivial, adaptive Customizer applications to illustrate its usage.", "outcome": "The authors report that the overhead of implementing the WSC architecture is small and tolerable, and the benefits outweigh the implementation cost.", "future_impact": "This new middleware could enable significantly improved performance, reliability, or security of wireless client interactions with web services by allowing for dynamic and adaptive content customization.", "venue": "WWW", "year": 2002, "title": "A web middleware architecture for dynamic customization of content for wireless clients"} +{"pid": "650c0146-30a1-40e4-9aa4-1d56992e7ac0", "context": "In online advertising, advertisers use budget management to control their cumulative expenditures. There are multiple ways to achieve this goal, and each one steers the system toward a different equilibrium. However, existing research primarily focuses on online stochastic optimization techniques or game-theoretic equilibria of such situations.", "key_idea": "The authors aim to compare the influence of various budget management strategies on the system equilibrium in terms of the seller's profit and buyers' utility. They examine six strategies: probabilistic throttling, thresholding, bid shading, reserve pricing, and multiplicative boosting.", "method": "The authors assert the existence of a system equilibrium for these methods in a general setting, proving dominance relations between them in a simplified situation. They empirically compare the system equilibria of these strategies using real ad auction data in sponsored search and randomly generated bids.", "outcome": "The authors verify the existence of a system equilibrium for the different budget management strategies in a generalized setting and establish dominance relations for the simplified setting. The empirical study using real-life data corroborates the theoretical findings about the relative performances of budget management strategies.", "future_impact": "This study helps to understand the impact of different budget management strategies on the balance between the seller's profit and buyers' utility which could guide choice of strategies in future research and implementations.", "venue": "WWW", "year": 2017, "title": "Budget Management Strategies in Repeated Auctions"} +{"pid": "7e667026-5fa5-4d43-8f07-aa6a69bb35d1", "context": "Despite being the most widely-used mechanism for authenticating users on the web, passwords face critical security concerns such as phishing and theft of password databases. These issues are made worse by users' tendency to reuse passwords across different services. Current solutions only address one of these concerns, do not protect passwords against rogue servers, and do not provide verifiable evidence of their server-side adoption.", "key_idea": "A novel solution called SafeKeeper is presented that ensures the secrecy of passwords in web authentication systems and can protect against strong adversaries, including external phishers and rogue servers.", "method": "SafeKeeper is developed and deployed using widely available hardware-based trusted execution environments like Intel SGX. It requires only minimal changes for integration into web platforms such as WordPress. An 86-participant user study, systematic analysis, and experiments were conducted to assess its usability, security, and deployability.", "outcome": "The experiments and user study revealed that SafeKeeper imposes negligible performance overhead, and is relatively inexpensive and easy to deploy, showing its strong usability, security, and deployability.", "future_impact": "SafeKeeper, which is available as open-source, has the potential to improve web authentication by mitigating critical security concerns related to passwords, offering a robust solution that can be easily adopted by popular web platforms.", "venue": "WWW", "year": 2018, "title": "SafeKeeper: Protecting Web Passwords using Trusted Execution Environments"} +{"pid": "96fe1e86-03db-4b71-aa07-8fb7b224f5f6", "context": "Community-contributed collections of rich media on the web have been an underused resource for generating representative and diverse views of the world's landmarks, which is a common search task.", "key_idea": "The authors propose an approach using a combination of context- and content-based tools, as well as location, metadata, image tags, and visual features, to automatically generate representative sets of images for location-driven features and landmarks.", "method": "The presented approach involves extracting tags that represent landmarks and using unsupervised methods to extract representative views and images for each landmark. The system is evaluated by using a real-life dataset consisting of 110,000 images from the San Francisco area.", "outcome": "The paper unveils an approach that enables automatic generation of representative sets of images for landmarks, but does not provide specific measurable results.", "future_impact": "The approach presented in this study has the potential to scale up and provide better search and representation for landmarks worldwide.", "venue": "WWW", "year": 2008, "title": "Generating diverse and representative image search results for landmarks"} +{"pid": "6274c5a75aee126c0f6f5867", "context": "Rumors on social media platforms such as Twitter can harm social stability and people's day-to-day life. Traditional rumor detection methods based on text features and the structural information of rumor propagation trees often suffer from poor generalization and lack of robustness against adversarial rumors and noise.", "key_idea": "A novel Graph Adversarial Contrastive Learning (GACL) method is proposed, that uses contrastive learning as part of the loss function to perceive differences between conversational threads of the same class and different classes, and an Adversarial Feature Transformation (AFT) module to generate adversarial samples.", "method": "The authors tested the GACL method using three public benchmark datasets. The adversarial samples generated by the AFT module were used as hard negative samples in contrastive learning to improve the robustness and effectiveness of the model.", "outcome": "Experimental results show that the GACL method outperforms other state-of-the-art models in rumor detection on social media.", "future_impact": "The proposed GACL method may enhance the performance of rumor detection in complex cases, offering potential improvements in maintaining online information integrity and social stability.", "venue": "WWW", "year": 2022, "title": "Rumor Detection on Social Media with Graph Adversarial Contrastive Learning"} +{"pid": "b9143609-c463-4d57-b17b-a5c2cdbdbaa2", "context": "High resolution and accurate colorimetric images from the new generation of image acquisition systems in art galleries are difficult to browse over the web due to their size and the need for color consistency.", "key_idea": "The authors are developing a system for web-based browsing of high-resolution, colorimetric images by holding images as tiled pyramidal JPEG TIFF files at a server and transmitting only the required portions of an image to a Java client, ensuring accurate and consistent color.", "method": "The system is being tested as part of the Euro-Canadian Viseum project, which has set up an ATM network from Vancouver to Berlin, Paris (the Louvre), and London (The National Gallery) to provide secure image archive access for museums.", "outcome": "This system allows for screen comparisons of images from different sites, where any differences in appearance are due to real differences between the objects, which is critical for art and conservation applications.", "future_impact": "This system could potentially revolutionize how one accesses and compares high-resolution images from different sources over the web, particularly for applications in art and conservation.", "venue": "WWW", "year": 1998, "title": "High resolution colorimetric image browsing on the Web"} +{"pid": "34fa0c1e-211c-4362-8165-324599c2e73d", "context": "The classic multi-vehicle minimum latency problem involves routing vehicles located at depots to serve requests located at different points in a metric space so as to minimize total latency. However, it does not consider scenarios with point-to-point requests and release-time constraints that model real-world taxi rides and deliveries.", "key_idea": "This paper proposes to study variants of the multi-vehicle minimum latency problem that can handle point-to-point requests with source-destination pairs and release-time constraints.", "method": "The authors demonstrate constant-factor approximation algorithms for all considered variants using a linear programming framework. They also conduct an empirical study of heuristics based on these theoretical algorithms using a real data set of taxi rides.", "outcome": "The paper provides the first known results for the discussed variants of the multi-vehicle minimum latency problem. Their algorithms show promise in using a linear programming framework.", "future_impact": "The study presents a new approach to optimizing latency in online ride and delivery services. It can potentially lead to improved service efficiency by allowing more realistic task handling.", "venue": "WWW", "year": 2018, "title": "Minimizing Latency in Online Ride and Delivery Services"} +{"pid": "60641c769e795e72406b65d9", "context": "Deep learning-based video manipulation methods, including deepfakes, are increasingly accessible and their detection is becoming crucial. Existing deep learning-based detection methods lack transferability and generalizability, performing poorly on types of deepfakes they were not trained on, including real-world deepfakes.", "key_idea": "The proposed solution to this problem is a Convolutional LSTM-based Residual Network (CLRNet) that adopts a unique model training strategy to better handle unknown and unseen deepfakes. The key idea is to develop a generalized approach to detect multiple types of deepfakes, including deepfakes from unknown generation methods such as DeepFake-in-the-Wild (DFW) videos.", "method": "The authors evaluate CLRNet through extensive experiments, including an evaluation with a high-quality DeepFake-in-the-Wild dataset collected from the internet with numerous videos and over 150,000 frames.", "outcome": "The CLRNet achieves far better generalization when detecting various benchmark deepfake methods (97.57% on average). It also outperforms existing state-of-the-art defense methods by achieving 93.86% detection accuracy on high-quality DFW videos.", "future_impact": "The introduction of CLRNet may suggest an improving ability to cope with the detection of unknown and unseen deepfakes, potentially enhancing defences against deepfake manipulation.", "venue": "WWW", "year": 2021, "title": "One Detector to Rule Them All: Towards a General Deepfake Attack Detection Framework"} +{"pid": "b6ff9efa-b2ef-4fb2-9d5c-eba04ffcadb5", "context": "People seek majority opinions online but often have to read through large numbers of documents returned by search services, or wait a significant amount of time for responses when they upload their own questions on social media sites.", "key_idea": "The authors propose Searching Majority Opinions System (SEMO), a new majority opinion-based search system that uses QA threads from SNS and cQA websites.", "method": "The authors tested the system in the movie domain and developed a prototype of SEMO in order to prove its usefulness for finding majority opinions. They also addressed the data sparsity problem using a novel query component expansion approach.", "outcome": "SEMO was implemented successfully, providing real-time results based on majority opinions for opinion-finding queries. The system is available online for public use.", "future_impact": "SEMO brings a significant change in the opinion-finding query search domain and can help people make decisions more conveniently by giving real-time response to opinion-based queries.", "venue": "WWW", "year": 2016, "title": "SEMO: Searching Majority Opinions on Movies using SNS QA Threads"} +{"pid": "bb3d1416-f37f-4e7b-9fd6-504706281782", "context": "Designing interactive systems for older adults often involves them from the early stages of development, but the exact methods and results of such an approach have not been thoroughly explored.", "key_idea": "This work introduces the co-design of a smartpen and paper calendar-based reminder system for older adults, leveraging multimodal interfaces with daily routines.", "method": "Co-design sessions were conducted with older adults, using experience prototypes to explore interactions between paper-based calendars and multimodal reminders using a smartpen.", "outcome": "Initial results demonstrate successful techniques linking paper interaction and scheduling reminders, such as using smartpen annotations and the location of written reminders on a paper diary to schedule digital reminders.", "future_impact": "The observations about physical aspects of paper diaries, such as layouts and binding, as discussed by older adults may inform future design of similar systems for older users.", "venue": "CHI", "year": 2013, "title": "Designing a smartpen reminder system for older adults"} +{"pid": "60a4e58991e011f4291ce47a", "context": "As digital fabrication machines become widespread, online communities are forming that provide a space for diverse practitioners to share their work, troubleshoot, and socialize. These communities pioneer increasingly novel fabrication workflows, which may not necessarily align with traditional manufacturing models.", "key_idea": "The authors conduct a qualitative study of the #PlotterTwitter community, who develop custom hardware and software tools to create artwork with computer-controlled drawing machines, known as plotters, which defy traditional interpretations of digital fabrication workflows.", "method": "The authors document and analyze emergent themes where the traditional interpretation of digital fabrication workflows fails to capture important nuances and nascent directions within the #PlotterTwitter community.", "outcome": "The study find that #PlotterTwitter makers champion creative exploration of interwoven digital and physical materials over a predictable series of steps challenging long-running views of digital fabrication.", "future_impact": "The findings propose design implications for future frameworks and toolkits to account for this breadth of practice, which could reshape perceptions and practices in digital fabrication.", "venue": "CHI", "year": 2021, "title": "Tools, Tricks, and Hacks: Exploring Novel Digital Fabrication Workflows on #PlotterTwitter"} +{"pid": "854f0245-db0e-42ba-8295-00faed83e3ed", "context": "With an increasing interaction of users with a variety of mobile apps under different usage contexts, it has become critical to involve users in the app design process. However, existing apps and app stores only offer limited ways for user interaction, like posting reviews and sending email reports. Moreover, developers are dealing with information overload due to a large number of app comments.", "key_idea": "The authors propose to handle this information overload faced by developers because of a plethora of app comments by carrying out a simple content analysis of app reviews from the developer's perspective, and introducing an algorithm that automatically identifies informative and relevant reviews reflecting user involvement.", "method": "They first undertook a content analysis of app reviews from a developers' point of view. Then, an algorithm was proposed for identifying informative app reviews that reflected user involvement and participation. This algorithm was then evaluated preliminarily.", "outcome": "The preliminary evaluation results document the efficiency of the proposed algorithm.", "future_impact": "Their preliminary work potentially facilitates the first steps towards building a unified platform aimed at enhancing user involvement across various apps, thereby improving app design and user satisfaction.", "venue": "CHI", "year": 2013, "title": "Facilitating developer-user interactions with mobile app review digests"} +{"pid": "f97a0eef-4b24-41d4-80e3-da0e99d57a1c", "context": "Traditional volunteering involves individuals working directly for a charity, however, this model may not be suitable or convenient for everyone.", "key_idea": "The study introduced PledgeWork, a novel form of volunteering where individuals complete tasks provided by a third party task provider to make indirect donations to charities.", "method": "A mixed methods study was conducted to evaluate user acceptance of online volunteering, quantity and quality of work performed as a volunteer, and potential benefits for charities, by comparing the performance and attitudes of volunteer workers and paid workers towards PledgeWork.", "outcome": "The study found that PledgeWork improved the quality of simple tasks and that most participants expressed interest in using the PledgeWork platform to contribute to a charity. The study also uncovered current issues with volunteering and online donations.", "future_impact": "The emergence of innovative volunteering models such as PledgeWork could potentially enhance the accessibility of volunteering and provide new opportunities for charities to raise funds and resources.", "venue": "CHI", "year": 2019, "title": "PledgeWork: Online Volunteering through Crowdwork"} +{"pid": "df3e7b8d-508a-4a8d-b52b-c375a348cb3f", "context": "Many individuals manage a complex array of digital information in their lives, including volunteer coordinators at nonprofit organizations who manage data about millions of volunteers, yet current information management systems do not meet their needs.", "key_idea": "The study examines the information management practices of volunteer coordinators and characterizes their self-made, ad-hoc solutions, termed as 'homebrew databases'.", "method": "The authors conduct a qualitative study of the information management practices of volunteer coordinators in nonprofit organizations.", "outcome": "The study identified resource constraints, diverse and fluid information needs, stakeholders, and work contexts that drive the information management strategies of these coordinators. It also highlighted deficiencies in the appropriateness and usability of databases and information management systems.", "future_impact": "The recognition of 'homebrew databases' and the challenges arising from their use could influence the future development of more suitable information management systems.", "venue": "CHI", "year": 2011, "title": "Homebrew databases: complexities of everyday information management in nonprofit organizations"} +{"pid": "635bcab490e50fcafd33cb1d", "context": "Design systems provide detailed guidelines on UI items\u2019 appearance and behavior and serve as a primary reference for interface development for languages like Arabic and Hebrew. However, investigating the correct direction of all items could involve costly research and this challenge has never been addressed before.", "key_idea": "For this challenge, the authors carried out five focus groups with HCI professionals and identified a set of prioritization criteria for bidirectionally designed UI items.", "method": "The authors conducted focus groups with HCI professionals to identify a set of prioritization criteria and used these criteria to prioritize the UI items for future empirical evaluation of their preferred directionality.", "outcome": "The identified criteria allowed the authors to prioritize the UI items for future empirical evaluation of their preferred directionality.", "future_impact": "The findings could help to provide a structured approach for future research to prioritize UI items for bidirectional language design systems and can also give some insights from their experience of using focus groups for the prioritization task.", "venue": "CHI", "year": 2022, "title": "Prioritizing UI Items for Bidirectional Design System Development: A Suggested Set of Criteria and its Implementation"} +{"pid": "5ea2c2ea9fced0a24b1ddd77", "context": "Human-Computer Interaction (HCI) research has recently been focused on transient, non-digital methods of displaying information in public places, such as chalk infographics. Despite their advantages, such as being easy to deploy, easy to interact with, and more sustainable, they have the limitation of only having local scale and impact.", "key_idea": "This study introduces the concept of using urban robots, specifically a slow-moving robot named Woodie that uses chalk sticks to draw on the ground, as an approach to scale up these physicalised displays.", "method": "The authors deployed Woodie for three weeks in a quiet urban alley, and they collected data via observations, video logs, and interviews.", "outcome": "Their findings revealed that Woodie was successful in attracting people's attention and facilitating collaborative, creative placemaking. Additionally, Woodie evoked emotional responses and was perceived as a living being.", "future_impact": "The study highlights potential opportunities that robotic displays can offer for future designs of pervasive urban public displays.", "venue": "CHI", "year": 2020, "title": "Stop and Smell the Chalk Flowers: A Robotic Probe for Investigating Urban Interaction with Physicalised Displays"} +{"pid": "51256eb0-063e-41a9-b821-7d6932b75455", "context": "Medical fundraising campaigns are migrating to online platforms in the form of crowdfunding, removing reliance on traditional organizations' methods for achieving credibility. Individual fundraisers must establish their own credibility, hence potential donors seek new factors to assess the credibility of these campaigns.", "key_idea": "The study investigates current practices in assessing the credibility of medical crowdfunding campaigns. The authors identify the distinctive validation practice of collective endorsements in online medical crowdfunding, which is formed by the online presence of close connections and external online communities.", "method": "Authors conducted a mixed-methods study analyzing data from social media and semi-structured interviews to investigate the practices of assessing credibility of online medical crowdfunding campaigns.", "outcome": "The study discovered eleven factors associated with the perceived credibility of medical crowdfunding, with three being unique to medical crowdfunding. One distinctive validation is the collective endorsement formed by the online presence of close connections and external online communities in online medical fundraising campaigns.", "future_impact": "The findings can be used by fundraisers to improve their campaigns' perceived credibility by leveraging the concept of collective endorsements.", "venue": "CHI", "year": 2016, "title": "The Power of Collective Endorsements: Credibility Factors in Medical Crowdfunding Campaigns"} +{"pid": "3ed583ef-7131-41f4-bafb-aba100d2b4e8", "context": "Latency during video scrubbing significantly degrades navigation performance, as deduced from a conducted study using abstracted video content.", "key_idea": "The authors introduce Swift, a technique for real-time scrubbing of online videos that overlays a small, low-resolution version of the video during scrubbing, and reverts to the high-resolution video when scrubbing is completed or paused.", "method": "The authors conducted a comparison study of the Swift technique against traditional online video players using a set of realistic live motion videos and content-specific search tasks.", "outcome": "Swift technique reduces navigation completion times by up to 72% even with a relatively low latency of 500ms.", "future_impact": "The authors anticipate that the Swift technique can be easily implemented using modern HTML5 web standards, potentially heightening user experience in online video scrubbing.", "venue": "CHI", "year": 2012, "title": "Swift: reducing the effects of latency in online video scrubbing"} +{"pid": "905284b6-285f-4715-b4a8-8a0afffa96a4", "context": "People's management of personal and familial information is changing as they increasingly turn to digital channels for sharing, storing, and reflecting on their lives and experiences.", "key_idea": "This study focuses on the creation of four interactive, provocative systems, derived from previous research, to investigate how digital systems influence the management, curation, and recall of personal and familial digital information.", "method": "The authors conduct sessions with 12 adults from Pittsburgh, PA, using these created systems in combination with interviews to understand their role in resurfacing personal and familial digital information.", "outcome": "The findings suggest that there is an opportunity to create systems that facilitate the curation and transmission of digital content and make sense of the differences between how systems and people recall and represent their experiences.", "future_impact": "This work could lay the groundwork for the design of future systems, capable of openly mediating the curation and transmission of digital content, that are more in sync with how humans recall and represent their experiences.", "venue": "CHI", "year": 2015, "title": "Curatorial Agents: How Systems Shape Our Understanding of Personal and Familial Digital Information"} +{"pid": "c1addf24-c032-455e-9388-48efb270f0af", "context": "Prior to this research, there was no established device that interpreted user-intent based on interpreting signals from capacitive sensor patterns and orientation sensors.", "key_idea": "The study introduces Chiron, a wearable device for the hand, that reads digital and analog signals from capacitive sensor patterns and orientation sensors in order to interpret user intent, enabling intuitive shape modelling and control.", "method": "The study explores two cases with Chiron: (a) an unconventional and low-cost method for intuitive shape modeling and control, and (b) ergonomically designing sensor patterns from conductive ink for reading localized finger interactions.", "outcome": "Chiron was able to use signals from capacitive sensor patterns and orientation sensors to control and model shapes intuitively, and the interpretation of localized finger interactions such as swiping or pinching was successful.", "future_impact": "The researchers envision future applications of the thumb-based interaction mechanism exploited in Chiron's design, although they do not specify what these might be.", "venue": "CHI", "year": 2014, "title": "Chiron: interpreting signals from capacitive patterns and inertial sensors for intuitive shape modeling"} +{"pid": "cb8af689-05c3-41d2-9f65-f2088565b78b", "context": "The rampant spread of fake news has prompted a need for novel literacy education tools and the understanding of audience perceptions on the same.", "key_idea": "The study introduces a game, consisting of two main iterative designs, designed to serve as a literacy education tool for distinguishing fake news from real news.", "method": "The game underwent two main design iterations and the user interactions from these versions were assessed, with more than 500,000 plays and data from over 45,000 players collected and analyzed.", "outcome": "The data gathered indicated that people above 70 years of age are better at identifying fake news, and that higher education correlates to better performance at discerning real news from fake. It was also observed that the time taken to identify fake from real news varies.", "future_impact": "This case study indicates that such game designs hold potential to collect data useful to non-game contexts and can help further study the nature of perception of fake news.", "venue": "CHI", "year": 2019, "title": "Factitious: Large Scale Computer Game to Fight Fake News and Improve News Literacy"} +{"pid": "0b9066d3-4500-4d64-a403-173b9bbb973b", "context": "Web pages are represented in different ways depending on the user interaction context, like search results or browser history links. Previous research has focused on optimizing these representations for specific tasks, however, the importance of consistency across different tasks has not been addressed.", "key_idea": "The paper introduces the concept of a compact web page representation that bolsters the process of identifying new, relevant web pages and rediscovering previously viewed pages.", "method": "The authors examine how different web page representations are deployed across varied use cases.", "outcome": "The study reveals that consistency in representation across tasks is important in web page interactions.", "future_impact": "This work opens up paths for further research on designing and implementing web page representations that account for diverse user needs and interaction contexts.", "venue": "CHI", "year": 2009, "title": "Visual snippets: summarizing web pages for search and revisitation"} +{"pid": "6287042a5aee126c0f5b6df0", "context": "The transition to renewable energy increases the risks of disruptions when electricity supply does not meet demand. HCI has explored how digital technologies can mitigate such problems in households through support for reducing or shifting electricity use.", "key_idea": "The paper proposes the notion of 'household energy resilience' and suggests that faster transitions to renewable energy may be possible if households adapt to disturbances in power supply, rather than just trying to mitigate them.", "method": "The authors study 21 Swedish households and their experiences of and ideas on managing disruptions in electricity supply, leading to the identification of three resilience strategies: response diversity, creating opportunities to develop resilience and building community energy resilience.", "outcome": "Three strategies for household energy resilience are identified from the study: 1) response diversity, 2) creating opportunities to develop resilience, and 3) building community energy resilience.", "future_impact": "The findings suggest that HCI can support these strategies both by providing tools to increase resilience and by carefully designing technology and services to be more resilient in themselves, influencing future design and use of technologies in a renewable-energy-dependent society.", "venue": "CHI", "year": 2022, "title": "Exploring Renewable Energy Futures through Household Energy Resilience"} +{"pid": "5ffc8d52-9bc1-43f1-889d-c6573d2c4b5f", "context": "Much of software development involves reengineering, often driven by usability problems. However, existing reengineering planning processes do not adequately address software usability concerns.", "key_idea": "The authors propose a new method for reengineering planning guided by usability evaluation, termed 'Reengineering Planning Process Guided by Usability Evaluation' (RPP-U), to incorporate usability evaluation methods into reengineering planning.", "method": "RPP-U is structured around five phases where usability evaluation methods are applied, including Project Justification. Here, various usability testing methods such as CIF, Heuristic Evaluation, Cognitive Walkthrough, and GOMS Model are used to analyze software quality, maintenance required, and business value", "outcome": "No specific outcome is mentioned in the abstract.", "future_impact": "The primary anticipated impact of this research is to enable more effective reengineering planning approaches by integrating usability evaluations into the planning process.", "venue": "CHI", "year": 2004, "title": "Reengineering planning process guided by usability evaluation"} +{"pid": "1bdff54b-dc38-48b8-93df-2c59079c1013", "context": "Video mediated communication (VMC) systems are emerging, but their role in supporting everyday activities, specifically related to the dynamics of private and public spaces, is yet to be explored in detail.", "key_idea": "The authors propose the concept dwelling of the future, comHOME, exploring the integration of different communication zones where the residents can be seen or heard, using VMC systems in the context of private and public spaces.", "method": "The authors provide video-based scenarios demonstrating the usage of comHOME for everyday activities.", "outcome": "A main observation from the making of the video is its beneficial role as a complementary method in a complex design process due to the focus on the user perspective.", "future_impact": "The architectural space, combined with information and communication technology (ICT), could form an enhanced interface to the digital world, aiding the creation of future dwelling spaces.", "venue": "CHI", "year": 2000, "title": "Private and public spaces: the use of video mediated communication in a future home environment"} +{"pid": "00fe473f-540a-4ef9-91bf-e330b47898f5", "context": "Designers are increasingly using online crowds for feedback; however, these contributors may lack the expertise, context, and sensitivity to provide effective critique. Rubrics help feedback providers but require domain experts to write them and may not generalize across different design domains.", "key_idea": "The paper proposes a semi-automated method to support feedback providers by analyzing feedback language using natural language features such as specificity and sentiment. This method includes a critique style guide with feedback examples to help providers revise their feedback through self-assessment.", "method": "The authors conducted two studies: the first involved 52 students from two design courses receiving feedback from 176 online providers, from which an algorithm extracted a set of natural language features. The second study tested the validity of the guide through a between-subjects experiment where providers wrote feedback on design solutions with or without the guide.", "outcome": "The study found that natural language features correlated with the ratings and accurately predicted these ratings. When using the style-based guide, providers produced feedback that was perceived as more helpful.", "future_impact": "The semi-automated method can be used to improve the quality of critique offered by online feedback providers across different design domains.", "venue": "CHI", "year": 2017, "title": "Critique Style Guide: Improving Crowdsourced Design Feedback with a Natural Language Model"} +{"pid": "9ac4420f-1f85-478d-9514-2a27bebddc65", "context": "While there is substantial coverage of methodological and design-related topics in the Human-Computer Interaction (HCI) literature, there is an observed lack of theoretical focus in the rapidly increasing work on user experience (UX).", "key_idea": "The authors highlight the need for a theoretical discussion around UX research and practice.", "method": "By analyzing 122 individual items on theories collected in a CHI 2011 special interest group session on UX theories and theoretical frameworks, the data set was filtered and categorized in several iterations, resulting in 56 items distributed over 7 major theory categories and linked to 9 relevant disciplines.", "outcome": "The categorization of items led to an initial mapping of the field, pointing towards directions for further conceptual and theoretical clarification. This helped to explore the multi-disciplinary nature of UX.", "future_impact": "These results can aid in building a more solid foundation for UX research and practice.", "venue": "CHI", "year": 2012, "title": "In search of theoretical foundations for UX research and practice"} +{"pid": "9ef1cade-50a3-4491-b42c-db9eb331d2b9", "context": "Multi-national projects face challenges in finding and coordinating resources, managing logistics in different countries, defining research methodology, controlling project cost, and dealing with cross-cultural issues. Different approaches can be used to manage such ventures which include centralized, decentralized and cooperative strategies.", "key_idea": "The authors discuss their experience in managing international usability teams using the cooperative strategy.", "method": "The authors share their experience in managing a multilingual project with teams based in four countries. They give detailed insights related to the strategy applied, specifically addressing the setup and data analysis stages, from the perspectives of a global project manager and a local coordinator.", "outcome": "The report details both the challenges faced and the valuable lessons learned during the execution of the project. Both positive and negative lessons from the project are listed.", "future_impact": "This paper could serve as a valuable resource for others to learn from the authors' experiences in managing multi-national projects using a cooperative strategy.", "venue": "CHI", "year": 2006, "title": "Managing international usability projects: cooperative strategy"} +{"pid": "d3d9abdd-7b9e-4287-8091-2d15000d9568", "context": "The current state of file retrieval is typically slow and inefficient, often lacking useful tools to assist users in finding target files and folders quickly and effectively.", "key_idea": "BIGFile is introduced, a new file retrieval technique based on the Bayesian Information Gain framework. This provides users with interface shortcuts to assist them in finding target files or folders efficiently.", "method": "The algorithm's computational efficiency was demonstrated through simulation and compared against AccessRank's prediction algorithm. Additionally, a further real-life experiment was conducted to compare the efficiency of BIGFile with ARFile (AccessRank instantiated in a split interface) and a Finder-like list view as a baseline.", "outcome": "The simulations and experiment showed that BIGFile was the most efficient technique, being up to 44% faster than ARFile and 64% faster than Finder.", "future_impact": "This study may inspire further research in improving file retrieval mechanisms and inform future designs. The unanimous preference for split interfaces may also influence user interface design in other areas.", "venue": "CHI", "year": 2018, "title": "BIGFile: Bayesian Information Gain for Fast File Retrieval"} +{"pid": "5bf221bf-1370-4ac2-8fe3-9bab51a8caf8", "context": "Nearly 40 million workers in the USA, a third of the working population, are low-wage, meaning they make less than $11.65 per hour. These workers face the pervasive and detrimental challenge of wage violations, also known as wage theft, which is any illegal activity by an employer that denies benefits or wages to employees.", "key_idea": "Based on the interviews with 24 low-wage workers who experienced wage theft and sought justice, the researchers identified key sociotechnical practices engaged in by these workers to address their wage theft, such as identifying wage and payment discrepancies, tracking and documenting work, and pursuing wage claims.", "method": "The research was based on interviews with 24 low-wage workers who had experienced wage theft and sought justice regarding their work practices, challenges and the ways they use information technology.", "outcome": "The researchers identified three key sociotechnical practices: 1) identifying wage and payment discrepancies; 2) tracking and documenting work; and 3) pursuing wage claims as methods to address wage theft.", "future_impact": "The study prompts HCI (Human-Computer Interaction) researchers to reflect on the possibility and limits of several key design recommendations that could interrupt uneven social, economic, and information relations in the low-wage workplace.", "venue": "CHI", "year": 2017, "title": "Low-Wage Precarious Workers' Sociotechnical Practices Working Towards Addressing Wage Theft"} +{"pid": "0007af11-65dc-4e32-a362-e08f6336c012", "context": "Automatic localization of cultural resources and UIs is crucial for the survival of minority languages, for which there are often insufficient parallel corpora or no corpus at all to build machine translation systems.", "key_idea": "The authors propose a method to compensate for resource-scarce languages by leveraging a family of languages closely related to the speaker's native language to construct translations in a coherent mix of these languages.", "method": "Experimental results are presented, testing the approach through user comprehension and acceptance.", "outcome": "The translations generated by the new method can be easily understood and are a useful aid for users who are not proficient in foreign languages. Also report insights on user acceptance towards the method.", "future_impact": "This work has significant implications for Human-Computer Interaction by establishing a language that can improve how applications communicate to their users.", "venue": "CHI", "year": 2012, "title": "Monsieur, azonnal k\u00f6vessen engem bitte!: an automatically generated interlanguage tailored to speakers of minority but culturally influenced languages"} +{"pid": "9ef08706-be2a-43fb-9e1d-88660fd6a920", "context": "In the context of managing the huge quantity of email people receive each day, attention is selectively given to certain messages over others based on the initial perusal of the email. However, the specific surface features drawing attention to an email are not well-understood.", "key_idea": "The authors investigate how top-level cues about an email's content influence the attention given to the email, with a proposed curiosity theory that suggests uncertainty about message content increases attention to a message.", "method": "Initially, a think-aloud study was conducted to understand the rationale behind email prioritizing actions. Subsequently, an experiment was conducted analyzing how factors like message importance, subject line specificity, workload and personal utility impacts attention towards emails.", "outcome": "The outcome of the experiment suggests that uncertainty about the message content increases the likelihood of attention towards a message. However, this influence of uncertainty diminishes in the face of enhanced task and personal utility cues and increased demand.", "future_impact": "The results of this study are anticipated to be beneficial for the design of intelligent email systems, email client interfaces, and strategies to reduce email strain.", "venue": "CHI", "year": 2011, "title": "Should I open this email?: inbox-level cues, curiosity and attention to email"} +{"pid": "5ea411e39fced0a24bae0cdb", "context": "Satellites have useful lifetimes of only a few decades that could be extended if consumable resources could be replaced. Unmanned orbital robots that are currently semi-autonomous and require monitoring and control by highly trained ground teams are being explored to perform satellite servicing operations. However, this creates a unique and tightly constrained operating environment with very particular interface needs.", "key_idea": "The authors have developed a tele-robotic interface specific to NASA's Restore-L mission, with special attention to its unique circumstances and operating constraints.", "method": "The authors describe the process of developing a tele-robotic interface, taking into account technical and mission parameters, as well as the social and operational contexts.", "outcome": "A tele-robotic interface specific to NASA's Restore-L mission was developed, considering requirements, constraints, technical, and mission parameters, and social and operational contexts.", "future_impact": "The study articulates the need for a design framework that better connects the technical and mission parameters with the social and operational context of tele-robotic interfaces, which could influence future design approaches.", "venue": "CHI", "year": 2020, "title": "Tele-robotic Interface Design in Context: A Case for Recursive Design"} +{"pid": "62abf7095aee126c0f620ce7", "context": "Decision support alerts can assist clinicians in determining appropriate interventions for critically injured patients. However, the design of these alerts is critical because it can impact their adoption and effectiveness.", "key_idea": "This study explores the design of decision support alerts for cognitive aids used in time- and safety-critical medical events.", "method": "The authors conducted interviews with 11 trauma team leaders to get their thoughts and reactions to potential alert designs.", "outcome": "From the findings, the authors identified three implications for designing alerts for cognitive aids that support team-based, time-critical decision making.", "future_impact": "The findings and implications of this study can be further explored in future work to improve the design of decision support alerts for cognitive aids.", "venue": "CHI", "year": 2022, "title": "\u201dPop-Up Alerts are the Bane of My Existence\u201d: Designing Alerts for Cognitive Aids Used in Time-Critical Medical Settings"} +{"pid": "7509b9c2-3c0f-41dc-b04c-feaa0f451372", "context": "Systems that improve access to prior knowledge and promote reuse can improve design efficiency and outcomes. However, such systems have not been widely adopted, indicating a lack of understanding about user needs in early design stages.", "key_idea": "The authors conducted a study focusing on early design to understand the practices and attitudes towards knowledge management and reuse, which complements previous research that majorly focused on knowledge reuse during later design and implementation phases.", "method": "The study involved interviews and surveys with professional designers in creative domains.", "outcome": "The study revealed new findings and implications related to knowledge management and reuse during early design.", "future_impact": "The findings have the potential to enhance the benefits of knowledge management systems and reuse in early design activities.", "venue": "CHI", "year": 2009, "title": "Understanding knowledge management practices for early design activity and its implications for reuse"} +{"pid": "f3545bf5-e9bc-4131-aaa7-1980f9e8925c", "context": "Building help systems for interactive applications is a labor-intensive task and, as of yet, there is no established automated method for generating such systems.", "key_idea": "This paper introduces a model-based interface design system that can automatically generate hypertext-based help about data presented in application displays, commands to manipulate data, and interaction techniques to invoke commands.", "method": "The system described in this paper generates help automatically from the model used to design the interface, and includes a component for refining automatically-generated help texts.", "outcome": "The proposed system allows for several levels of customization, including programming-by-example techniques to let developers directly edit help windows that the system produces, and to refine help generation rules.", "future_impact": "The system's provision of a framework for developers to easily refine automatically-produced help may significantly reduce the effort involved in creating help systems for interactive applications.", "venue": "CHI", "year": 1994, "title": "Automatic generation of help from interface design models"} +{"pid": "f7c45412-7675-4e16-9630-4cc0ae5bb4a4", "context": "Information and Communication Technology (ICT) has been contributing to sustainable development, and its application varies in different regions such as Brazil, U.S., and Europe.", "key_idea": "This work seeks to conduct a systematic review of how ICT contributes to sustainability and how this contribution varies in Brazil, U.S., and Europe.", "method": "The authors adopt a systematic review method to understand both how ICT contributes to sustainability and its application in different geographical contexts.", "outcome": "The systematic review provided some insights on areas which have received more investments and which ones need more attention in terms of ICT application for sustainable development.", "future_impact": "The outputs from this systematic review may be used to formally devise indicators representing the areas that need more attention regarding ICT application for sustainability.", "venue": "CHI", "year": 2013, "title": "Uma revis\u00e3o sistem\u00e1atica sobre tecnologias de informa\u00e7\u00e3o e comunica\u00e7\u00e3o aplicadas \u00e0s pr\u00e1ticas de sustentabilidade"} +{"pid": "5ea2c2ea9fced0a24b1ddbc9", "context": "Early-detection of Autism Spectrum Disorder (ASD) in preschoolers is a challenge due to missed early detections, late- or misdiagnoses in Pakistan.", "key_idea": "The authors propose EAST, an intermediary interactive tablet-based app meant to screen preschoolers for Autism Spectrum Disorder through play-based activities.", "method": "The authors undertook an iterative co-design process, involving medical professionals, parents and teachers, to design and refine EAST. They then tested EAST on both typically developed and autistic children. They evaluated them based on various parameters such as time taken, the number of wrong attempts, or incorrect answers and audiovisual feedback.", "outcome": "The authors found that EAST performs well in terms of its acceptability, usability, and validity. Moreover, it delivers insights into a child's behavior, thereby fostering collaboration among parents, teachers, and medical professionals.", "future_impact": "The developed tool, EAST, holds potential in improving early detection and diagnosis of Autism Spectrum Disorder, enabling collaboration between parents, teachers, and medical professionals.", "venue": "CHI", "year": 2020, "title": "EAST: Early Autism Screening Tool for Preschoolers"} +{"pid": "0829ce90-9409-4584-8501-2d10664be7d5", "context": "Before this study, gaze information was not extensively used as an augmented input along with keyboard and mouse for user interaction.", "key_idea": "The Gaze-enhanced User Interface Design (GUIDe) project explores the effective use of gaze information as an additional input method for interaction with computing devices.", "method": "The authors shared three practical applications of gaze information for pointing and selection, application switching, and scrolling.", "outcome": "The GUIDe project's gaze-based interaction techniques were found not to overload the visual channel and provide a natural, universally-accessible and general-purpose use of gaze information.", "future_impact": "The new methods of user interaction presented in this paper could significantly enhance the way users interact with everyday computing devices in the future.", "venue": "CHI", "year": 2007, "title": "GUIDe: gaze-enhanced UI design"} +{"pid": "056aa5b9-640d-4f52-8df7-104426de4466", "context": "Before this study, the user experience of IBM Lotus Notes required advancements due to usability concerns.", "key_idea": "The authors discuss their approach of involving the community of Notes users in the redesign process of the software, emphasizing a dialogue-rich process.", "method": "The methods the authors used for this redesign of IBM Lotus Notes user experience were crafted to facilitate extensive dialogue with the user community.", "outcome": "The redesign of the IBM Lotus Notes user experience was successfully implemented with the involvement of its user community.", "future_impact": "The findings could influence future software redesign efforts, showing potential benefits and challenges when including end-users in the process.", "venue": "CHI", "year": 2009, "title": "Open by design: how IBM partnered with the user community in the redesign of lotus notes"} +{"pid": "624bb3a35aee126c0fea50aa", "context": "Single-switch users with motor impairments typically use row-column scanning for choosing items arranged in a grid. Although the Nomon system might provide a faster and easier alternative by allowing arbitrary arrangement of selections, its effectiveness over longer time periods or in tasks beyond writing has not been thoroughly quantified.", "key_idea": "The authors attempt to quantify the performance of the Nomon system over time and beyond writing tasks. In this context, the authors also develop a webcam-based switch to simulate the response times of a motor-impaired user, enabling better calibration and initialisation of their method.", "method": "The authors conducted experiments over 10 sessions using the webcam switch, with users performing typing and picture-selection tasks using both Nomon and traditional row-column scanning. Evaluation and feedback were also collected from a motor-impaired switch user.", "outcome": "The experiments demonstrated that users could type faster and more easily with Nomon than with row-column scanning, with the benefits of Nomon being even more pronounced in picture-selection tasks.", "future_impact": "The user feedback and experimental results support the potential for Nomon to improve single-switch communication interfaces, influencing future designs.", "venue": "CHI", "year": 2022, "title": "A Performance Evaluation of Nomon: A Flexible Interface for Noisy Single-Switch Users"} +{"pid": "9a3e5af8-64d8-45a0-b30d-636e9d6d33c4", "context": "The current generation of desktop computers and networks has made streaming audio and video more popular, with a small investment allowing presentations or lectures to be multicast, enabling passive viewing from different locations. However, awareness of and interaction with remote viewers during presentations can be challenging.", "key_idea": "A lightweight system called TELEP has been designed to enhance awareness of remote viewers during presentations and allow interaction between remote viewers and the speaker.", "method": "The authors surveyed experienced viewers of multicast presentations and used their feedback to design and modify the TELEP system.", "outcome": "The authors report on the successful design, use, and modification of the TELEP system, enabling greater awareness of remote viewers during presentations and allowing interaction.", "future_impact": "The design trade-offs discussed in the study could stimulate further improvements to the TELEP system and overall remote presentation and interaction technology.", "venue": "CHI", "year": 2000, "title": "Presenting to local and remote audiences: design and use of the TELEP system"} +{"pid": "5ea2c2ea9fced0a24b1ddce3", "context": "Traditional haptic experiences generally involve feedback mechanisms at the full, human-body scale, limiting the range of experiences that can be realistically reproduced.", "key_idea": "The authors propose Miniature Haptics, a new approach that applies miniaturized haptic feedback to hand-based, embodied avatars, enabling the exploration of haptic experiences that weren't previously possible due to practical restrictions.", "method": "The authors conducted a human factors study to explore how people map their full-body skeletal model to their hands and developed a miniature football haptic display to test user experience. Two miniature motion platforms were also developed for different gaming experiences.", "outcome": "The results from the user study show that Miniature Haptics significantly improved the realism and enjoyment of the experience and are preferred by users. It was effective for rapidly changing ground height in platform jumping games and changing terrain slope.", "future_impact": "Miniature Haptics could make it possible to explore novel haptic experiences that were not practical before this study.", "venue": "CHI", "year": 2020, "title": "Miniature Haptics: Experiencing Haptic Feedback through Hand-based and Embodied Avatars"} +{"pid": "be63e9ca-fc29-43e3-bde7-c7a6945e5d13", "context": "In competitive First-Person Shooter (FPS) games, enjoyment is compromised when players' skill levels differ significantly. Existing player balancing techniques have limitations when addressing substantial skill level differences.", "key_idea": "The authors developed new player balancing schemes to handle a range of FPS skill differences and utilized in one-on-one deathmatches using a commercial-quality FPS game.", "method": "The authors tested the new player balancing schemes in one-on-one deathmatches using a commercial-quality FPS game developed with the UDK engine.", "outcome": "The new balancing schemes effectively balanced players, even those with large skill differences. Notably, the techniques that balanced best were also rated most enjoyable by players despite being the most noticeable.", "future_impact": "This study provides game developers with feasible player balancing techniques for realistic FPS games, expanding the audience for this popular genre. Moreover, it highlights the idea that successful balancing depends on the application method, not just the specific manipulation.", "venue": "CHI", "year": 2015, "title": "Now You Can Compete With Anyone: Balancing Players of Different Skill Levels in a First-Person Shooter Game"} +{"pid": "bad9d04d-936d-4f09-96f1-da9d6fea3e2f", "context": "Many people with dementia residing in long-term care cannot access experiences beyond their physical premises due to various restrictions such as location, mobility constraints, and legal mental health act restrictions. Recent research focuses on non-pharmacological interventions to improve the quality of life for these individuals.", "key_idea": "The authors proposed the use of Virtual Reality (VR) as a tool to provide 360\u00b0-video based experiences for individuals with moderate to severe dementia in a locked psychiatric hospital.", "method": "The paper explores VR as an intervention by observing the impacts of these interactions on individuals with dementia in a locked psychiatric hospital.", "outcome": "The paper discusses the appeal of Virtual Reality for people with dementia and the impact such interactions have had on these individuals.", "future_impact": "The paper identifies opportunities, pitfalls, and recommendations for future use of VR in healthcare services as a virtual alternative for experiences that are otherwise inaccessible for people with dementia in locked settings.", "venue": "CHI", "year": 2019, "title": "Bring the Outside In: Providing Accessible Experiences Through VR for People with Dementia in Locked Psychiatric Hospitals"} +{"pid": "5fa88ec0-08ae-452a-881d-f5181167067b", "context": "Ocean salinity is critically important for understanding climate change. However, analyzing it presents a unique visualization challenge due to the importance of data found in narrow data ranges, which varies with global location. Moreover, current editing tools do not easily enable a scientist to explore the subtleties of salinity and changes in salinity values.", "key_idea": "This paper proposes an interactive colormap workflow and tool that allows scientists to apply sophisticated colormaps to scalar data, thereby enabling more detailed exploration of ocean salinity.", "method": "The paper presents a novel workflow, which is enabled by an interactive colormap tool, for interactively applying sophisticated colormaps to scalar data.", "outcome": "The paper demonstrates an intuitive and immediate interaction of the scientist with the data, as enabled by the interactive colormap workflow and tool.", "future_impact": "The more refined salinity visualization enabled by the proposed solution can be expected to help in the improved understanding of ocean currents and climate patterns in the future.", "venue": "CHI", "year": 2016, "title": "Interactive Colormapping: Enabling Multiple Data Range and Detailed Views of Ocean Salinity"} +{"pid": "60a4e58991e011f4291ce424", "context": "15% of the world population with disabilities relies on mobile apps, underscoring the need for accessible software. Existing automated accessibility assessment techniques primarily detect violations of predefined guidelines and produce a large amount of accessibility warnings. However, these techniques often overlook how software is actually used by users with disabilities.", "key_idea": "The paper presents a novel accessibility testing approach for Android apps called Latte. Unlike earlier methods, Latte reuses tests written for app functionality to evaluate its accessibility, while taking into account the use cases and how they are executed by disabled users using assistive services.", "method": "Latte extracts use cases corresponding to each test and executes them as a disabled user would. Its effectiveness was evaluated on real-world Android apps.", "outcome": "The empirical evaluation showed that Latte is effective in detecting more valuable defects compared to prior techniques.", "future_impact": "The introduction of Latte could improve the accessibility of Android applications, leading to more inclusive software for users with disabilities. It may inspire future research into use-case based and user-centric automated testing frameworks.", "venue": "CHI", "year": 2021, "title": "Latte: Use-Case and Assistive-Service Driven Automated Accessibility Testing Framework for Android"} +{"pid": "60a7892691e0110affd71cbc", "context": "The development of an interface for remote vital signs self-monitoring at a large-scale isolation facility for COVID-positive patients presented several challenges: lack of time, lack of access to end users, changing requirements, high risk of infection and supply chain limitations.", "key_idea": "The paper utilizes five commonly prescribed activities of user experience design - Empathise, Define, Ideate, Prototype and Test - to explore how user experience designers can navigate similar challenges.", "method": "The authors present a case study, where these activities were applied in the design process under disrupted conditions.", "outcome": "The authors describe how their work was affected by the disrupted conditions and draw insights from their experience in designing the interface for remote vital signs self-monitoring.", "future_impact": "The authors recommend a focus on Empathy, Prototyping and Communication for user experience practitioners and educators, whose training may be needed in similarly mission-critical, time-constrained circumstances.", "venue": "CHI", "year": 2021, "title": "A Case Study of User Experience Design in a Disrupted Context: Design and Development of a Vital Signs Self-monitoring System"} +{"pid": "8a2f2c4d-4672-4500-bcb2-6e6d172d41b8", "context": "The existing understanding of collaborative information visualization is insufficient, particularly how different aspects of system design and group processes impact the efficacy of collaborative problem solving.", "key_idea": "The study examined how the level of options for visualizations and the nature of collaboration (i.e., group versus individual effort) affect the speed and accuracy of problem-solving", "method": "An empirical study was conducted in which the effects of different visualization options and the nature of collaboration on problem-solving speed and accuracy were observed and measured.", "outcome": "Evidence showed that systems with fewer options for visualizations yielded faster and more correct responses. Group collaboration made problem-solving more accurate but slower. It was also observed that the benefits of collaboration mainly came from validating results, rather than from planning and system use.", "future_impact": "The findings of this study suggest that designing tools to help translate and confirm visualizations could greatly enhance the benefits of collaborative information visualization.", "venue": "CHI", "year": 2003, "title": "Are there benefits in seeing double?: a study of collaborative information visualization"} +{"pid": "adbf4cb9-af07-42f0-8e45-09f69e062afa", "context": "Stroke patients often have limited access to rehabilitation after discharge from hospital, forcing them to self-regulate their recovery. Previous research has indicated that several musical approaches can be used effectively in stroke rehabilitation.", "key_idea": "The study proposes using specialized digital musical interfaces for stroke rehabilitation.", "method": "The authors conducted workshops involving stroke patients who played music in groups and individually using specialized digital musical interfaces, and collected their feedback to refine prototypes and understand the potential benefits of music making.", "outcome": "93% of participants stated they thought that the music workshops were potentially beneficial for their rehabilitation.", "future_impact": "This study contributes to the field of HCI by exploring the role of computer-based systems in stroke rehabilitation, suggesting a potential area for future research and therapeutic techniques.", "venue": "CHI", "year": 2015, "title": "Can Specialised Electronic Musical Instruments Aid Stroke Rehabilitation"} +{"pid": "803f5c61-d46a-4d1b-a231-af9f54990ad5", "context": "Earlier versions of the pCubee, a personal perspective-corrected hand-held cubic display, did not include certain interaction mechanisms and hardware features.", "key_idea": "The paper presents an improved version of pCubee that includes a complete hardware redesign and novel interaction mechanisms such as physics simulation, directly-mapped stylus, and cross-screen multi-touch input.", "method": "The authors have redesigned the pCubee hardware and introduced novel interaction mechanisms, which are presumably tested with users.", "outcome": "Users can interact with virtual objects in the display through various methods, offering improved visual experience. The redesigned system is also lightweight and compact, making it easy for manipulation.", "future_impact": "Applications of the updated pCubee may include visualization for static or dynamic contents, 3D object manipulation, and tangible entertainment.", "venue": "CHI", "year": 2014, "title": "The new pCubee: multi-touch perspective-corrected cubic display"} +{"pid": "977a43df-c766-4f3f-9688-0efc9bb37ca8", "context": "Video display terminal (VDT) technology is increasingly used in office workplaces. However, research on the implications of VDT use for the well-being of office workers has shown conflicting results regarding the type, magnitude, and causes of adverse changes in health, comfort, and satisfaction. Studies up to date are limited by their methodological approaches, mainly focusing on the physical ergonomic demands associated with VDT use, while others emphasize on job content, psychosocial and related operational-organizational factors.", "key_idea": "The authors propose to undertake a systematic study considering both work and workplace design factors and their interactions in relation to VDT use, in order to understand and mitigate the stresses brought about by VDT use in office environments.", "method": "The abstract doesn't specify the details of the specific methods that the authors use in this study.", "outcome": "The abstract doesn't specify the outcomes or results of this study.", "future_impact": "The study may provide a more comprehensive understanding of VDT-use related stresses, which could inform better design and policies for workplace environments, although this is not explicitly stated in the abstract.", "venue": "CHI", "year": 1982, "title": "A systems analysis of stress-strain in VDT operation"} +{"pid": "d73d8c32-d867-49e3-9eed-f18728027dfd", "context": "Input by free motion of fingers is an emerging method, facilitated by advancements in hand tracking. It is expressive, potentially fast, and can be used across many settings as it does not require physical contact or visual feedback.", "key_idea": "The paper seeks to provide a detailed analysis of the performance and anatomical characteristics of finger motion to inform the design of high-performance input methods.", "method": "An experiment was performed using a commercially available sensor to scrutinize the speed, accuracy, individuation, movement ranges, and individual differences of each finger. The findings were applied to text entry by computationally optimizing multi-finger gestures in mid-air using a novel objective function that considers performance, anatomical factors, and learnability.", "outcome": "The experiment revealed differences of up to 50% in movement times and provided indices quantifying the individuation of single fingers. Optimization case studies showed entry rates of up to 22 words per minute.", "future_impact": "The authors anticipate that the insights from this study will inform the design of high-performance input methods, with considerations for performance, anatomical factors, and learnability.", "venue": "CHI", "year": 2015, "title": "Investigating the Dexterity of Multi-Finger Input for Mid-Air Text Entry"} +{"pid": "a85cb790-c494-4e21-b93d-dd7f1077c78d", "context": "Currently, most wearable computing devices are not conveniently accessible or require significant effort to interact with, often requiring the use of both hands.", "key_idea": "The authors propose a new type of wearable computing device in the form of an interactive identity badge. The badge has an embedded LCD to present information and a retractable string for single-handed interaction.", "method": "The authors developed a prototype of the interactive badge and explored various usage scenarios and interaction techniques.", "outcome": "The prototype interactive badge allows single-handed, lightweight, and immediate access to a variety of information.", "future_impact": "This research might encourage further exploration and development of wearable computing devices with single-handed interaction.", "venue": "CHI", "year": 2013, "title": "An interactive belt-worn badge with a retractable string-based input mechanism"} +{"pid": "17cf7f33-44fa-41b1-90e9-028b85561392", "context": "Scatterplots often use size, shape, and color as dimensions to encode multivariate datasets because these dimensions are considered separable. However, recent work shows that the size of marks significantly impacts color difference perceptions, leading to questions about the actual separability of these channels.", "key_idea": "This paper investigates the impact of mark shape, size, and color on data interpretation in multiclass scatterplots, providing insights into how these visual channels intersect in perception.", "method": "The authors conduct a series of crowdsourced experiments to measure how mark shape, size, and color influence data interpretation in multiclass scatterplots.", "outcome": "The experiments reveal that mark shape significantly influences color and size perception, and that separability among these channels functions asymmetrically, with shape having a stronger influence on size and color perceptions in scatterplots than size and color have on shape.", "future_impact": "The data and models resulting from this study can support designers in anticipating viewer perceptions, thereby enabling them to build more effective visualizations.", "venue": "CHI", "year": 2019, "title": "Measuring the Separability of Shape, Size, and Color in Scatterplots"} +{"pid": "60a4e58991e011f4291ce43d", "context": "Warm-up games are common in a variety of domains, but there is little scholarly research on their role in physical training.", "key_idea": "The study explores potential goals and benefits of warm-up games and the possibilities of including technology in them by investigating a collection of warm-up games.", "method": "The study involves investigating a series of warm-up games collected from online sources and physical education teachers, as well as design research as part of a technology-supported circus training course. They also conducted experimental design interventions, augmenting warm-up games with wearable technology.", "outcome": "The study uncovers primary aims and benefits of warm-up games, which involve preparing participants physically, socially, and mentally. It also shows how the inclusion of open-ended technology can support these goals.", "future_impact": "This research could lead to broader opportunities for technology inclusion in warm-up games.", "venue": "CHI", "year": 2021, "title": "Physical Warm-up Games: Exploring the Potential of Play and Technology Design"} +{"pid": "f6478da1-ad86-4d2f-a003-977275811b73", "context": "Current interfaces made with 2D paper substrates lack extensibility and usability, presenting challenges to those with specific requirements for input devices.", "key_idea": "The authors propose a DIY process for creating customized paper keyboards with kinesthetic feedback for touchscreens, combining kirigami structures for tactile sensation and double-layered circuits for adjustable input layouts.", "method": "To create a keyboard with appropriate stroke and reaction force, the authors adopted the Rotational Erection System (RES). The designed keys are then realized using kirigami techniques and printable double-layered circuits.", "outcome": "The paper describes the successful creation of a keyboard with kinesthetic sensations comparable to that of traditional keyboards, indicating the feasibility and effectiveness of the proposed design methodology.", "future_impact": "The authors anticipate that their proposed easy-to-customize keyboard can be beneficial for individuals with specific requirements for input devices, potentially improving the usability and adaptability of paper-based interfaces.", "venue": "CHI", "year": 2019, "title": "Kirigami Keyboard: Inkjet Printable Paper Interface with Kirigami Structure Presenting Kinesthetic Feedback"} +{"pid": "60a7892691e0110affd71d00", "context": "Spatial experience is an important subject in various fields, and in Human-Computer Interaction (HCI) it has been mostly investigated at the urban scale, focusing mostly on the personal meaning or aesthetic and embodied experiences in the space.", "key_idea": "This workshop looks to harness spatial experiences for building and interacting with technologies in everyday environments, particularly in smart cities, and seeks to understand human scale spatial experiences and its future implications.", "method": "The authors present a speculative design approach to collaboratively discover new ways to understand and capture human scale spatial experience with researchers and practitioners from various fields.", "outcome": "The outcome of the workshop was the development of a sketch of concrete solutions that could better capture critical features of human scale spaces, and open up unique possibilities for aspects such as urban play.", "future_impact": "The authors aim that the outcome of this work will contribute a road map for future HCI research on human scale spatial experience and its application.", "venue": "CHI", "year": 2021, "title": "2VT: Visions, Technologies, and Visions of Technologies for Understanding Human Scale Spaces"} +{"pid": "3aa2ee25-b4b0-4a48-a3cf-cdad36c36c31", "context": "Charities are subject to stringent transparency and accountability requirements from government and funders to ensure that they are conducting work and spending money appropriately. Charities have unique characteristics that make them different from regular organizations.", "key_idea": "The study investigates how the transparency obligations of a charity manifest through work and how the workers of a charity reason about transparency and accountability as an everyday practice.", "method": "The authors conducted ethnographic fieldwork and workshops at a charity over a seven month period to understand the transparency and accountability practices within the organization.", "outcome": "Findings indicated that charities engage in presenting different accounts of their work; workers view their legal transparency obligations differently from their accountability to their everyday community; and how their labour does not translate well to outcome measures or metrics.", "future_impact": "The study has implications for the design of future systems that would aid organisations in producing accounts of their work as part of their everyday practice.", "venue": "CHI", "year": 2018, "title": "Accountability Work: Examining the Values, Technologies and Work Practices that Facilitate Transparency in Charities"} +{"pid": "2021220c-e672-4ac2-9c9c-d79c729f0235", "context": "Current touch screen strategies have limitations in speed, accuracy, and user satisfaction, particularly when dealing with small, densely-packed targets.", "key_idea": "This study investigates the merits of more complex touch strategies that utilize continuous streams of touch data instead of the simpler, prevalent strategies.", "method": "The researchers conducted an experiment comparing the speed, accuracy, and user satisfaction of three different touch screen strategies.", "outcome": "The results of the experiment showed that a touch strategy that provides continuous feedback until a selection is confirmed results in fewer errors compared to other strategies.", "future_impact": "The study's results may influence the design of future touch screens, particularly those containing small, densely-packed targets.", "venue": "CHI", "year": 1988, "title": "Improving the accuracy of touch screens: an experimental evaluation of three strategies"} +{"pid": "37588b9b-6243-4f2b-ab20-4a6c39740508", "context": "Content on computer screens is often inaccessible to users because it is hidden, occluded by other windows, outside the viewport, or overlooked. Current software only provides limited support to visualize hidden content and rarely supports search synchronization crossing application boundaries.", "key_idea": "The authors introduce two novel visualization methods to guide users to hidden content: a method that generates awareness for occluded or out-of-viewport content using see-through visualization, and another that shows off-screen indicators and an on-demand smart preview for content that is either outside the screen's viewport or for data sources not opened at all.", "method": "The authors validate their methods in a user study, where they measure the speed of localization of hidden content compared to traditional search functionality.", "outcome": "The user study demonstrated that the proposed visualization techniques enable a faster localization of hidden content compared to traditional search functionality.", "future_impact": "The proposed techniques can assist users in information retrieval tasks, potentially improving efficiency and reducing the chances of overlooking important content.", "venue": "CHI", "year": 2014, "title": "Show me the invisible: visualizing hidden content"} +{"pid": "623d5d0c6750f864fe44bd29", "context": "Immersive interactive technologies like VR have been used to evoke positive emotions through preset settings of light, color, and scenery, however, the potential for users to independently create their own virtual environments (VE) for emotional engagement and well-being hasn't been sufficiently addressed.", "key_idea": "The paper presents Mood Worlds, a VR application that allows users to visualize their emotions by self-creating a VE.", "method": "The authors conduct an exploratory evaluation on Mood Worlds with 16 participants to measure its impact on emotional engagement, positive emotions, and well-being.", "outcome": "The exploratory evaluation concluded that Mood Worlds is an effective tool supporting emotional engagement. The study also found that an autonomous creation process in VR increases positive emotions and well-being.", "future_impact": "The study opens up opportunities and design requirements in using VR as positive technology for well-being.", "venue": "CHI", "year": 2022, "title": "Mood Worlds: A Virtual Environment for Autonomous Emotional Expression"} +{"pid": "07d48278-b48c-4ff2-b210-7292587f5a66", "context": "Usability evaluation for web applications on mobile devices is an ongoing research area and various methods like Heuristic Evaluation, Cognitive Walkthrough, and Web Design Perspectives-based Usability Evaluation (WDP) are in use.", "key_idea": "The authors conducted a comparative study of three different methods (Heuristic Evaluation, Cognitive Walkthrough, and WDP) for usability evaluation in Web mobile applications to ascertain the best one.", "method": "The authors carried out this study as a class activity involving students of Human-Computer Interaction (HCI).", "outcome": "The results obtained from this study are critical for the development of a specific technology targeting cost-effective usability evaluation of mobile applications.", "future_impact": "This research could lead to the creation of a specific technology focusing on the most cost-effective and efficient usability evaluation of mobile applications.", "venue": "CHI", "year": 2010, "title": "Aplicando t\u00e9cnicas de inspe\u00e7\u00e3o de usabilidade para avaliar aplica\u00e7\u00f5es m\u00f3veis"} +{"pid": "a4fd60b7-af7c-4511-bed6-2172d7c41360", "context": "The issue of accurately and precisely triggering input at a specific time point has been largely undiscussed in gaze interaction research. The 'Midas touch' problem, the accidental triggering of input when looking at an interface, is prevalent.", "key_idea": "The authors propose the exploration of voluntary eye convergences as a novel interaction technique for accurately and precisely timing gaze input, solving the 'Midas touch' problem.", "method": "A novel clock paradigm is introduced to study input timing and to compare voluntary convergences with common gaze dwelling.", "outcome": "It was found that voluntary convergences are significantly more accurate and precise in timing gaze inputs compared to frequent gaze dwelling techniques.", "future_impact": "The results suggest that voluntary convergences are well-suited for applications in which the timing of user input is critical, hence complementing existing gaze techniques focused on speed and spatial precision.", "venue": "CHI", "year": 2016, "title": "On the Verge: Voluntary Convergences for Accurate and Precise Timing of Gaze Input"} +{"pid": "80d3eca9-e645-41f2-8d40-32358154d996", "context": "Despite organization schemes, many people still feel overwhelmed when it comes to family coordination. Also, coordination is not typically done through the family calendar; rather it is a tool that provides an awareness of activities and changes that in turn enables coordination.", "key_idea": "To help overcome this overwhelm, the authors propose LINC, an inkable family calendar designed for the kitchen, which should provide tools that enable families to use their own coordination routines.", "method": "The authors used a participatory design process involving interviews, paper prototyping, and a formative evaluation to develop LINC.", "outcome": "The authors proposed key implications for digital family calendars and family coordination systems based on the research on LINC.", "future_impact": "These implications can potentially influence the design and function of digital family calendars and family coordination systems further.", "venue": "CHI", "year": 2006, "title": "\"LINC-ing\" the family: the participatory design of an inkable family calendar"} +{"pid": "4e984a5b-21ff-4eba-87cb-7b2df619fbdb", "context": "The context of the paper is to understand how the location of the image display affects the performance of an endoscopic surgery task.", "key_idea": "The key idea was to test the performance of an endoscopic surgery task when the image is displayed on a monitor in front of the subject versus when it's projected directly above the workspace.", "method": "Two tasks were performed under two display conditions and the timing results were analyzed. An endoscopic camera view of the workspace was displayed on a monitor in front of the subject or projected directly above the workspace.", "outcome": "Performance was faster when the superimposed display was used first and image quality was perceived to be superior on the monitor. Also, the subjects' ability to calibrate the display space with the workspace is discussed.", "future_impact": "The results of the experiment have implications for the design and arrangement of surgical operating theatres.", "venue": "CHI", "year": 1999, "title": "Superimposing display space on workspace in the context of endoscopic surgery"} +{"pid": "2106f524-b243-4440-8cc9-8ca8e9c841f8", "context": "Earlier research showed that females are less likely to use unfamiliar software features in end-user debugging environments, which pose a serious problem as these features may be key in debugging problems. In contrast, males are known for tinkering in unfamiliar environments.", "key_idea": "The authors are interested in exploring whether encouraging tinkering with new features would help females overcome factors such as low self-efficacy that have previously been identified as hindrance to their use of unfamiliar software features.", "method": "The authors conducted an experiment with males and females in an end-user debugging setting, examining how tinkering behavior impacts several measures of their debugging success.", "outcome": "The study found that the factors of tinkering, reflection, and self-efficacy can combine in multiple ways to impact debugging effectiveness differently for males and females.", "future_impact": "This work could lead to a better understanding of how gender impacts the use of debugging tools and influence the design of future tools that are inclusive and accommodating of different troubleshooting strategies.", "venue": "CHI", "year": 2006, "title": "Tinkering and gender in end-user programmers' debugging"} +{"pid": "af92be11-9f74-4f2e-98cb-76d40388dece", "context": "Several Software Development Kits (SDKs) exist for the development of Augmented Reality (AR) applications, and a qualitative analysis is needed to determine their suitability for augmented educational settings.", "key_idea": "The paper focuses on evaluating these main SDKs for the development of marker-based AR software to use in-class teaching situations. This evaluation is based on a case study in a public university in Brazil.", "method": "The SDKs were evaluated based on the accuracy of their ability to detect markers with different characteristics.", "outcome": "The study concludes that Vuforia and EasyAR are the superior choices among the evaluated SDKs.", "future_impact": "The conclusion could potentially guide educators and software developers in making informed decisions when choosing an SDK for developing AR applications aimed at improving classroom teaching methodologies.", "venue": "CHI", "year": 2019, "title": "Evaluation of augmented reality SDKs for classroom teaching"} +{"pid": "340a413f-aa82-4c93-9e45-a62cd1b7bb82", "context": "The extensive and complex process of government budgeting hinders taxpayers' understanding of budget allocation and prevents their participation in public discussions.", "key_idea": "This study introduces 'Factful', a web-based annotative article reading interface that enhances news articles with fact-checking support and provides data-driven context of budgetary information leveraging open government data.", "method": "Factful's effectiveness was evaluated through a lab study where participants used the interface to engage in discussions around budgetary issues.", "outcome": "Participants using Factful discussed more critically and made more fact-based supporting statements, building a rich context around budget facts beyond what was initially presented in the article.", "future_impact": "Factful presents a potential model for facilitating informed, fact-oriented discussions on budgetary issues online, by leveraging open government data.", "venue": "CHI", "year": 2015, "title": "Factful: Engaging Taxpayers in the Public Discussion of a Government Budget"} +{"pid": "638eb2ef90e50fcafd58b283", "context": "In sequence labeling (SL) tasks, handling long-tail cases is complex and the conventional models are inadequate confronted with these situations.", "key_idea": "This paper introduces a new approach called graph neural networks sequence labeling (GNN-SL) which improves the vanilla SL model by augmenting the output with similar tagging examples retrieved from the entire training set.", "method": "The authors construct a heterogeneous graph and use graph neural networks (GNNs) to transfer information between the retrieved tagging examples and the input word sequence. A variety of experiments on Named Entity Recognition (NER), Part of Speech Tagging (POS), and Chinese Word Segmentation (CWS) tasks are conducted to assess the performance of GNN-SL.", "outcome": "GNN-SL achieved state-of-the-art results on the CWS task with scores of 96.9 on PKU, 98.3 on CITYU, 98.5 on MSR, and 96.9 on AS. Moreover, it achieves results comparable to state-of-the-art performances on NER and POS datasets.", "future_impact": "This presented approach can be a new way to improve predictions quality in various sequence labeling tasks.", "venue": "ACL", "year": 2022, "title": "GNN-SL: Sequence Labeling Based on Nearest Examples via GNN"} +{"pid": "64ae668b3fda6d7f068448dd", "context": "Extracting Adverse Drug Events (ADEs) and related suspect drugs from unstructured data sources like clinical study reports, patient health records, social media posts, etc., is a complex task due to linguistic relations and unavailability of large corpus of labeled datasets.", "key_idea": "The paper introduces ADEQA, a question-answer-based approach that uses quasi supervised labeled data and sequence-to-sequence transformers for extracting ADEs, drug suspects, and relationships between them.", "method": "The proposed model (ADEQA) was applied to a public ADE corpus to establish relationships between ADEs and suspect drugs.", "outcome": "With the introduction of the ADEQA model, state-of-the-art results were achieved, as evidenced by an F1 score of 94% on establishing relationships between ADEs and their respective suspects.", "future_impact": "The use of natural language generation (NLG) based models like ADEQA can reduce the need for extensive token level labeling and significantly lower the adoption barrier.", "venue": "ACL", "year": 2023, "title": "ADEQA: A Question Answer based approach for joint ADE-Suspect Extraction using Sequence-To-Sequence Transformers"} +{"pid": "649a5e2ad68f896efad84927", "context": "Hedging is a conversational strategy but its automated generation has not been successfully addressed in previous work. Most notably, this strategy is often found in contexts of education, such as tutoring.", "key_idea": "The authors develop a model for generating hedges in conversation, which involves fine-tuning language models trained on tutoring data, and then reranking to select the candidate that best matches the expected hedging strategy using a hedge classifier.", "method": "The authors apply their approach to a natural peer-tutoring corpus that consists of disfluencies, repetitions, and repairs and conduct an error analysis for both their approaches.", "outcome": "The results show that generation in an environment with many disfluencies, repetitions, and repairs is feasible with the reranking method.", "future_impact": "The error analysis reveals challenges of balancing between social and task-oriented goals in conversation, which could inform future design of conversational systems attempting to accomplish mixed goals in conversation.", "venue": "ACL", "year": 2023, "title": "How About Kind of Generating Hedges using End-to-End Neural Models?"} +{"pid": "46f8948e-de3c-4210-9b31-81aac5a04d87", "context": "Acquisition of prosody is essential for language learners apart from vocabulary and grammar. However, it received less attention in instruction. There is a need for a solution that could enable automatic identification and feedback on learners' prosodic errors.", "key_idea": "The authors proposed an investigation into automatic pitch accent labeling for non-native speech.", "method": "The authors used an acoustic-based context model for binary pitch accent recognition. The model was trained on within-group data and was also tested in cross-group training.", "outcome": "The acoustic-based context model achieved accuracies over 79% on binary pitch accent recognition when trained on within-group data. In cross-group training, where native and near-native training data was used, it showed no significant loss of accuracy on non-native test speech.", "future_impact": "The findings illustrate the potential for automatic feedback in computer-assisted prosody learning.", "venue": "ACL", "year": 2009, "title": "Investigating Pitch Accent Recognition in Non-native Speech"} +{"pid": "62393e7e5aee126c0f125f03", "context": "Traditional propositional analysis focuses on extracting belief-driven sentiment, known as stances. However, converting it to extract concerns, which are topical issues with moral dimensions/endorsements, has not been systematized for adaptation across tasks and domains.", "key_idea": "The authors developed a generalized paradigm that adapts propositional analysis to new tasks and domains. By drawing an analogy between stances and concerns, the authors propose an approach that combines semi-automatic resource building for domain-dependent concern types extraction and an entirely automatic procedure for domain-independent moral dimensions and endorsement values extraction.", "method": "The authors used semi-automatic resource building and automatically extracted domain-independent moral dimensions and endorsement values. Then, they selected terms from propositional structures for lexical expansion based on semantic similarity, creating new moral dimension lexicons. The performance was evaluated using a ground truth developed based on expert annotators.", "outcome": "The proposed method improved recall by 231% compared to the baseline while only losing 10% in precision. The F1 measure improved by 66% over the baseline and achieved 97.8% of human performance.", "future_impact": "The lexically based approach provides significant savings over human labor and model-building methods, offering a cost-effective alternative. The authors have made the expanded moral dimension/value lexicon, annotation guidelines, and ground truth available to the community for future research.", "venue": "ACL", "year": 2022, "title": "From Stance to Concern: Adaptation of Propositional Analysis to New Tasks and Domains."} +{"pid": "fca32ab0-f3c2-42d3-854b-2b9b0c1aa82f", "context": "Large scale information retrieval systems need a mechanism to refine the flood of incoming documents that are subjected to detailed analysis, focusing on those that are of critical relevance for the users.", "key_idea": "The paper presents the idea of using semantic codes from a machine readable dictionary, called Subject Field Codes (SFCs), to filter sizable sets of documents for their overall thematic appropriateness to a specific topic of interest.", "method": "Each word in a document is tagged with the appropriate, disambiguated SFC, producing a summary-level semantic representation of the document's contents. The SFCs in the document are normalized into a vector representing that document's contents. Queries are similarly turned into SFC vectors. The similarity between the SFC vectors of incoming documents and the query's SFC vector is calculated, and documents with a certain degree of similarity to the query vector are then flag for further detailed analysis.", "outcome": "The automatic and efficient method of SFC assignment was implemented and tested. The empirical testing has showed that using the Subject Field Coder is a reasonable way of classifying documents from a wide array of incoming documents.", "future_impact": "The research points to the potential of using SFCs for enhancing document filtering in large-scale information retrieval systems, which could pave the way for more refined document representation and matching techniques.", "venue": "ACL", "year": 1993, "title": "Document Filtering Using Semantic Information from a Machine Readable Dictionary"} +{"pid": "8a33a530-b212-4a3f-b968-ef0cdadea2bb", "context": "Existing technologies do not infer the underlying semantic and pragmatic structure of documents. There is a lack of applications to render documents versatile and intelligent contents.", "key_idea": "The authors propose the Global Document Annotation (GDA) project which uses a tag set to automatically infer the structure of documents. The paper discusses an automatic text summarization based on GDA that features a domain/style-free algorithm and personalization, reflecting the interests and preferences of readers.", "method": "The method calculates the importance score of a text element using spreading activation on an intradocument network connecting text elements via thematic, rhetorical, and coreferential relations.", "outcome": "The proposed method is capable of dynamically generating summaries of various sizes and includes a summary browser that supports personalization.", "future_impact": "If WWW users were to tag their documents as part of content authoring using GDA, it could promote the development and spread of NLP/AI applications.", "venue": "ACL", "year": 1998, "title": "Automatic Text Summarization Based on the Global Document Annotation"} +{"pid": "dcff1cd6-476e-4946-982b-d308d45a971d", "context": "CFG (Context-Free Grammar) filtering techniques have been applied to LTAG (Lexicalized Tree Adjoining Grammar) and HPSG (Head-Driven Phrase Structure Grammar) systems but there seems to be no comparison of their efficiencies in the literature.", "key_idea": "An empirical comparison between CFG filtering techniques for LTAG and HPSG is conducted, with the objective of demonstrating that an approximation of HPSG produces a more effective CFG filter than that of LTAG.", "method": "A demonstration is carried out comparing both CFG filters, with an investigation conducted to uncover the reason behind any differences observed.", "outcome": "It is found that an approximation of HPSG produces a more effective CFG filter than that of LTAG", "future_impact": "The findings of the study could lead to improved optimization of CFG filters in LTAG and HPSG-based systems in the future.", "venue": "ACL", "year": 2003, "title": "Comparison between CFG Filtering Techniques for LTAG and HPSG"} +{"pid": "64a78f1fd68f896efa01e9d2", "context": "The current state of natural language inference models consists of advancements in distributed representation and neural networks but their reliance on real reasoning versus spurious correlations is unclear. Adversarial attacks have been used to expose weaknesses in these models.", "key_idea": "The authors explore a fundamental problem of developing attack models based on logic formalism and propose NatLogAttack, a framework for systematically attacking natural language inference models using natural logic, a classical logic formalism developed for natural language inference.", "method": "NatLogAttack is designed to perform label-preserving and label-flipping attacks. It is compared with existing attack models.", "outcome": "The authors found that the proposed NatLogAttack generates superior adversarial examples with fewer visits to the victim models. Victim models are found more vulnerable under label-flipping setting.", "future_impact": "NatLogAttack provides a tool to probe the existing and future natural language inference models' capacity from a new perspective, and prompts further exploration of logic-based attacks for understanding the desired property of reasoning.", "venue": "ACL", "year": 2023, "title": "NatLogAttack: A Framework for Attacking Natural Language Inference Models with Natural Logic"} +{"pid": "80d2d9c8-1af6-402f-9a55-df2b3b79861f", "context": "Abstractive text summarization of news requires a way of representing events, yet the effective ways of extracting event patterns for this task have not been fully explored.", "key_idea": "This paper introduces three event pattern extraction methods applicable to abstractive summarization: heuristics-based, compression-based, and memory-based methods. The memory-based method stands out as it allows generating more grammatical and informative sentences.", "method": "Three different event pattern extraction techniques are compared: heuristics-based, compression-based, and memory-based, while a new data structure and a search method are introduced to cope with the vast search space required by the memory-based approach.", "outcome": "The memory-based method enables significantly more grammatical and informative sentence generation for abstractive summarization, compared to the other methods, albeit at the expense of computational resources due to the significant requirement of search space.", "future_impact": "Adoption of the memory-based method in abstractive summarization could facilitate generation of more informative and grammatically correct sentences, potentailly leading to new advancements in the field.", "venue": "ACL", "year": 2014, "title": "Modelling Events through Memory-based, Open-IE Patterns for Abstractive Summarization"} +{"pid": "b8b2b9ff-7a23-4b43-a2e6-719ef6058875", "context": "Before this study, no open sentiment analysis evaluations had been conducted in Russian, and the state of the art in sentiment analysis in Russian was undefined.", "key_idea": "The researchers have conducted the first open sentiment analysis evaluations in Russian as part of the annual TREC-like ROMIP competition.", "method": "Various test and train collections for tasks, such as sentiment classification in blogs and newswire, opinion retrieval, were used. The evaluation used track tasks and evaluation metrics.", "outcome": "The study resulted in the creation of several test and train collections for tasks like sentiment classification in blogs and newswire, and opinion retrieval.", "future_impact": "This initiative is projected to contribute significantly to the advancement of sentiment analysis evaluations in the Russian language.", "venue": "ACL", "year": 2013, "title": "Evaluating Sentiment Analysis Systems in Russian"} +{"pid": "575a89d3-52ed-4dbe-bca9-608abfe2344f", "context": "The extraction of sentiments expressed towards named entities is a significant challenge, especially when both accuracy and speed of extraction and classification are of high importance.", "key_idea": "The authors propose a system that uses a combination of rule-based classification, statistics, and machine learning to extract and classify opinions expressed towards named entities.", "method": "The authors employ a service-oriented architecture that provides a flexible interface for the end-user and allows the development of applications with a wide range of use-cases. The experiment measures extraction results and sentiment calculus in one version, while the other version validates internal metrics.", "outcome": "The paper presents preliminary results of the system but does not provide specific details about its performance and effectiveness.", "future_impact": "This work may enable the creation of applications that can be used for tasks like aggregating consumer feedback on commercial products or measuring public opinion on political issues using blog and forum data.", "venue": "ACL", "year": 2011, "title": "Sentimatrix -- Multilingual Sentiment Analysis Service"} +{"pid": "64563889d68f896efacf53be", "context": "Current methods of adapting large language models (LLMs) for the task of radiology report summarization (RRS) can be expensive and inefficient due to the need for extensive fine-tuning.", "key_idea": "The paper proposes lightweight strategies for domain adaptation of LLMs specifically for RRS, with approaches including pretraining and prompting variations, and parameter-efficient fine-tuning methods such as prefix tuning and LoRA.", "method": "The authors apply and evaluate various methods of domain adaptation for RRS on the MIMIC-III dataset and conduct a radiologist reader study and qualitative analysis.", "outcome": "The best performance on the MIMIC-III dataset was achieved through pretraining on clinical text and parameter-efficient fine-tuning, which only fine-tuned 0.32% of the model's parameters, standing in contrast to the inefficiency of end-to-end fine-tuning. The authors also reported findings from the reader study and qualitative analysis.", "future_impact": "These findings highlight the significance of domain adaptation in radiology report summarizations, offering key insights for the development of more effective natural language processing solutions for clinical tasks.", "venue": "ACL", "year": 2023, "title": "RadAdapt: Radiology Report Summarization via Lightweight Domain Adaptation of Large Language Models"} +{"pid": "0cfc4102-3a0f-414f-994a-bc4b14294b59", "context": "The ACL2 theorem prover is a complex system with large libraries which can be extended by hundreds of thousands of lines of additional modeling tools, specifications, and proof scripts in industrial verification efforts. High quality, current documentation is crucial for teams working on such expansive projects.", "key_idea": "The authors introduce XDOC, a scalable documentation tool for ACL2 that incorporates the documentation of ACL2 itself, the community books, and an organization\u2019s internal formal verification projects.", "method": "They applied this tool in generating a comprehensive, publicly available ACL2+Books Manual for all ACL2 users, and an extended manual to cover Centaur Technology's internal books.", "outcome": "XDOC was used to produce the ACL2+Books Manual, which brings improved documentation to all ACL2 users. It also was employed to generate an extended manual for internal use at Centaur Technology.", "future_impact": "Other organizations using ACL2 are expected to adopt the method to develop extensive manuals that cater to their specific needs.", "venue": "ACL", "year": 2014, "title": "Industrial-strength documentation for ACL2"} +{"pid": "6232a74d5aee126c0fe13f03", "context": "Mainstream machine learning paradigms for NLP often presume that the target task is predefined and static with supervision mainly coming from a set of labeled examples. These presumptions raise the question of how to build a system capable of continually learning new tasks from instructions.", "key_idea": "This work introduces a new learning paradigm, ConTinTin (Continual Learning from Task Instructions), where a system should sequentially learn new tasks from their textual instructions. This involves generating expected outputs from instruction, transferring knowledge from previous tasks to assist future tasks, and retaining or improving performance on prior tasks after learning new tasks.", "method": "The approach, named InstructionSpeak, uses two strategies - learning from negative outputs and revisiting instructions from previous tasks - to improve both forward-transfer and backward-transfer. The method is tested on a stream of over 60 tasks, each with an instruction.", "outcome": "The study's outcome is not explicitly stated in the abstract, although the new learning problem and approach (ConTinTin paradigm and InstructionSpeak method) are introduced and applied to over 60 tasks.", "future_impact": "This work contributes to a better understanding of this new learning problem and provides analyses for the community to explore continual learning from instructions, thus potentially impacting and broadening the NLP learning paradigms.", "venue": "ACL", "year": 2022, "title": "ConTinTin: Continual Learning from Task Instructions"} +{"pid": "60badf1591e01102e59b6afc", "context": "Existing recognizing textual entailment (RTE) datasets do not focus on figurative language, and current state-of-the-art RTE models might not sufficiently capture this aspect of language.", "key_idea": "The authors develop a new collection of RTE datasets that focus specifically on figurative language, including simile, metaphor, and irony.", "method": "The authors evaluate state-of-the-art models trained on popular RTE datasets using these newly created figurative language RTE datasets.", "outcome": "The evaluation results indicate that state-of-the-art RTE models struggle to perform pragmatic inference and reasoning about world knowledge when dealing with figurative language.", "future_impact": "The authors anticipate that their datasets will provide a challenging testbed for evaluating and improving future RTE models.", "venue": "ACL", "year": 2021, "title": "Figurative Language in Recognizing Textual Entailment."} +{"pid": "e73698fc-6cc0-4641-b75a-77ee156b7bf7", "context": "In The Preposition Project (TPP), 13 prepositions have now been analyzed and considerable data made available.", "key_idea": "By analyzing the coverage of these senses in The Preposition Project, the authors propose specific inheritance mechanisms within the preposition sense inventory which provides a potentially greater breadth and depth than other inventories of the range of semantic roles.", "method": "The authors analyze the coverage of senses in The Preposition Project. These prepositions, among the most common words in English, contain 211 senses.", "outcome": "The authors' analysis shows the viability of the proposed inheritance mechanisms and their ability to provide a basis for the rationalization of the range of preposition meaning. The rationalization can be used for developing a data-driven mapping of a semantic role hierarchy.", "future_impact": "The findings and methodology can be used to outline the broad structure of a WordNet-like representation of preposition meaning, complete with self-contained disambiguation tests.", "venue": "ACL", "year": 2006, "title": "Coverage and inheritance in the preposition project"} +{"pid": "648fd298d68f896efa163a9c", "context": "High-quality data is essential for conversational recommendation systems. Existing works have limitations as they require heavy human efforts for manually labeling or designing and extending recommender dialogue templates, they suffer from limited scale, and can contain uninformative or inappropriate recommendations due to the limited experience and knowledge of annotators.", "key_idea": "This paper presents AUGUST, a novel automatic dataset synthesis approach that can generate both large-scale and high-quality recommendation dialogues through a data2text generation process, where unstructured recommendation conversations are generated from structured graphs based on real-world user-item information.", "method": "The authors generate recommendation conversations using rich personalized user profiles from traditional recommendation datasets, external knowledge from knowledge graphs, and conversation skills contained in human-to-human conversational recommendation datasets. They validate the synthesized data with extensive experiments, particularly under low-resource scenarios.", "outcome": "Experiment results validate the benefit of automatically synthesized data under low-resource scenarios, demonstrating the promising potential to develop more effective conversational recommendation systems.", "future_impact": "The authors anticipate that their novel data synthesis approach could facilitate the development of more effective conversational recommendation systems.", "venue": "ACL", "year": 2023, "title": "AUGUST: an Automatic Generation Understudy for Synthesizing Conversational Recommendation Datasets"} +{"pid": "63a2794d90e50fcafd2945e1", "context": "Previous work has shown that pretrained Masked Language Models (MLMs) can be expanded to new languages by learning a new set of embeddings, keeping the transformer body frozen. Nonetheless, despite learning a small subset of parameters, this approach is not compute-efficient.", "key_idea": "The authors propose mini-model adaptation as a compute-efficient alternative. This approach builds a shallow mini-model from a fraction of a large model's parameters, and then trains new language-specific embeddings over the mini-model for rapid cross-lingual transfer.", "method": "The authors explore two approaches, MiniJoint and MiniPost, to learn mini-models and conduct experiments on XNLI, MLQA and PAWS-X.", "outcome": "The mini-model adaptation strategy matches the performance of the standard approach whilst using up to 2.4 times less compute.", "future_impact": "The proposed mini-model adaptation can make the process of extending pretrained models to new languages more efficient, which could lead to improvements in cross-lingual transfer tasks.", "venue": "ACL", "year": 2022, "title": "Mini-Model Adaptation: Efficiently Extending Pretrained Models to New Languages via Aligned Shallow Training"} +{"pid": "3cdb81c5-f846-4a0f-9977-c38845739f71", "context": "Prior to this study, there was a need for a flexible tool to experiment with and develop solutions for the Textual Entailment problem.", "key_idea": "The authors propose a general-purpose open source package for recognizing Textual Entailment that is configurable and allows for quick solution prototyping.", "method": "The authors develop an open-source package implementing a collection of algorithms and a modular architecture. This facilitates setting up an experimental working environment for the Textual Entailment problem.", "outcome": "The new open-source package allows for fast prototyping of solutions and configuration of a working environment for RTE tasks.", "future_impact": "This tool can serve as a valuable resource for tackling the Textual Entailment problem, it can be used for didactic purposes and has the potential to foster a collaborative environment to promote research in the field.", "venue": "ACL", "year": 2010, "title": "An Open-Source Package for Recognizing Textual Entailment"} +{"pid": "95998615-32d5-4fce-929c-c7977853f33e", "context": "Preordering of a source language sentence to match target word order has proved to be useful for improving machine translation systems. However, when training reordering models on larger corpora of sentence-aligned data, manual word alignments are often not available, and machine-generated alignments can be noisy.", "key_idea": "This paper proposes a novel approach that uses information from a reordering model to improve reorderings given noisy machine alignments, and also improve word alignments, effectively mitigating the effect of noisy machine alignments.", "method": "The created method is applied to a larger corpus of sentence-aligned data. It generates alignments and trains a reordering model on the NIST MT-08 Urdu-English evaluation set.", "outcome": "The proposed approach generates alignments that are 2.6 f-Measure points better than a baseline supervised aligner, and trains a reordering model that gives an improvement of 1.8 BLEU points over a reordering model that only uses manual word alignments, and a gain of 5.2 BLEU points over a standard phrase-based baseline.", "future_impact": "This work offers a strategy for dealing with the challenge of noisy machine alignments in translation data, potentially improving the performance of machine translation systems trained on large, automatically-aligned corpora in the future.", "venue": "ACL", "year": 2013, "title": "Cut the noise: Mutually reinforcing reordering and alignments for improved machine translation"} +{"pid": "da7b1a20-f741-4e59-b4f0-3489e344b5f3", "context": "Chinese word segmentation is the first step in any Chinese NLP system and typically relies on a lexicon and handcrafted linguistic resources.", "key_idea": "This paper presents a new algorithm for segmenting Chinese texts that operates without the use of any lexicon or hand-crafted linguistic resource, instead deriving the required statistical data directly from raw Chinese corpora.", "method": "The authors conduct preliminary experiments with this new segmentation algorithm, using mutual information and differences in t-scores between characters as the statistical data for segmentation.", "outcome": "Preliminary experimentation indicates that the proposed segmentation algorithm has acceptable accuracy.", "future_impact": "The authors anticipate that their new approach could improve the performance of existing segmenters, especially in dealing with unknown words and adapting to various domains. The algorithm could also function as a standalone segmenter in some NLP applications.", "venue": "ACL", "year": 1998, "title": "Chinese Word Segmentation without Using Lexicon and Hand-crafted Training Data"} +{"pid": "645c5e3fd68f896efa22c582", "context": "Large Language Models (LLMs) often learn correlations between labels and features that are irrelevant to the task, leading to poor generalization on out-of-distribution data.", "key_idea": "The paper proposes explanation-based finetuning as an approach to mitigate LLMs' reliance on spurious correlations. Unlike standard finetuning, the model is also finetuned to generate a free-text explanation supporting its answer.", "method": "The model is finetuned on artificially constructed training sets containing different types of spurious cues and tested on a test set without these cues. The method's robustness against spurious cues is tested across four classification tasks: ComVE, CREAK, e-SNLI, and SBIC.", "outcome": "Compared to standard finetuning, this method is more robust against spurious cues, showing improved accuracy on four classification tasks: ComVE (+1.2), CREAK (+9.1), e-SNLI (+15.4), and SBIC (+6.5). It is also demonstrated that the method works with explanations generated by the model.", "future_impact": "The method's ability to work with model-generated explanations implies its potential applicability to more datasets without human-written explanations.", "venue": "ACL", "year": 2023, "title": "Explanation-based Finetuning Makes Models More Robust to Spurious Cues"} +{"pid": "646d863cd68f896efa09eebb", "context": "The main focus of research on automatic readability assessment (ARA) has shifted towards using expensive deep learning-based methods with the primary goal of increasing models' accuracy, but this approach is rarely applicable for low-resource languages where traditional handcrafted features are still widely used.", "key_idea": "The authors propose taking a step back from the technical component and focus on how linguistic aspects such as mutual intelligibility or degree of language relatedness can improve ARA in a low-resource setting.", "method": "The authors collect short stories written in three languages in the Philippines, Tagalog, Bikol, and Cebuano, to train readability assessment models and explore the interaction of data and features in various cross-lingual setups.", "outcome": "The results show that the inclusion of CrossNGO, a novel specialized feature exploiting n-gram overlap applied to languages with high mutual intelligibility, significantly improves the performance of ARA models compared to the use of off-the-shelf large multilingual language models alone.", "future_impact": "When both linguistic representations are combined, the authors achieve state-of-the-art results for Tagalog and Cebuano, and establish baseline scores for ARA in Bikol, possibly influencing future research directions in the field.", "venue": "ACL", "year": 2023, "title": "Automatic Readability Assessment for Closely Related Languages"} +{"pid": "5efdb17791e011a13faee537", "context": "The automatic identification and categorization of misogyny is an overlooked task in natural language processing, despite its importance for identifying hate speech in social Internet interactions.", "key_idea": "The authors' study tackles this sentence classification task from a representation learning perspective, using both a bidirectional LSTM and BERT optimized with different metric learning loss functions including contrastive loss, triplet loss, center loss, congenerous cosine loss and additive angular margin loss.", "method": "The authors optimized a bidirectional LSTM and BERT using various types of loss functions including contrastive loss, triplet loss, center loss, congenerous cosine loss, and additive angular margin loss.", "outcome": "The study achieves a new state-of-the-art for the task using the fine-tuned BERT model. They also found that almost all loss functions perform equally well in this setting, matching the regular cross entropy loss.", "future_impact": "The authors have made their code openly available, which can facilitate further research and easy reproducibility of the results.", "venue": "ACL", "year": 2020, "title": "A Metric Learning Approach To Misogyny Categorization"} +{"pid": "30427b4a-90c3-4c5d-8707-8afe0bf1508b", "context": "Many current information extraction techniques require supervised training data, limiting their applicability.", "key_idea": "The paper presents an approach to learn effective models for certain field structured extraction tasks, such as classified advertisements and bibliographic citations, using primarily unsupervised learning and small amounts of prior knowledge.", "method": "The proposed approach uses hidden Markov models (HMMs) as the underlying generative model. Prior knowledge of desired solutions is exploited to improve the quality of the learned structures.", "outcome": "The paper found that unsupervised methods can attain accuracies with 400 unlabeled examples comparable to those attained by supervised methods on 50 labeled examples, and that semi-supervised methods can make good use of small amounts of labeled data.", "future_impact": "The proposed approach to field structured extraction tasks could point to new ways of addressing limitations in the need for labelled data in information extraction techniques.", "venue": "ACL", "year": 2005, "title": "Unsupervised Learning of Field Segmentation Models for Information Extraction"} +{"pid": "628748ea5aee126c0ffd4bc0", "context": "The problem of learning fixed-length vector representations of characters in novels is addressed in this study. Previous advances in word embeddings were effective in learning entity representations from short texts, but failed on longer documents because they didn't capture full book-level information.", "key_idea": "This paper proposes two novel methods for representing characters: graph neural network-based embeddings from a full corpus-based character network, and low-dimensional embeddings constructed from the occurrence pattern of characters in each novel.", "method": "The quality of these character embeddings is tested using a new benchmark suite designed to evaluate character representations, which encompasses 12 different tasks.", "outcome": "The techniques combined with text-based embeddings resulted in the best character representations, outperforming text-based embeddings in four of the tasks.", "future_impact": "The researchers made their dataset publicly available to stimulate additional work in this area.", "venue": "ACL", "year": 2022, "title": "Learning and Evaluating Character Representations in Novels"} +{"pid": "64927546d68f896efa88a104", "context": "Fine-tuning has been proven to be an effective way to transfer the learned knowledge of pre-trained language models to downstream tasks, yet it often overfits the target data and degrades generalization. Current studies attribute it to catastrophic forgetting without identifying what knowledge is actually transferable.", "key_idea": "The authors propose framing fine-tuning into a causal graph to identify the missing causal effects from the pre-trained data, and they introduce a unified objective for fine-tuning to bring back the causality, which they claim can mitigate negative transfer while preserving knowledge from pre-trained language models (PLMs).", "method": "The authors implement the proposed method on commonsense QA with a heuristic estimation. The approach was tested against state-of-the-art fine-tuning methods on six commonsense QA datasets.", "outcome": "In the experiments, their method outperformed all state-of-the-art fine-tuning methods on all six commonsense QA datasets.", "future_impact": "The method can be implemented as a plug-in module which potentially enhances the performance of existing QA models.", "venue": "ACL", "year": 2023, "title": "Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference"} +{"pid": "60a4e22091e011e398b0cf59", "context": "Many types of text style transfer can be achieved with only small, precise edits, but existing methods focus on single-span edits or require parallel style text pairs for training.", "key_idea": "The study proposes a coarse-to-fine editor for style transfer that concurrently edits multiple spans in the source text using Levenshtein edit operations, and to train the model, an unsupervised data synthesis procedure is introduced.", "method": "Text is converted to style-agnostic templates using style classifier attention, and the slots in these templates are filled using fine-tuned pretrained language models. The proposed method is tested on sentiment and politeness transfer tasks.", "outcome": "The proposed method outperforms existing generation and editing style transfer methods on sentiment (Yelp, Amazon) and politeness (Polite) transfer. It also results in higher quality parallel style pairs and improves model performance compared to previous methods on unsupervised data synthesis.", "future_impact": "The research potentially offers an approach to improve other text generation models as it provides a way to perform multiple-span edits and an unsupervised data synthesis procedure.", "venue": "ACL", "year": 2021, "title": "LEWIS - Levenshtein Editing for Unsupervised Text Style Transfer."} +{"pid": "9e4d48d4-44cc-48c9-a0b8-a6b5cb574de2", "context": "Natural Language Processing technologies typically require a computer science background and overlook the needs of students who may wish to learn through hands-on experiences. Existing systems also grapple with issues of platform dependency and restricted computer lab times.", "key_idea": "The authors have developed web interfaces for a range of Natural Language Processing technologies, designed for use by students without a computer science background. These interfaces allow students to experiment, view system outputs and inner workings, and modify the system's knowledge bases.", "method": "The authors discuss their approach and lessons learned from developing these web interfaces for Natural Language Processing technologies allowing students to observe and modify the behavior of the system.", "outcome": "Web interfaces have been successfully built for numerous Natural Language Processing technologies, bypassing platform dependency issues and constraints related with computer lab timings.", "future_impact": "These web interfaces can substantially boost hands-on learning experiences for students without computer science backgrounds, enabling them to extensively interact with and learn from Natural Language Processing technologies.", "venue": "ACL", "year": 2005, "title": "Web-Based Interfaces for Natural Language Processing Tools"} +{"pid": "39cba680-82de-49f9-b823-9789b463a637", "context": "Efforts to represent meaning in natural language processing systems have not effectively utilized the insights of lexicography. Previous attempts have resulted in progress, but have not come to fruition. Many in the field have concluded that the contents of a dictionary have little value for linguistics or artificial intelligence due to its size and complexity.", "key_idea": "The author believes that by developing a comprehensive model that brings together the detail of a dictionary, it is possible to identify the requirements for a semantic representation of an entry in the lexicon for use in natural language processing systems.", "method": "The author attempts to develop a comprehensive model to bring together the details of a dictionary to identify the requirements for a semantic representation in a lexicon.", "outcome": "The author describes what they have learned from their attempt to integrate lexicographical insights into natural language processing systems.", "future_impact": "This work may influence future efforts in natural language processing by prompting researchers to consider dictionary details and lexicography when developing semantic representations.", "venue": "ACL", "year": 1980, "title": "REQUIREMENTS OF TEXT-PROCESSING-LEXICONS"} +{"pid": "6476d20cd68f896efaf727d9", "context": "Membership Inference attacks (MIAs) aim to predict whether a data sample was part of the training data of a machine learning model, and are widely used for assessing the privacy risks of language models. However, the prevalent method of simple thresholding of the model score can lead to high false-positive rates, and while recent reference-based attacks can improve the accuracy of MIAs, they involve the unrealistic assumption that an adversary has access to samples closely resembling the original training data.", "key_idea": "The authors propose 'neighbourhood attacks' where they compare model scores for a given sample to scores of synthetically generated neighbour texts, thereby eliminating the need for access to the training data distribution.", "method": "To validate their proposed method, they analyze the efficacy of their attack compared to other existing methods, including reference-free methods as well as reference-based methods with both perfect and imperfect knowledge about the training data distribution.", "outcome": "The neighbourhood attacks proposed in this paper outperformed the existing reference-free and reference-based attacks with imperfect knowledge, and proved to be competitive with reference-based attacks that have perfect knowledge about the training data distribution.", "future_impact": "This study demonstrates the need for a reevaluation of the threat model of adversarial attacks, as the new approach reveals greater vulnerability than previous models anticipated.", "venue": "ACL", "year": 2023, "title": "Membership Inference Attacks against Language Models via Neighbourhood Comparison"} +{"pid": "0fed45d4-1c3e-460f-ab62-7840d3814f29", "context": "Current sentiment classification of product reviews relies on pre-defined sentiment lexicons, which don't always accurately capture the sentiment of the text due to situational differences and personal attitudes.", "key_idea": "The authors propose using maximally occurring phrases in text as a more representative and context-sensitive feature for sentiment classification.", "method": "The authors explore the use of maximally occurring phrases in text as features for sentiment classification, looking for representative phrases used typically in positive and negative reviews.", "outcome": "The paper does not provide clear, measurable outcomes in the abstract.", "future_impact": "The implication of this study is that adopting a more nuanced, context-dependent approach to sentiment analysis might improve the quality of sentiment classification, possibly avoiding reliability on pre-defined sentiment lexicons in future.", "venue": "ACL", "year": 2011, "title": "Automatic Sentiment Classification of Product Reviews Using Maximal Phrases Based Analysis"} +{"pid": "6344dee690e50fcafd24e90b", "context": "Large-scale diffusion neural networks have marked a significant development in text-to-image generation but their interpretability is limited.", "key_idea": "The authors propose DAAM, a method to perform a text-image attribution analysis on Stable Diffusion, a recently open-sourced model. DAAM generates pixel-level attribution maps by upscaling and aggregating cross-attention word-pixel scores in the denoising subnetwork.", "method": "DAAM is evaluated by testing its semantic segmentation ability on nouns, and its generalized attribution quality on all parts of speech is rated by humans. The authors study the role of syntax in the pixel space and several semantic phenomena using DAAM.", "outcome": "Findings include identifying that cohyponyms worsen generation quality and descriptive adjectives attend too broadly.", "future_impact": "This paper, being the first to interpret large diffusion models from a visuolinguistic perspective, opens up future lines of research in the area of text-to-image generation.", "venue": "ACL", "year": 2022, "title": "What the DAAM: Interpreting Stable Diffusion Using Cross Attention"} +{"pid": "dff09cc8-1c33-497a-8b4b-4f15f23d488b", "context": "Compiling a bilingual lexicon of nouns and proper nouns from unaligned, noisy parallel texts of Asian/Indo-European language pairs is a complex task that has not been adequately addressed.", "key_idea": "The authors propose a pattern matching method which leverages tagging information of one language and represents word frequency and position information for high and low-frequency words in two different vector forms.", "method": "The paper introduces new anchor point finding and noise elimination techniques for its pattern matching method. The results are also demonstrated by using them in the compilation of domain-specific noun phrases.", "outcome": "The proposed method achieves a precision of 73.1% in compiling a bilingual lexicon from unaligned, noisy parallel texts.", "future_impact": "The results of the study can be utilized for the compilation of domain-specific noun phrases, enriching the linguistic resources in those specific fields.", "venue": "ACL", "year": 1995, "title": "A Pattern Matching Method for Finding Noun and Proper Noun Translations from Noisy Parallel Corpora"} +{"pid": "f6f0bb46-aeb5-4efd-82ce-2d40efaffe8b", "context": "The BioNLP 2009 shared task on event extraction had posed challenges that required an effective solution.", "key_idea": "This paper proposes two strong baselines for the BioNLP 2009 event extraction task; a re-implemented rule-based approach and a new approach that replaces rule-based components with support vector machine classifiers.", "method": "The authors first explored the effects of domain-adapted parsing on the task using the rule-based approach. They then replaced the rule-based components with support vector machine classifiers.", "outcome": "Both approaches achieved good performances and are reproducible baselines. The approach using support vector machine classifiers achieved performance near the state-of-the-art without using any external resources.", "future_impact": "The authors concluded with suggestions for future work with respect to the task representation, which may inspire potential improvements for the event extraction task in BioNLP.", "venue": "ACL", "year": 2010, "title": "Two Strong Baselines for the BioNLP 2009 Event Extraction Task"} +{"pid": "60dfc80a91e01129379b3774", "context": "Quality Estimation (QE) is the task of predicting Machine Translation quality without reference translations. While large multilingual models have had success in this area, their high inference time, disk, and memory requirements limit their practical use. Models trained on distilled pre-trained representations still remain too large for many applications.", "key_idea": "The authors propose to directly transfer knowledge from a robust QE teacher model to a much smaller and shallower architecture. This approach is combined with data augmentation.", "method": "The authors use the proposed direct knowledge transfer approach combined with data augmentation to create lightweight models and compare their performance with that of models using distilled pre-trained representations.", "outcome": "The lightweight QE models developed using the proposed approach perform competitively with models using distilled pre-trained representations, with 8x fewer parameters.", "future_impact": "The proposed approach may enable the development of more efficient and practical quality estimation models, expanding the possibilities for real-world applications of machine translation quality prediction.", "venue": "ACL", "year": 2021, "title": "Knowledge Distillation for Quality Estimation."} +{"pid": "54c85509-350e-4365-8b61-065ea8de7279", "context": "The WordNet sense inventory and sense-tagged corpus require a coarse-grained version for easier utilization.", "key_idea": "The authors proposed and developed a coarse-grained English all-words task as a part of SemEval-2007.", "method": "The authors prepared a coarse version of the WordNet sense inventory and used a sense-tagged corpus for the task.", "outcome": "The authors have presented the results of the participating systems operating under the new coarse-grained task.", "future_impact": "The authors intend to explore future directions and likely enhancements in relation to the proposed coarse-grained approach.", "venue": "ACL", "year": 2007, "title": "SemEval-2007 Task 07: Coarse-Grained English All-Words Task"} +{"pid": "631aaf0690e50fcafdafa218", "context": "Unfaithful summaries have been widely investigated in abstractive summarization but less attention has been paid to the same issue in extractive summarization.", "key_idea": "This work defines a typology with five types of broad unfaithfulness problems that can appear in extractive summaries - incorrect coreference, incomplete coreference, incorrect discourse, incomplete discourse, and misleading information. The authors then propose a new metric, ExtEval, to detect these problems.", "method": "The authors asked humans to label these problems in 1500 English summaries produced by 15 diverse extractive systems. To detect these problems automatically, they compared ExtEval with 5 existing faithfulness evaluation metrics for summarization.", "outcome": "The study found that 33% of the summaries have at least one of the five issues and that current faithfulness evaluation metrics for summarization have poor correlations with human judgment. ExtEval, a new metric proposed by the authors, is shown to have the best performance for detecting unfaithful extractive summaries.", "future_impact": "This work can increase the awareness of unfaithfulness problems in extractive summarization and help future work to evaluate and resolve these issues.", "venue": "ACL", "year": 2022, "title": "Extractive is not Faithful: An Investigation of Broad Unfaithfulness Problems in Extractive Summarization"} +{"pid": "506400a0-42b9-4123-ac32-de6e54a5a891", "context": "In statistical machine translation, word alignment is a known challenge not necessarily informed by the topical contents of document-pairs.", "key_idea": "The authors propose a novel bilingual topical admixture (BiTAM) formalism for word alignment in statistical machine translation. This formalism assumes that parallel sentence-pairs within a document-pair constitute a mixture of hidden topics, and each word-pair follows a topic-specific bilingual translation model.", "method": "Three BiTAM models are proposed to capture topic sharing at different levels of linguistic granularity. With those models, the authors perform word-alignment leveraging topical contents of document-pairs and evaluate the results using efficient variational approximation algorithms for inference and parameter estimation.", "outcome": "The preliminary experiments show that the proposed models improve word alignment accuracy and lead to better translation quality.", "future_impact": "The inferred latent topics from the BiTAM models can facilitate coherent pairing of bilingual linguistic entities that share common topical aspects, potentially improving future machine translation tools.", "venue": "ACL", "year": 2006, "title": "BiTAM: Bilingual Topic AdMixture Models for Word Alignment"} +{"pid": "64ae66963fda6d7f06845237", "context": "Translation difficulty arises from translation ambiguity, and it is typically measured by recording the diversity of responses from human translators and the time taken, but such measurements are expensive and don't scale.", "key_idea": "In this study, the authors propose the use of word alignments computed over large scale bilingual corpora to predict lexical translation difficulty.", "method": "The authors develop and evaluate predictors using behavioral data from translations provided both in and out of context.", "outcome": "The authors report results that improve upon a previous embedding-based method, indicating that their proposed method is more effective in predicting translation difficulty.", "future_impact": "This work helps deepen the understanding of cross-lingual differences and causes of translation difficulty.", "venue": "ACL", "year": 2023, "title": "Predicting Human Translation Difficulty Using Automatic Word Alignment"} +{"pid": "e8998baf-391e-4674-a9fe-b87fd32ca3a3", "context": "Concepts in speech recognition usually rely on conventional automatic speech recognition systems, or use text transcriptions and other conventional linguistic annotations.", "key_idea": "The authors came up with a method for discovering word-like acoustic units in the continuous speech signal and then tying them to semantically relevant image regions without the usage of automatic speech recognition, or text transcripts or any linguistic annotations.", "method": "Using a collection of images and spoken audio captions, the proposed method was implemented to identify word-like units in speech signals and connect them with corresponding image regions.", "outcome": "The model is able to effectively detect spoken instances of words within utterances and associate them with their corresponding image regions containing the related objects, showcasing successful spoken language acquisition by the computer.", "future_impact": "The proposed method represents a distinct form of spoken language acquisition, where the computer not only learns to categorize spoken words by sound, but also imbues them with semantic meaning by grounding them in images, potentially expanding the realm of language and image processing.", "venue": "ACL", "year": 2017, "title": "Learning Word-Like Units from Joint Audio-Visual Analylsis"} +{"pid": "64ae66693fda6d7f06842ba2", "context": "Extractive summarization relies heavily on context information. However, certain pattern factors related to sentence importance such as sentence position or specific n-gram tokens are not generalized across datasets thus affecting the performance of extractive summarization models when transferred to new datasets.", "key_idea": "The paper proposes applying disentangled representation learning to extractive summarization to separate the two key factors, context and pattern. This approach aims to improve generalization ability in low-resource settings.", "method": "Two groups of losses are proposed for encoding and disentangling sentence representations into context representations and pattern representations which can be used solely or in a fine-tuned combination depending on the data availability. This approach is tested experimentally on three summarization datasets from different domains.", "outcome": "The experimental results show the effectiveness of the proposed approach enabling better performance of extractive summarization models in low-resource settings.", "future_impact": "The improved performance of extractive summarization models in low-resource settings thanks to this approach could potentially influence the development of future models in this field.", "venue": "ACL", "year": 2023, "title": "Separating Context and Pattern: Learning Disentangled Sentence Representations for Low-Resource Extractive Summarization"} +{"pid": "621d8ecc5aee126c0f73b2a1", "context": "Entity linking, which involves linking entity mentions in a document to a knowledge base, has been the focus of many previous studies. These studies have primarily focused on Wikipedia-derived knowledge bases, with little work done on Wikidata which is the most extensive crowdsourced knowledge base.", "key_idea": "The authors propose a novel strategy for candidate retrieval within the realm of entity linking. This approach is based on entity profiling where a sequence-to-sequence model is used to generate the profile of a target entity, which consists of its title and description. The profile is then used to query the indexed search engine and retrieve candidate entities.", "method": "The authors index Wikidata entities and their textual fields into a text search engine like Elasticsearch. Given a mention and its context, a sequence-to-sequence model is used to generate the entity's profile, which is then used to fetch candidate entities from the search engine. A highly effective hybrid method for candidate retrieval is designed by combining this approach with the traditional approach of using a Wikipedia anchor-text dictionary.", "outcome": "Using a simple cross-attention reranker, the proposed entity linking framework achieves state-of-the-art results on three Wikidata-based datasets and demonstrates strong performance on TAC-KBP-2010.", "future_impact": "While not directly stated in the abstract, the authors imply that their novel approach opens up numerous real-world applications due to the extensive scale of Wikidata.", "venue": "ACL", "year": 2022, "title": "Improving Candidate Retrieval with Entity Profile Generation for Wikidata Entity Linking"} +{"pid": "e4fdb43f-1ce1-4352-8789-2e6ca88de6af", "context": "Processing natural language messages requires understanding implicit linguistic information, a challenge that existing systems struggle to deal with effectively.", "key_idea": "The authors propose a Prolog system PUNDIT, which consists of distinct syntactic, semantic and pragmatic modules. The novelty is to let syntax and semantics recognize missing linguistic entities as implicit entities, so they can be labelled, and a reference resolution task can find specific referents for these entities.", "method": "The method involves interplay between the syntactic, semantic, and pragmatic modules of PUNDIT to recognize implicit entities and then label and find referents for these entities.", "outcome": "The paper describes an approach for making implicit linguistic information explicit becomes a subset of the tasks performed by reference resolution. It doesn't present measurable outcomes.", "future_impact": "The success of this approach in understanding natural language can have potential implications in the field of natural language processing as it can help in improving models' performance that relies on implicit information.", "venue": "ACL", "year": 1986, "title": "RECOVERING IMPLICIT INFORMATION"} +{"pid": "d14bca70-1411-4da0-8f02-6d2af85565df", "context": "Before this study, Japanese technical terms were translated to English using commercial technical term dictionaries, but the need for a wider and more dynamic source of translation is required.", "key_idea": "The authors propose a system to extract English translations of Japanese technical terms by using the web as a bilingual dictionary, utilizing partially bilingual documents in the web.", "method": "The authors' method obtains translation candidates based on the distance of Japanese and English terms in web documents. A preliminary experiment was conducted to test their system.", "outcome": "The abstract doesn't provide specific results from the preliminary experiment.", "future_impact": "The authors do not explicitly state the future impact.", "venue": "ACL", "year": 2001, "title": "Using the web as a bilingual dictionary"} +{"pid": "fe9d0b7e-2a74-41bd-98d1-b00148119b32", "context": "Subjectivity and concreteness are known dimensions of lexical meaning, but their relationship has not been quantified in large corpora and applied to improve semantic classification tasks.", "key_idea": "The authors propose a corpus-based method to quantify the lexical subjectivity of adjectives. They also suggest exploiting the relationship between noun concreteness and adjective subjectivity to improve semantic classification of adjective-noun pairs.", "method": "The authors use a corpus-based method to quantify the lexical subjectivity of adjectives and correlate it with noun concreteness. This relationship is then leveraged to semantically classify adjective-noun pairs.", "outcome": "The use of the lexical subjectivity of adjectives and its correlation with noun concreteness proved effective in enhancing the performance of semantic classification of adjective-noun pairs.", "future_impact": "The found relationship between lexical subjectivity and noun concreteness and its application to semantic classification may influence future research involving lexical meaning and its application in natural language processing tasks.", "venue": "ACL", "year": 2014, "title": "Concreteness and Subjectivity as Dimensions of Lexical Meaning"} +{"pid": "6735896b-d7ac-4eff-b00f-34bf47085aa9", "context": "Supersenses, coarse-grained semantic categories, are useful for tasks like question answering or machine translation, but there has been no successful integration of supersenses into distributional word representations so far.", "key_idea": "The authors introduce a novel joint embedding model of words and supersenses, which aims to provide insights into the relationship between words and supersenses in the same vector space.", "method": "The authors use the joint embedding model in a deep neural network model to investigate the benefits of supersense enrichment in a range of downstream classification tasks.", "outcome": "Embeddings developed using the joint embedding model allow for significant improvements in a range of downstream classification tasks due to supersense enrichment.", "future_impact": "The joint embedding model and the findings about the advantages of integrating supersenses into distributional word representations could provide direction for further research in the field and advancements in pertinent tasks.", "venue": "ACL", "year": 2016, "title": "Supersense Embeddings: A Unified Model for Supersense Interpretation, Prediction, and Utilization"} +{"pid": "5eda6302-5c9a-4222-8c6a-399ed58388b4", "context": "In Chinese, zero anaphors occur frequently. Their identification and recovery in discourse processing present challenges due to their ambiguity and complex syntactic structures. There is a widely accepted belief that extra-linguistic knowledge is needed for resolving zero anaphors.", "key_idea": "The author proposes the use of syntactic tagging for detection of anaphora and introduces two principles based on linguistic information for recovering zero anaphors. The first principle is to return to the opening statement for recovery (topic continuity principle) and the second is to look for anaphors in the most recent element (recency principle), which can be modified by certain adverbial expressions.", "method": "The proposal is a theoretical framework, so there's no specified method for validation mentioned in the abstract.", "outcome": "No outcomes or results are provided in the abstract.", "future_impact": "The anticipated impact is not explicitly stated in the abstract, so it is marked as 'N/A'.", "venue": "ACL", "year": 2000, "title": "Zero Anaphors in Chinese Discourse Processing"} +{"pid": "cc71def1-95dc-4744-97cf-ab0c0c81f88d", "context": "The homogeneity of large parallel corpora in statistical analyses has been considered given, even though there might be differences in various sub-parts of the corpus.", "key_idea": "The authors took a novel approach of measuring the similarity between various sub-parts of a parallel corpus using discourse connective counts and compare with a general measure of lexical similarity based on \u03c72.", "method": "Measuring techniques were applied to a large parallel corpus to identify the similarity using general lexical and discourse connectives, and to reveal their differences.", "outcome": "The proposed method showed that discourse connectives provide a more sensitive measure, revealing differences that are not visible with the general measure. Evidence was also found for the existence of specific characteristics defining translated texts as opposed to non-translated ones, due to a universal tendency for explicitation.", "future_impact": "The paper suggests that more accurate measurements of similarity can be achieved by using discourse connectives, indicating that this could be a useful tool for further studies on corpus homogeneity and revealing specific characteristics of translated texts.", "venue": "ACL", "year": 2011, "title": "How Comparable are Parallel Corpora? Measuring the Distribution of General Vocabulary and Connectives"} +{"pid": "64ae66fb3fda6d7f0684b24f", "context": "Curriculum learning, in the existing literature, usually involves monotonic curricula and often use easy-to-hard or hard-to-easy transition, which may risk underperforming. The existing approaches lack an effective discovery mechanism to find optimal curricula.", "key_idea": "The authors introduce a new problem called curriculum discovery and propose a framework for curriculum learning capable of discovering effective curricula based on prior knowledge about sample difficulty, using annotation entropy and loss as measures of difficulty.", "method": "The authors showed the effectiveness of the proposed method using experiments across several Natural Language Processing (NLP) tasks.", "outcome": "The top-performing discovered curricula are often non-monotonic, and the curricula discovered for smaller datasets and models perform well on larger datasets and models.", "future_impact": "The introduced problem of curriculum discovery and the proposed framework can pioneer the advancement of more effective curriculum learning methods in various domains.", "venue": "ACL", "year": 2023, "title": "HuCurl: Human-induced Curriculum Discovery"} +{"pid": "6221834e5aee126c0f23c325", "context": "Early exiting allows instances to exit at different layers based on the estimated difficulty of instances. Traditional approaches employ heuristic metrics such as entropy to measure this difficulty, which leads to issues related to generalization and threshold tuning. Some have attempted to use 'learn-to-exit' modules to predict instance difficulty, but there are still questions about how well it can be learned.", "key_idea": "The study proposes a Hash-based Early Exiting approach (HASHEE), which normatively assigns each token to a specific exiting layer using hash functions. This approach contrasts with previous methods by requiring neither internal classifiers nor extra parameters, thus offering higher efficiency.", "method": "The validity of HASHEE is confirmed using experiments across several task types, including classification, regression, and generation tasks.", "outcome": "HASHEE outperforms previous state-of-the-art early exiting methods, achieving higher performance with fewer FLOPs and less inference time.", "future_impact": "The efficiency and superior performance of HASHEE has the potential to enhance future efforts in language understanding and generation research.", "venue": "ACL", "year": 2022, "title": "A Simple Hash-Based Early Exiting Approach For Language Understanding and Generation"} +{"pid": "648151f0d68f896efa687863", "context": "While Sentence Simplification is a valuable technique that can benefit language learners and children significantly, the current research primarily focuses on English. The lag in Chinese sentence simplification development is predominantly due to lack of data.", "key_idea": "This paper serves to alleviate this limitation by introducing CSS, a new dataset for assessing sentence simplification in Chinese, with manual simplifications collected from human annotators.", "method": "The authors perform data analysis to demonstrate the differences between English and Chinese sentence simplifications, testing several unsupervised and zero/few-shot learning methods on CSS and analyzing their automatic evaluation and human evaluation results.", "outcome": "Their analyses display the divergence between English and Chinese sentence simplifications, establish the performance of various learning methods on CSS, and reflect upon the results drawn from automatic and human evaluations.", "future_impact": "Beyond the immediate outcome, the paper also suggests the possibility of Large Language Models serving as high-quality Chinese sentence simplification systems, subject to further evaluation against CSS.", "venue": "ACL", "year": 2023, "title": "A New Dataset and Empirical Study for Sentence Simplification in Chinese"} +{"pid": "5e91957f91e011505f40a387", "context": "Practical applications of abstractive summarization models are limited by frequent factual inconsistencies with respect to their input. Existing automatic evaluation metrics for summarization are largely insensitive to such errors.", "key_idea": "The paper introduces a new automatic evaluation protocol, QAGS, that is designed to identify factual inconsistencies in a generated summary through asking questions about a summary and its source.", "method": "To evaluate QAGS, the authors collect human judgments of factual consistency on model-generated summaries for the CNN/DailyMail and XSUM summarization datasets.", "outcome": "The authors found that QAGS has substantially higher correlations with these judgments than other automatic evaluation metrics. Additionally, it provides answers and questions generated while computing QAGS, indicating which tokens of a summary are inconsistent.", "future_impact": "QAGS is anticipated to be a promising tool in automatically generating usable and factually consistent text.", "venue": "ACL", "year": 2020, "title": "Asking and Answering Questions to Evaluate the Factual Consistency of Summaries"} +{"pid": "6482a38ed68f896efa8db3ed", "context": "The growth of scientific publications, especially during the COVID-19 pandemic, necessitates tools for effective comprehension of the latest advancements. An essential part of understanding this literature is research aspect classification, which categorizes sentences in abstracts into Background, Purpose, Method, and Finding.", "key_idea": "This study explores the impact of different datasets on model performance for the crowd-annotated CODA-19 research aspect classification task and evaluates the effectiveness of large language models (LLMs), such as LLaMA, GPT-3, ChatGPT, and GPT-4.", "method": "The study investigates model performance using the large, automatically curated PubMed 200K RCT dataset for the CODA-19 research aspect classification task. In addition, the authors compare the performance of large language models like LLaMA, GPT-3, ChatGPT, and GPT-4.", "outcome": "The study finds that using the PubMed 200K RCT dataset does not improve performance for the CODA-19 task. It also observes that GPT-4 does not outperform the SciBERT model fine-tuned on the CODA-19 dataset, emphasizing the importance of task-aligned datasets.", "future_impact": "The findings of this paper underline the importance of using dedicated and task-aligned datasets for target tasks in research aspect classification, which could influence the direction of future work in this area.", "venue": "ACL", "year": 2023, "title": "Good Data, Large Data, or No Data? Comparing Three Approaches in Developing Research Aspect Classifiers for Biomedical Papers"} +{"pid": "5eb7896cda5629cf2443067b", "context": "A decade has passed since the establishment of SPMRL, an initiative that unified research efforts to tackle the challenges of Statistical Parsing for Morphologically-Rich Languages (MRLs).", "key_idea": "The authors outline the lessons learned from pre-neural era MRL parsing and suggest that incorporating symbolic ideas from the SPMRL era into modern neural architectures could be beneficial.", "method": "The authors generalize strategies for designing Neural Models for MRLs (NMRL) and test these strategies by performing multi-tagging on Hebrew, a morphologically-rich, high-fusion language.", "outcome": "The authors find preliminary support for their strategies in the task of multi-tagging in Hebrew.", "future_impact": "The outlined strategies for creating NMRL could potentially improve NLP for MRLs, pushing the field to a new level.", "venue": "ACL", "year": 2020, "title": "From SPMRL to NMRL: What Did We Learn (and Unlearn) in a Decade of Parsing Morphologically-Rich Languages (MRLs)?"} +{"pid": "6456389bd68f896efacf6ade", "context": "Even though transformer-based systems have improved accuracies with fewer training examples, challenges still persist in data acquisition for rare-class tasks where the class label is very infrequent. Active learning has been proposed to alleviate these challenges, but the selection strategy has not been thoroughly evaluated.", "key_idea": "The authors propose combining transfer learning and active learning solutions to tackle the issue of dissonance detection in rare class problems. They also propose a new acquisition strategy termed probability-of-rare-class (PRC).", "method": "The authors evaluate their approach on the task of collecting language samples of cognitive dissonance from social media, using models trained on closely related tasks.", "outcome": "The results show that the proposed PRC strategy is effective in guiding annotations and improving model accuracy, and that transfer-learning can support the early performance of the learner, though it doesn't help in active learning iterations.", "future_impact": "These insights could guide the design of future systems for the detection of rare classes, particularly in fields such as social media data analysis.", "venue": "ACL", "year": 2023, "title": "Transfer and Active Learning for Dissonance Detection: Addressing the Rare-Class Challenge"} +{"pid": "49990260-4d5b-4932-a65c-6e9081873e4c", "context": "Previously, the approach to assessing text difficulty level, particularly for second-language learners and applied to the Interagency Language Roundtable (ILR) proficiency scale, was not defined in a language-independent manner.", "key_idea": "The authors propose a new language-independent baseline for text difficulty assessment using z-normalized shallow length features and TF-LOG weighted vectors on bag-of-words for multiple languages.", "method": "The proposed baseline approach is tested using Support Vector Machines and the Margin-Infused Relaxed Algorithm, with mean squared error as the performance metric.", "outcome": "Through the experimentation, the authors identified which features are most predictive of a given level.", "future_impact": "This research could establish a standardized, language-independent approach to assessing text difficulty level specifically intended for second-language learners.", "venue": "ACL", "year": 2013, "title": "A Language-Independent Approach to Automatic Text Difficulty Assessment for Second-Language Learners"} +{"pid": "64ae66b43fda6d7f06846951", "context": "Benchmark datasets are currently used to interpret model performance without quantified measures indicating their adequacy for evaluating model generalization across various datasets.", "key_idea": "The authors propose a simple novel metric based on a nearest neighbors approach to quantify the overlap between training and evaluation sets which characterizes the adequacy of an individual dataset to evaluate model generalization.", "method": "The authors map instances using both generic and task-specific embedding methods to identify nearest training examples for each evaluation example and apply their method on eleven classification and extractive QA tasks.", "outcome": "The analysis reveals a wide range of train-evaluation overlap and illustrates how data collection methods and the difficulty of tasks can influence the level of overlap. It also helps in identifying challenging or potentially mislabeled examples.", "future_impact": "This analysis and resultant metric could provide insights for constructing datasets to evaluate generalization, which could improve the ability to interpret model performance and increase the accuracy and reliability of machine learning models.", "venue": "ACL", "year": 2023, "title": "Quantifying Train-Evaluation Overlap with Nearest Neighbors"} +{"pid": "64ae66773fda6d7f06843830", "context": "The Visual Word Sense Disambiguation (VWSD) task aims to find the most related image among 10 images to an ambiguous word in some limited textual context. The best official baseline achieves 37.20% and 54.39% macro-averaged hit rate and Mean Reciprocal Rank (MRR) respectively.", "key_idea": "The authors propose using AltCLIP features in conjunction with a 3-layer standard transformer encoder to compare the cosine similarity between the given phrase and different images. They also use a subset of LAION-5B to improve the model's generalization.", "method": "The authors train and validate their approach on the Visual Word Sense Disambiguation (VWSD) task, using AltCLIP features and a 3-layer standard transformer encoder.", "outcome": "The proposed model reaches a macro-averaged hit rate of 39.61% and a macro-averaged MRR of 56.78%, which are improvements over the baseline.", "future_impact": "The authors' code will be made publicly available on GitHub, which may encourage further advancements in the field.", "venue": "ACL", "year": 2023, "title": "PMCoders at SemEval-2023 Task 1: RAltCLIP: Use Relative AltCLIP Features to Rank"} +{"pid": "5db9293447c8f766461f0788", "context": "Estimation of symmetric properties of distributions from i.i.d. samples has typically been handled by the empirical estimation in the easy region and by complex estimators in the difficult region. Earlier universal estimation methods have been based on PML.", "key_idea": "The authors propose a general framework for symmetric property estimation that targets the difficult region by approximately computing the profile maximum likelihood (PML) distribution.", "method": "Algorithms leveraging pseudo PML distributions, derived within the authors' proposed framework, are used in the implementation.", "outcome": "The authors demonstrate that their symmetric property estimation framework is sample complexity optimal for many properties in a broader parameter regime than the previous universal estimation approaches based on PML.", "future_impact": "The algorithms based on pseudo PML distributions resulting from this framework may have practical applications in the field of parameter estimation.", "venue": "NeurIPS", "year": 2020, "title": "A General Framework for Symmetric Property Estimation"} +{"pid": "6284fc4b5aee126c0f3ebb88", "context": "Meta-learning has emerged as an important method for enhancing flexibility and sample efficiency in reinforcement learning. One key challenge in this area of research has been the lack of suitable benchmarks, as existing ones are either too simplistic or too undefinable to support in-depth analysis.", "key_idea": "The paper introduces a new benchmark for meta-RL research called 'Alchemy', a 3D video game developed in Unity. Alchemy features a latent causal structure that changes from one episode to another, enabling structure learning, online inference, hypothesis testing, and action sequencing based on abstract domain knowledge.", "method": "The authors apply a pair of reinforcement learning agents to the Alchemy platform and conduct an in-depth analysis of one of these agents.", "outcome": "Initial results reveal a clear and specific failure of meta-learning, thereby establishing 'Alchemy' as a challenging benchmark for meta-reinforcement learning.", "future_impact": "The authors are releasing 'Alchemy', with its analysis tools and sample agent trajectories, as a public resource, encouraging further research and development in the meta-reinforcement learning field.", "venue": "NeurIPS", "year": 2021, "title": "Alchemy: A benchmark and analysis toolkit for meta-reinforcement learning agents."} +{"pid": "63a413f690e50fcafd6d1d34", "context": "Test-time adversarial attacks have posed serious challenges to the robustness of machine-learning models, and in many settings the adversarial perturbation need not be bounded by small $\\ell_p$-norms. The challenges are specifically notable in program analysis and security tasks.", "key_idea": "The authors introduce the concept of relational adversaries, who create adversarial examples in a reflexive-transitive closure of a logical relation, and propose a learning framework, called 'normalize-and-predict', which leverages input normalization to achieve provable robustness against such adversaries.", "method": "The authors apply their 'normalize-and-predict' framework to tasks of source code authorship attribution and malware detection, exploring its impact on the robustness-accuracy trade-off against relational adversaries.", "outcome": "The results show that the proposed learning framework significantly improves the robustness of models against relational adversaries, outperforming adversarial training, a noteworthy defense mechanism, by a wide margin.", "future_impact": "Given its strong performance in enhancing robustness, the proposed 'normalize-and-predict' framework may be applicable to a broader range of tasks and scenarios dealing with adversarial attacks, specifically those instigated by relational adversaries.", "venue": "NeurIPS", "year": 2022, "title": "Robust Learning against Relational Adversaries"} +{"pid": "8a6102a8-5083-427f-b31c-8b0a7d05711f", "context": "Visual attention, the ability to dynamically restrict processing to a subset of the visual field, is argued by researchers to play a crucial role in intermediate level visual tasks. However, the mechanism of visual attention and its functionality in different visual areas is still a topic of extensive research.", "key_idea": "This paper proposes VISIT, a novel neural network model of visual attention, which is specifically designed to model the search for target objects in scenes containing multiple distractors.", "method": "The study concentrates on the biological plausibility of VISIT and its relationship to the primary visual cortex, pulvinar, superior colliculus and posterior parietal areas.", "outcome": "VISIT successfully reflects the known psychophysical data on visual search and visual attention, and also corresponds with much of the physiological data on attention.", "future_impact": "The proposed model provides a novel perspective on the functionality of a number of visual areas and can potentially lead to a better understanding of how visual attention works in various contexts.", "venue": "NeurIPS", "year": 1991, "title": "VISIT: A Neural Model of Covert Visual Attention"} +{"pid": "61a8829b6750f82b17638ad6", "context": "Previous works have proved the Turing-completeness of recurrent neural networks (RNNs), but these proofs relied on the assumption of neurons with unbounded precision, which is not practical or biologically plausible.", "key_idea": "The authors propose a dynamically growing memory module for RNNs made of neurons of fixed precision, which recruits new neurons when needed and releases them when memories become irrelevant.", "method": "They prove the Turing completeness of a 54-neuron bounded-precision RNN with growing memory modules through a theoretical analysis, and the result is verified by simulating a Universal Turing Machine.", "outcome": "This work proves that a 54-neuron bounded-precision RNN with growing memory modules can simulate a Universal Turing Machine, with time complexity linear in the simulated machine's time and independent of the memory size.", "future_impact": "This work is extendable to stack-augmented RNNs and provides a theoretical foundation for further research on the capabilities of both unbounded-precision and bounded-precision RNNs.", "venue": "NeurIPS", "year": 2021, "title": "Turing Completeness of Bounded-Precision Recurrent Neural Networks."} +{"pid": "3a1edc11-c634-413b-a9ae-21a8711689b6", "context": "In extreme classification problems, learning algorithms need to handle an extremely large label set. These problems require efficient and effective solutions, as researchers have recently proposed the logarithmic time and space (LTLS) and error correcting output coding (ECOC) with loss-based decoding approaches.", "key_idea": "The research introduces a flexible and efficient approach for extreme classification that leverages output codes induced by graphs and allows efficient loss-based decoding. The unique aspect is the potential to provide a tradeoff between accuracy, model size and prediction time using only training data.", "method": "The authors theoretically and experimentally validate their approach, which involves finding the optimal tradeoff point using only the training data.", "outcome": "The experimental study proves the validity of the authors' assumptions and claims, showing that the proposed method can provide competitive performance when compared to state-of-the-art algorithms.", "future_impact": "As the study introduces a flexible and efficient approach with theoretical bounds for extreme classification problems, it could lead to advancements in large scale learning applications that require extreme classification.", "venue": "NeurIPS", "year": 2018, "title": "Efficient Loss-Based Decoding On Graphs For Extreme Classification"} +{"pid": "5e7495c591e0111c7cee14d4", "context": "Image classifiers are typically evaluated based on their test set accuracy. However, high accuracy can mask subtle failures in models. Convolutional neural networks (CNNs) achieving high scores on popular benchmarks can still exhibit underlying pathologies, functioning with high accuracy even when semantically salient features are absent.", "key_idea": "The paper introduces the concept of overinterpretation where a model provides a high-confidence decision without salient supporting input features. Also, the paper proposes a new method, Batched Gradient SIS, for discovering sufficient input subsets for complex datasets.", "method": "The authors tested different neural networks trained on CIFAR-10 and ImageNet to highlight their tendency to overinterpret. They used the newly proposed Batched Gradient SIS method to demonstrate the sufficiency of border pixels in ImageNet for training and testing.", "outcome": "Neural networks trained on CIFAR-10 and ImageNet exhibited overinterpretation, making confident predictions even when 95% of input images were masked. The study further showed that border pixels in ImageNet were sufficient for achieving high test accuracy.", "future_impact": "The authors indicate two methods, ensembling and input dropout, as means to mitigate the issue of overinterpretation in image classifiers, implying future research in these areas could lead to more robust and reliable models.", "venue": "NeurIPS", "year": 2021, "title": "Overinterpretation reveals image classification model pathologies."} +{"pid": "5cbe97e7-79bf-419a-a69e-1c551baef7cc", "context": "The problem of learning a series of tasks in a fully online Meta-Learning setting is studied. The aim is to use similarities amongst tasks to adapt an inner online algorithm and thus incur a low cumulative error over the tasks.", "key_idea": "The authors propose treating the online Mirror Descent meta-algorithm as an inner algorithm. This inner algorithm continually adapts under the influence of the minimum regularized empirical risk viewed as the meta-loss.", "method": "Keeping the process fully online, the authors approximate the meta-subgradients using the online inner algorithm. Their method is theoretically evaluated by deriving a cumulative error bound, and this analysis is converted to a statistical setting through online-to-batch arguments.", "outcome": "An upper bound on the approximation error is found, which aids in deriving the cumulative error bound for the proposed method. Preliminary numerical experiments confirm the theoretical findings.", "future_impact": "The authors show two examples of framework instantiation where the meta-parameter is either a common bias vector or a feature map, which may serve as a foundation for future studies.", "venue": "NeurIPS", "year": 2019, "title": "Online-Within-Online Meta-Learning"} +{"pid": "3edf1717-5bd4-4ef5-946e-fe41baf7d6d9", "context": "Neural network approaches to pattern recognition are attractive due to their use of a discrimination-based training method. However, traditional multilayer perceptrons do not provide mathematically correct probability distributions, and use a squared error criterion instead of a probability-based score.", "key_idea": "The authors propose modifying the output layer of a multilayer perceptron to provide correct probability distributions, and replacing the squared error criterion with a probability-based score. This results in a Maximum Mutual Information training, which aligns with successful approaches for speech recognition.", "method": "The network construction demonstrates this through recognition computations of a given kind of stochastic model-based classifier.", "outcome": "The authors show how modifying a multilayer perceptron leads to Maximum Mutual Information training, and propose the 'Alphanet', an HMM-based word discriminator as an example.", "future_impact": "The proposed method may lead to the discrimination-based training of parameters of models, improving their performance in applications like speech recognition.", "venue": "NeurIPS", "year": 1989, "title": "Training Stochastic Model Recognition Algorithms as Networks can Lead to Maximum Mutual Information Estimation of Parameters"} +{"pid": "f3edc987-c51a-4352-b32d-7e6f9d8c66ef", "context": "Autonomous agents in real-world environments need to master a variety of skills. While a single short skill can be learned quickly, learning every task independently is impractical. Traditional goal or demonstration conditioned policies, while capable of sharing knowledge between tasks, may fail to generalize to more complex tasks.", "key_idea": "The paper introduces Compositional Plan Vectors (CPVs) that represent trajectories as the sum of the subtasks within them, enabling a policy to perform compositions of tasks without additional supervision.", "method": "The authors implement CPVs within a one-shot imitation learning framework, without the need for additional supervision or information about task hierarchy.", "outcome": "The implementation of CPVs allows a demonstration-conditioned policy to generalize to tasks that sequence twice as many skills as the tasks seen during training. Like word2vec embeddings in NLP, CPVs can support simple arithmetic operations - adding CPVs for two different tasks can command an agent to compose both tasks, with no additional training needed.", "future_impact": "CPVs could revolutionize the way autonomous agents perceive, learn, and execute tasks by enabling task compositionality, potentially leading to more efficient learning and the ability to handle more complex tasks.", "venue": "NeurIPS", "year": 2019, "title": "Compositional Plan Vectors"} +{"pid": "598a8486-44bf-4599-8fda-e3a40c05012d", "context": "The causal notions embodied in the concept of Granger causality have been argued to belong to a different category than those of Judea Pearl's Causal Model, and their relationship has been obscure so far.", "key_idea": "The paper aims to show that the concepts of Granger causality and Pearl's Causal Model are closely linked. This is done by relating each to the notions of direct causality embodied in settable systems, an extension of the Pearl Causal Model.", "method": "They show the relationship between the concepts by providing practical methods to test for direct causality using tests for Granger causality.", "outcome": "The authors successfully demonstrate the close linkage between Granger causality and Pearl's Causal Model by relating each to the notions of direct causality in settable systems.", "future_impact": "The elucidation of the relationship between these two causality concepts and their linkage to direct causality in settable systems could potentially lead to improved understanding and application in areas leveraging causal models.", "venue": "NeurIPS", "year": 2009, "title": "Linking granger causality and the pearl causal model with settable systems"} +{"pid": "5994170b-ba8c-4e7e-a62c-09511714d617", "context": "CNNs are used for the recognition of novel categories from few examples, with their transferability properties inspiring further research.", "key_idea": "The authors introduce an unsupervised meta-training stage that involves exposing multiple top layer units of a CNN to many unlabeled real-world images and an unsupervised margin maximization inferring low-density separators.", "method": "The authors combine unsupervised meta-training with CNNs by incorporating low-density separator (LDS) modules into the top layers of a standard CNN architecture. The effectiveness of the approach is evaluated through tests in scene classification, fine-grained recognition, and action recognition using small training samples.", "outcome": "The incorporation of LDS modules into standard CNN architecture significantly improved performance in scene classification, fine-grained recognition, and action recognition tasks when used with small training samples.", "future_impact": "The proposed approach of combining unsupervised meta-training with CNNs could change how novel categories are recognized using smaller sample sets, thus improving the efficiency in such tasks.", "venue": "NeurIPS", "year": 2016, "title": "Learning from small sample sets by combining unsupervised meta-training with CNNs"} +{"pid": "7d90529c-4a59-49d3-9de2-562f1327f6b8", "context": "In the statistical learning framework, previous results suggested that $O(n)$ random features are needed for ridge regression with random features.", "key_idea": "The authors investigate the possibility of achieving $O(1/\\sqrt{n})$ learning bounds with $O(\\sqrt{n}\\log n)$ random features, rather than $O(n)$ features as suggested by previous research.", "method": "The authors provide theoretical analysis and proofs to quantify the learning rates and the number of random features needed to achieve them.", "outcome": "They show that $O(1/\\sqrt{n})$ learning bounds can be achieved with only $O(\\sqrt{n}\\log n)$ random features and also present conditions under which faster learning rates might require more random features.", "future_impact": "The findings of this study shed light on statistical computational trade-offs in large scale kernelized learning, showing that random features can effectively reduce the computational complexity while preserving optimal generalization properties.", "venue": "NeurIPS", "year": 2016, "title": "Generalization Properties of Learning with Random Features"} +{"pid": "fb589651-4c47-432a-92cd-415d956a475b", "context": "The Perceptron is a theoretical construct relevant for understanding networks of spiking neurons in the brain. However, how the computational power of the Perceptron can be accomplished by the plasticity mechanisms of real synapses is unclear.", "key_idea": "The paper introduces spike-timing-dependent plasticity with an anti-Hebbian form for excitatory synapses, and a spike-timing-dependent plasticity of Hebbian shape for inhibitory synapses, as a solution for realizing the original Perceptron Learning Rule.", "method": "The authors propose a mechanism where the introduced plasticity mechanisms act in concert with the hyperpolarisation of the post-synaptic neurons.", "outcome": "The authors show that with these simple yet biologically realistic dynamics, Tempotrons and Chronotrons are learned. The proposed mechanism also enables incremental associative learning from a continuous stream of patterns.", "future_impact": "The mechanism might be integral to the acquisition of long-term memories in the cortex, thus underscoring the importance of the interactions of synaptic plasticity mechanisms with the dynamics of participating neurons in learning processes.", "venue": "NeurIPS", "year": 2013, "title": "Perfect Associative Learning with Spike-Timing-Dependent Plasticity"} +{"pid": "61a8803c6750f87bf8702007", "context": "ReLU Bayesian neural networks (BNNs) are known to underestimate uncertainty and thus be asymptotically overconfident when dealing with data far away from the training set. This issue arises since the output variance of a BNN with finitely many features is quadratic in the distance from the data region.", "key_idea": "The paper proposes an extension for ReLU BNNs with infinite ReLU features via a Gaussian process (GP) that helps the model to become asymptotically maximally uncertain far away from the data, and this does not affect BNNs' predictive power near the data.", "method": "The authors use a Gaussian process to extend finite ReLU BNNs with infinite ReLU features and investigate the performance of this extended model.", "outcome": "The resulting model approximates a full GP posterior, can mitigate overconfidence in ReLU nets around far-away data, and it can be applied post-hoc to any pre-trained ReLU BNN at a low cost.", "future_impact": "The successful demonstration of this Bayesian treatment may inspire more research into improving uncertainty estimation in BNNs and other machine learning models.", "venue": "NeurIPS", "year": 2021, "title": "An Infinite-Feature Extension for Bayesian ReLU Nets That Fixes Their Asymptotic Overconfidence."} +{"pid": "5f0c34e391e0115455a34c8a", "context": "Existing public video datasets for action recognition capture action videos primarily from a limited number of countries. This bias towards specific countries in data collection is asserted to have an adverse effect on the transferability of the models trained with such datasets.", "key_idea": "The authors introduce Anonymized Videos from Diverse countries (AViD), a new public video dataset for action recognition. AViD is a diverse collection of action videos from many different countries, with properly anonymized face identities to protect privacy.", "method": "The authors run experiments to illustrate the bias in existing datasets and their effect on model transferability. Additionally, they analyze the effectiveness of AViD as a dataset for pretraining models, comparing its performance against prior datasets.", "outcome": "The experiments demonstrate that models trained on biased datasets do not transfer well to other countries. However, models pre-trained on AViD perform comparably or better than those pre-trained on prior datasets, illustrating the superiority of AViD in addressing the issue of transferability.", "future_impact": "The development of the AViD dataset could benefit the training and pretraining of action recognition models universally, rather than being limited to certain countries, potentially improving the global applicability and performance of such models.", "venue": "NeurIPS", "year": 2020, "title": "AViD Dataset: Anonymized Videos from Diverse Countries"} +{"pid": "464cfb36-6c2e-4de9-a143-af35320e723a", "context": "Current understanding of learning observed in real neurons and the principles of self-organization are not well-connected to known biophysical processes. Independent Component Analysis (ICA) has been used to explain the learning of many properties of early visual receptive fields in rate coding models, but this understanding doesn't extend well to spike coding in networks.", "key_idea": "The authors propose a new method that aims to optimize a network sensitivity measure related to maximizing the mutual information between input spike timings and output spike timings in the context of learning observed in real neurons.", "method": "The authors derive a mathematical rule using probabilistic machine learning ideas and simulate its implementation on a single-layer feedforward spiking network to unmix, or demultiplex, mixed spike trains.", "outcome": "In initial simulations, the derived rule was shown to learn synaptic weights which can unmix, or demultiplex, mixed spike trains - it can recover independent point processes embedded in distributed correlated input spike trains.", "future_impact": "Such understanding can potentially lead to further insights into spike coding in networks and how it is supported by cellular biophysical processes in a principled probabilistic manner.", "venue": "NeurIPS", "year": 2004, "title": "Maximising Sensitivity in a Spiking Network"} +{"pid": "1b8e2931-536a-4cc4-97eb-a054cdce8bf2", "context": "The problem of maintaining the data-structures of a partition-based regression procedure in a setting where the training data arrives sequentially over time is a challenging issue.", "key_idea": "The study proposes a new method for maintaining these data structures in a time of O(log n) at any time step n to achieve a nearly optimal regression rate.", "method": "The authors validate this approach through theoretical proofs, using a defined measure of regression performance in terms of the unknown metric dimension d.", "outcome": "The study manages to maintain such a data struture in time O(log n) while achieving a nearly-optimal regression rate of O(n-2/(2+d)). A new regression lower-bound was proved that is independent of a given data size.", "future_impact": "This method could potentially provide a more suitable solution for maintaining the data-structures of a partition-based regression procedure in a streaming setting.", "venue": "NeurIPS", "year": 2013, "title": "Regression-tree Tuning in a Streaming Setting"} +{"pid": "5f7fdd328de39f08283980ad", "context": "Deep learning classifiers are playing crucial roles in decision making and therefore the constant behavior of these models is important for user trust. However, periodic retraining of deployed models can lead to inconsistencies in model outputs over time.", "key_idea": "This paper introduces formal definitions for consistency and correct-consistency of a learning model. The authors theorize that ensembles can improve the consistency and correct-consistency of individual learners, especially if their accuracies are greater than the average accuracy of the ensemble's components.", "method": "The authors validate their theory by testing their proposed dynamic snapshot ensemble method on three datasets and two state-of-the-art deep learning classifiers.", "outcome": "The experiments validate that the ensemble learner demonstrated improved levels of consistency and correct-consistency, compared to individual learners.", "future_impact": "This research provides formal definitions and understanding of model consistency, which could be used in future research to further improve the consistency and predictability of AI models.", "venue": "NeurIPS", "year": 2020, "title": "Wisdom of the Ensemble: Improving Consistency of Deep Learning Models"} +{"pid": "bb280b0b-ee35-4d2e-9dcb-eb3bd1438081", "context": "The problem of segmenting words from phoneme strings and learning the referents of words are two key aspects of language acquisition that are typically handled separately.", "key_idea": "The authors propose Bayesian non-parametric models, an extension of Adaptor Grammars embedded in PCFGs, which can simultaneously segment phoneme sequences into words and learn the relationship between non-linguistic objects to the words that refer to them.", "method": "Two kinds of linguistic information - word segmentation and word-object relationships - are studied through these models, with an examination of the inter-dependencies between the two.", "outcome": "The results show that modelling inter-word dependencies not only improves the accuracy of word segmentation but also word-object relationships, and a model that learns both simultaneously performs better in segmentation than one that only learns word segmentation.", "future_impact": "The results suggest a new interactive view of language acquisition that could harness such synergies for improved learning, inspiring future work along these lines.", "venue": "NeurIPS", "year": 2010, "title": "Synergies in learning words and their referents"} +{"pid": "654d9d86939a5f408273a7ab", "context": "Multiple Instance Learning (MIL) methods have been successfully applied in various domains, such as medical diagnosis based on whole slide images (WSIs). However, performance issues arise when these models are tested on data from different domains, for instance, WSIs from different hospitals.", "key_idea": "The authors introduce the Retrieval-AugMented MIL (RAM-MIL) framework, which integrates Optimal Transport (OT) as the distance metric for nearest neighbor retrieval. The development is driven by insights that reducing the input's intrinsic dimension can minimize the error in attention-based MIL, and there's a link between input intrinsic dimension and feature merging with retrieved data.", "method": "The authors conduct empirical evaluations on whole slide image classification, testing the proposed RAM-MIL framework in both in-domain scenarios, where the training and retrieval data are in the same domain, and out-of-domain scenarios, where the retrieval data comes from a different domain.", "outcome": "The RAM-MIL framework achieves state-of-the-art performance in both in-domain and out-of-domain scenarios. The use of the transportation matrix derived from OT provides interpretable retrieval results at the instance level, allowing visualization for human experts.", "future_impact": "The framework could improve the domain adaptability and interpretability of Multiple Instance Learning models, enhancing their usability across domains.", "venue": "NeurIPS", "year": 2023, "title": "Retrieval-Augmented Multiple Instance Learning"} +{"pid": "39a4570f-1917-4b2f-95ac-c0988e47ebb6", "context": "Semi-supervised methods use both labeled and unlabeled data to construct predictors. Despite showing promising empirical performance, their development is largely based on heuristics.", "key_idea": "This paper rethinks semi-supervised learning from the viewpoint of minimax theory, suggesting that some common methods using graph Laplacians do not lead to faster minimax rates of convergence.", "method": "The authors use the statistical tools of minimax analysis to evaluate existing semi-supervised learning methods and develop newer methods.", "outcome": "The analysis shows that some estimators that use unlabeled data do not have smaller risk than the ones using only labeled data. However, the new approaches developed lead to provably improved performance.", "future_impact": "The study's findings and proposed approaches can provide a new perspective on the problem of semi-supervised learning.", "venue": "NeurIPS", "year": 2007, "title": "Statistical Analysis of Semi-Supervised Regression"} +{"pid": "16bd89ef-a183-4585-92db-cfda71e4aa37", "context": "Standard statistical learning theory typically studies uniform bounds on the expected error, while other approaches like the luckiness framework exploit the serendipity of the training sample.", "key_idea": "The authors present a framework that leverages the specific learning algorithm used, only requiring to cover the functions which could have been learned using the fixed learning algorithm. This differs from standard approaches which cover all hypotheses in a given hypothesis space.", "method": "The authors demonstrate how their framework relates to the VC, luckiness and compression frameworks. They also utilize this framework in an application to the maximum margin algorithm for linear classifiers.", "outcome": "The presented framework results in a bound that takes advantage of both the margin and the distribution of the data in feature space.", "future_impact": "This novel framework with its methodology of exploiting the specific learning algorithm and serendipity of the training sample provides a new way to consider algorithmic learning. Though specific impacts are not mentioned, this could lead to more effective learning models in the future.", "venue": "NeurIPS", "year": 2001, "title": "Algorithmic Luckiness"} +{"pid": "654ce79e939a5f40826b10ec", "context": "Planning in an open-ended world like Minecraft presents significant challenges for multi-task embodied agents because it requires precise, multi-step reasoning due to the long-term nature of tasks. Additionally, current vanilla planning does not consider the achievability of the agent when ordering parallel sub-goals, leading to inefficiencies in the planned actions.", "key_idea": "The authors propose \"Describe, Explain, Plan and Select\" (DEPS), an interactive planning approach based on Large Language Models (LLMs) that includes error correction from feedback during planning and intelligent ordering of sub-goals using a learnable goal Selector module.", "method": "The authors evaluate DEPS using experiments in Minecraft, where it is used to accomplish a variety of tasks. They also test its effectiveness in other non-open-ended domains such as ALFWorld and tabletop manipulation.", "outcome": "The experiments showcase the successful application of DEPS, with it being the first zero-shot multi-task agent that can robustly accomplish over 70 Minecraft tasks, nearly doubling overall performance. Additional testing shows its effectiveness in both ALFWorld and tabletop manipulation tasks. Ablation and exploratory studies further demonstrate DEPS's superiority over other methods.", "future_impact": "Despite not being explicitly stated in the abstract, it seems that due to its successful implementation and superior performance, the DEPS approach could potentially serve as a new basis for future work on planning systems in open-ended environments.", "venue": "NeurIPS", "year": 2023, "title": "Describe, Explain, Plan and Select: Interactive Planning with LLMs Enables Open-World Multi-Task Agents"} +{"pid": "62ecacaa5aee126c0f0b6cf4", "context": "Pre-training models on large-scale datasets has become crucial in deep learning and has achieved success in many downstream tasks in natural language processing and 2D vision. However, translating such a pre-training-tuning paradigm to 3D vision is challenging due to the limited and hard-to-collect training data.", "key_idea": "The paper presents a new approach, Point-to-Pixel (P2P) prompting, which leverages pre-trained 2D knowledge in the 3D domain. Point clouds are transformed into colorful images using geometry-preserved projection and geometry-aware coloring, allowing for the adaption of pre-trained image models.", "method": "The pre-trained image model weights are kept frozen during end-to-end optimization of point cloud analysis tasks. Extensive experiments are conducted to validate the performance of the proposed method in 3D vision tasks such as ScanObjectNN, ModelNet classification, and ShapeNet Part Segmentation.", "outcome": "It's shown that better pre-trained image model will lead to consistently better performance in 3D vision when working with the P2P prompting method. An accuracy of 89.3% is achieved on the hardest setting of ScanObjectNN, surpassing conventional point cloud models with fewer trainable parameters. The method also demonstrates competitive performance on ModelNet classification and ShapeNet Part Segmentation.", "future_impact": "The success of Point-to-Pixel prompting hints at its potential to revolutionize the field of 3D vision by leveraging existing pre-training models and techniques from the field of 2D vision.", "venue": "NeurIPS", "year": 2022, "title": "P2P: Tuning Pre-trained Image Models for Point Cloud Analysis with Point-to-Pixel Prompting"} +{"pid": "e8de0dfd-a528-419f-8e52-a55bdcd1e469", "context": "Guided policy search algorithms are used to optimize complex nonlinear policies, such as deep neural networks, without directly computing policy gradients in the high-dimensional parameter space. These methods use supervised learning to train the policy to mimic a teacher algorithm, like a trajectory optimizer or a trajectory-centric reinforcement learning method. However, there is no clarity on how much the policy improves within a finite number of iterations. Plus, in existing methods, the projection onto the constraint manifold is not exact.", "key_idea": "The paper proposes that guided policy search algorithms can be interpreted as an approximate variant of mirror descent. Additionally, they introduce a new guided policy search algorithm that simplifies the process and provides improvement and convergence guarantees in simplified convex and linear settings.", "method": "The authors derive a new guided policy search algorithm and provide empirical results on several simulated robotic manipulation tasks.", "outcome": "The results show that their method is stable and achieves similar or better performance when compared to prior guided policy search methods, with a simpler formulation and fewer hyperparameters.", "future_impact": "The provision of improvement and convergence guarantees in simplified convex and linear settings represents a significant step towards more reliable guided policy search algorithms.", "venue": "NeurIPS", "year": 2016, "title": "Guided Policy Search as Approximate Mirror Descent"} +{"pid": "658b1fd9-e5da-48b6-be4a-ce103c677c10", "context": "Unsupervised learning methods for object detection and parsing typically struggle with being invariant to the scale and rotation of the objects.", "key_idea": "The authors propose an unsupervised method to learn a probabilistic grammar of an object from a set of training examples, which is invariant to scale and rotation.", "method": "The authors illustrate the method using thirteen objects from the Caltech 101 database and a hybrid class composed of faces, motorbikes, and airplanes. The method's performance is compared with alternative approaches via training datasets and testing datasets.", "outcome": "The proposed method delivers quick inference (under one second), accurate performance, and the parsing of the object. Individual objects can be recovered as different aspects of the grammar for the object class, including from a hybrid class where the specific object or its position, scale, or pose is unknown.", "future_impact": "The authors suggest that their approach is very general and can be potentially applied to a large range of objects and structures.", "venue": "NeurIPS", "year": 2006, "title": "Unsupervised Learning of a Probabilistic Grammar for Object Detection and Parsing"} +{"pid": "91e871fd-84f8-4c3a-a3a1-7e5230612720", "context": "Stochastic gradient descent is popular for large scale optimization but has slow convergence asymptotically due to the inherent variance.", "key_idea": "The authors introduce an explicit variance reduction method for stochastic gradient descent called stochastic variance reduced gradient (SVRG). Unlike the current methods, it does not require the storage of gradients, making it more applicable to complex problems.", "method": "The authors perform theoretical analysis on the proposed method for smooth and strongly convex functions.", "outcome": "The authors prove that SVRG enjoys the same fast convergence rate as those of stochastic dual coordinate ascent (SDCA) and Stochastic Average Gradient (SAG). The analysis was found to be simpler and more intuitive.", "future_impact": "The lack of requirement for gradient storage in SVRG makes it potentially more applicable to complex problems such as some structured prediction problems and neural network learning.", "venue": "NeurIPS", "year": 2013, "title": "Accelerating Stochastic Gradient Descent using Predictive Variance Reduction"} +{"pid": "eebf58b1-d7ab-4ee8-8e84-70861c652412", "context": "The efficient coding hypothesis suggests that neural receptive fields are adapted to the environment's statistics. This adaptation occurs on both evolutionary and developmental timescales.", "key_idea": "The authors focus on adaptation occurring during an organism's lifetime and examine several unsupervised feature learning algorithms that can explain receptive field properties in multiple primary sensory cortices.", "method": "The authors tested various unsupervised feature learning algorithms in portraying normal and altered receptive field properties as driven by experimentally modified environmental statistics.", "outcome": "The unsupervised feature learning algorithms were successful in demonstrating both normal and altered receptive field properties in response to changes in environmental statistics.", "future_impact": "The authors propose that these learning algorithms can serve to realize Mountcastle's original theory that a similar learning algorithm operates throughout primary sensory cortices, thereby forming phenomenological models of receptive field plasticity throughout an organism's lifetime.", "venue": "NeurIPS", "year": 2011, "title": "Unsupervised learning models of primary cortical receptive fields and receptive field plasticity"} +{"pid": "760a82b8-5636-43e2-bd6c-4072b8a6ebc9", "context": "In machine learning, overfitting is a common issue which results in reduced predictive performance on unseen data. The particular behavior that leads to overfitting in a perceptron learning task with a high degree of nonlinearity, in which the task is generated by another perceptron with identical architecture and nonlinear sigmoid output functions, is not thoroughly understood.", "key_idea": "This study observes and seeks to explain why a higher level of nonlinearity leads to overfitting, and it proposes a new method to avoid overfitting in a perceptron learning task.", "method": "The perceptrons, both with nonlinear sigmoid output functions, are examined and their behaviours observed in scenarios with varying levels of nonlinearity.", "outcome": "This study finds that a high level of nonlinearity leads to overfitting in this perceptron learning task.", "future_impact": "The newly proposed method to prevent overfitting could be understood and interpreted in two ways - learning with noise or cross-validated early stopping, potentially expanding its applications and implications in the broader context of machine learning.", "venue": "NeurIPS", "year": 1995, "title": "A Realizable Learning Task which Exhibits Overfitting"} +{"pid": "634e194190e50fcafd24e7df", "context": "Offline reinforcement learning (RL) uses historical data to improve future performance, but the algorithm and hyperparameters chosen can significantly impact the performance of decision policies. Current practices lack a systematic pipeline for these selections in real-world situations where only historical data is available.", "key_idea": "The authors introduce a systematic pipeline for training, comparing, selecting, and deploying the best policy in offline RL settings with limited datasets. Emphasis is given to the importance of multiple data splits for reliable algorithm-hyperparameter selection, a common approach in supervised learning but not thoroughly explored in offline RL.", "method": "The efficiency of the pipeline is demonstrated using a variety of offline RL algorithms across simulation domains in healthcare, education, and robotics.", "outcome": "The pipeline outperformed alternate approaches by producing higher-performing policies across various domains, highlighting the benefit of using multiple data splits, especially with small datasets.", "future_impact": "This work contributes to the development of a general-purpose meta-algorithm for automatic algorithm-hyperparameter selection in offline RL, potentially improving decision-making processes in various real-world applications.", "venue": "NeurIPS", "year": 2022, "title": "Data-Efficient Pipeline for Offline Reinforcement Learning with Limited Data"} +{"pid": "61a884096750f82b17638b2c", "context": "Differentiable representations of graphs or taxonomies are useful for a variety of machine learning tasks. Although vectors in Euclidean space can theoretically represent any graph, alternatives like complex, hyperbolic, order, or box embeddings are said to have better-suited geometric properties for modeling real-world graphs. However, these benefits are only observed in lower dimensions, with performance diminishing in higher dimensions.", "key_idea": "The authors introduce a novel box embeddings variant that uses a learned smoothing parameter. This variant is designed to achieve better representational capacity than vector models in low dimensions while avoiding performance saturation common to other geometric models in high dimensions.", "method": "The authors perform empirical evaluations of vector, hyperbolic, and region-based geometric representations on several families of synthetic and real-world directed graphs.", "outcome": "The novel variant of box embeddings demonstrated better representational capacity than vector models in lower dimensions while avoiding performance degradation in high dimensions. The results revealed correlations between different families of graphs, graph characteristics, model size, and embedding geometry, providing insights into the inductive biases of various differentiable graph representations.", "future_impact": "The findings could influence how graph representations are chosen and utilized for machine learning tasks, potentially leading to more effective models.", "venue": "NeurIPS", "year": 2021, "title": "Capacity and Bias of Learned Geometric Embeddings for Directed Graphs."} +{"pid": "224fd7ac-7176-4795-93c9-d77061a02b90", "context": "The proximal gradient method (PGM) is popularly used to minimize a convex smooth loss function with trace norm regularization in machine learning applications. However, this method is known to have a sublinear rate of convergence, regardless of the loss function type.", "key_idea": "This paper proposes that for a large class of loss functions, the convergence rate of the proximal gradient method is actually linear, not sublinear, even without any strong convexity assumption on the loss function.", "method": "A new Lipschitzian error bound for the trace norm-regularized problem is utilized in establishing the linear convergence rate of the proximal gradient method.", "outcome": "It was proven that the convergence rate of the proximal gradient method, for a large class of loss functions, is linear.", "future_impact": "This finding regarding the linear convergence rate of the proximal gradient method could have implications in various machine learning applications. Additionally, the introduced Lipschitzian error bound may offer value in other contexts and problems.", "venue": "NeurIPS", "year": 2013, "title": "On the Linear Convergence of the Proximal Gradient Method for Trace Norm Regularization"} +{"pid": "6184a0d25244ab9dcb28bf89", "context": "Sampling from an unnormalized probability distribution, a fundamental problem in machine learning, is typically addressed using variations of Markov Chain Monte Carlo (MCMC), which suffer from slow convergence. Auxiliary neural models can accelerate MCMC, but their training often involves high overhead.", "key_idea": "The authors propose a different approach, the Energy Sampling Hamiltonian (ESH) dynamics, which is a new deterministic Hamiltonian dynamics with non-Newtonian momentum. This dynamics model does not require a stochastic step differing from MCMC approaches and can be interpreted as a normalizing flow that samples an energy model without training.", "method": "The ESH dynamics are implemented using existing ODE solvers, and the performance is compared with that of a specialized solver developed by the authors. The dynamics are assessed based on the speed of convergence in training neural network energy models.", "outcome": "The proposed ESH dynamics are easier to solve when compared with existing ODE solvers. ESH dynamics demonstrate faster convergence than their MCMC counterparts, enabling more efficient, stable training of neural network energy models.", "future_impact": "The introduction of ESH dynamics could lead to a shift in the prevailing methods for sampling from unnormalized probability distributions, potentially replacing MCMC variants, and expedite the training of neural network energy models.", "venue": "NeurIPS", "year": 2021, "title": "Hamiltonian Dynamics with Non-Newtonian Momentum for Rapid Sampling."} +{"pid": "629ec1f95aee126c0fb7047c", "context": "In the space of recommendation models, data quality, heterogeneity, and sparsity are critical problems that have not been fully addressed, especially when dealing with extremely large and sparse user-item interaction matrices.", "key_idea": "The authors propose two concepts, $\\infty$-AE, an autoencoder with infinitely-wide bottleneck layers using the Neural Tangent Kernel, and Distill-CF, a method for creating highly efficient and accurate data summaries from the user-item interaction matrix for subsequent model training, inference, and architecture search.", "method": "The authors use the concept of differentiable Gumbel-sampling to handle the inherent data heterogeneity, sparsity, and semi-structuredness of large datasets. They validate their proposed model on datasets containing hundreds of millions of user-item interactions.", "outcome": "The proposed approaches significantly outperformed their respective state-of-the-arts and when used together, achieved between 96% and 105% of $\\infty$-AE's performance on the full dataset using just 0.1% of the original dataset.", "future_impact": "The success of the proposed approach opens up an exploration into the counter-intuitive question of whether more data is needed for better recommendation, potentially impacting how future recommendation systems are built.", "venue": "NeurIPS", "year": 2022, "title": "Infinite Recommendation Networks: A Data-Centric Approach"} +{"pid": "634f6ae490e50fcafdcb66a1", "context": "The recently-developed Associating Objects with Transformers (AOT) approach introduces hierarchical propagation into Video Object Segmentation (VOS) and has shown promising results. However, the increase of object-specific information leads to the loss of object-agnostic visual information in deep propagation layers.", "key_idea": "The authors propose the Decoupling Features in Hierarchical Propagation (DeAOT) method that decouples the hierarchical propagation of object-agnostic and object-specific embeddings by processing them in two independent branches. They also propose an efficient module for constructing hierarchical propagation, the Gated Propagation Module, specially designed with single-head attention.", "method": "The method is applied and validated by extensive experiments to compare the performance of AOT and DeAOT on benchmarks such as YouTube-VOS, DAVIS 2017, DAVIS 2016, and VOT 2020.", "outcome": "DeAOT significantly outperforms AOT in both accuracy and efficiency, achieving new state-of-the-art performance on four benchmarks (YouTube-VOS (86.2%), DAVIS 2017 (86.2%), DAVIS 2016 (92.9%), and VOT 2020 (0.622)).", "future_impact": "This new method of decoupling hierarchy features could potentially offer higher accuracy and efficiency in video object segmentation tasks and transform related research trends.", "venue": "NeurIPS", "year": 2022, "title": "Decoupling Features in Hierarchical Propagation for Video Object Segmentation"} +{"pid": "5f7fdd328de39f0828397f27", "context": "Deep reinforcement learning (RL) agents trained in a limited set of environments tend to overfit and fail to generalize to unseen testing environments. Data augmentation approaches like cutout and random convolution that were previously used to improve generalizability showed limited effectiveness due to local perturbation of the observations.", "key_idea": "The authors propose a new method, called mixreg, which trains agents on a mixture of observations from different training environments and imposes linearity constraints on the observation interpolations and the supervision (e.g. associated reward) interpolations, which effectively increase data diversity and help in learning smoother policies.", "method": "The authors tested the generalization of mixreg by conducting extensive experiments on the large-scale Procgen benchmark.", "outcome": "The results show that mixreg outperforms the well-established baselines on unseen testing environments by a large margin.", "future_impact": "Mixreg can be applied to both policy-based and value-based RL algorithms, indicating broad-applicability and potential for future advancements in generalization of reinforcement learning algorithms.", "venue": "NeurIPS", "year": 2020, "title": "Improving Generalization in Reinforcement Learning with Mixture Regularization"} +{"pid": "468ebd04-c374-414d-8a72-779a48e8d5df", "context": "Deep autoregressive sequence-to-sequence models have shown impressive performance in various tasks. However, inference for new inputs remains an inherently sequential process, despite different architecture classes providing trade-offs between computation amount needed per layer and the length of the critical path during training.", "key_idea": "The paper introduces a blockwise parallel decoding scheme, which leverages the ability of some architectures to score sequences in sublinear time, allowing multiple time steps to be predicted at once and then reverting to the longest prefix confirmed by the scoring model.", "method": "The authors tested their approach on state-of-the-art self-attention models for machine translation and image super-resolution, comparing it with a baseline greedy decoder.", "outcome": "The proposed approach achieves up to 2x reductions in iterations over a baseline greedy decoder, with no loss in quality. Fine-tuning and relaxing the acceptance criteria boost the reduction up to 7x, albeit with a slight decline in performance. The fastest models achieve a 4x speedup in wall-clock time.", "future_impact": "This work could influence future development of more efficient and faster algorithms for autoregressive models.", "venue": "NeurIPS", "year": 2018, "title": "Blockwise Parallel Decoding for Deep Autoregressive Models"} +{"pid": "d65292a6-71c2-4adb-98a0-1c4f90a5fad1", "context": "Identifying multiway block structure from large noisy tensors is a frequent challenge in applications such as genomics, recommendation systems, topic modeling, and sensor network localization. Existing methods are not optimal.", "key_idea": "The paper presents the Tensor Block Model and introduces a unified least-square estimation for multiway clustering - issues frequently faced in many practical fields.", "method": "Statistical convergence of the estimator is inspected, along with the accuracy of the associated clustering procedure. Regularization is used for identifying elevated blocks. The method is tested through simulation and application to two real datasets.", "outcome": "The paper proves the estimator's statistical convergence and shows that the clustering procedure achieves partition consistency. Several tests using simulations and real datasets indicate the proposed model performs better than previous methods.", "future_impact": "The proposal handles a broad range of data types and could potentially enhance the performance of various applications including genomics, recommendation systems, topic modeling, and sensor network localization.", "venue": "NeurIPS", "year": 2019, "title": "Multiway clustering via tensor block models"} +{"pid": "65262272939a5f4082badb86", "context": "Spiking Neural Networks (SNNs) have demonstrated potential for 2D visual recognition due to their extreme energy efficiency. However, the application of SNNs to 3D recognition, specifically point clouds, has not been thoroughly investigated.", "key_idea": "The study proposes Spiking PointNet, the first spiking neural model designed for efficient deep learning on point clouds. The paper identifies and addresses two major challenges hindering the application of SNNs in point clouds: the intrinsic optimization obstacle of SNNs and the expensive memory and computation cost of PointNet.", "method": "The authors present a 'trained-less but learning-more' paradigm for Spiking PointNet, training it with only a single time step but achieving better performance with multiple time step inferences. The effectiveness of Spiking PointNet is demonstrated through various experiments on ModelNet10 and ModelNet40 datasets.", "outcome": "The Spiking PointNet demonstrates superior performance compared to results obtained from direct multiple time step training. Notably, it outperforms its ANN counterpart, a rarity in the SNN field. Spiking PointNet also shows impressive speedup and storage saving during the training phase.", "future_impact": "The effectiveness and efficiency of Spiking PointNet may guide future research in the SNN field towards a more extensive application of SNNs in 3D recognition.", "venue": "NeurIPS", "year": 2023, "title": "Spiking PointNet: Spiking Neural Networks for Point Clouds"} +{"pid": "e1175ce1-c72a-4323-bdb4-a5b874be8b36", "context": "Prior to this work, time-series modeling of video traffic was primarily tackled using time domain methods.", "key_idea": "This paper introduces a new approach to time-series modeling of video traffic by modeling the wavelet coefficients in the wavelet domain rather than the time domain.", "method": "The authors build the model in the wavelet domain to capture both long-range and short-range dependencies, and use it to generate high quality video traffic. They then perform a performance analysis.", "outcome": "The study confirms that the wavelet model is a unified and computationally efficient method for developing the model and generating high quality video traffic.", "future_impact": "The proposed wavelet domain model has the potential to enable more sophisticated and efficient video traffic analysis.", "venue": "NeurIPS", "year": 1997, "title": "Wavelet Models for Video Time-Series"} +{"pid": "619d7d496750f82b1e89fff1", "context": "Predicting event memorability, that is, to what extent an event is recalled, is a significant challenge in memory research with implications for artificial intelligence.", "key_idea": "The authors examine whether event memorability relies on the event's context, as well as the intrinsic visual attributes of image cues. They propose a novel concept of a contextual event memory network (CEMNet) to predict event memorability.", "method": "A novel experiment protocol was designed, and a large-scale experiment was conducted with 47 elder subjects over 3 months. The subjects' memory of life events was tested in a cued recall process. Advanced visual analytics methods were used to build a first-of-its-kind event memorability dataset (R3) with detailed information about event context and visual semantic features.", "outcome": "The proposed CEMNet, which handles multi-modal input to predict item-wise event memorability, outperforms competitive benchmarks.", "future_impact": "The findings provide a deeper understanding of episodic event memory, and establish a new avenue for predicting human episodic memory.", "venue": "NeurIPS", "year": 2021, "title": "Predicting Event Memorability from Contextual Visual Semantics"} +{"pid": "63a413f790e50fcafd6d214b", "context": "Existing research in the field of non-negative rank-$K$ matrix modelling has mostly been focused on the case where all eigenvalues are non-negative ($m = 0$). Additionally, exploring when a rank-$K$ model can be rewritten as a DCMM (Discrete Curvilinear Motion Model) is another interest area due to their inherent interpretability and usefulness in practice.", "key_idea": "The key idea for this study is to extend the existing Non-negative Matrix Factorization (NMF) problem to the case where general $m$ eigenvalues are negative, using the novelty of proof ideas developed in this paper.", "method": "The researchers use new proof ideas to develop sharp results for solving the NMF problem in the general $m$ case. Additionally, they use these NMF results to analyze when a rank-$K$ model can be rewritten as a DCMM model, focusing on parameters in the most interesting range.", "outcome": "The study successfully presents extended results on the solvability of the NMF problem, and shows that any rank-$K$ model with parameters in the most interesting range can be rewritten as a DCMM model.", "future_impact": "The findings could influence future work in non-negative matrix modelling and can be instrumental in the development and interpretation of more effective rank-$K$ models, such as DCMM, in network modeling.", "venue": "NeurIPS", "year": 2022, "title": "A sharp NMF result with applications in network modeling"} +{"pid": "632bd2a990e50fcafdb7a283", "context": "Federated Learning (FL) allows decentralized clients to learn collaboratively without sharing their private data, but the excessive computation and communication demands pose challenges to current FL frameworks, especially when training large-scale models.", "key_idea": "The authors propose a lightweight framework in which clients jointly learn to fuse the representations generated by multiple fixed pre-trained models instead of training a large-scale model from scratch. Moreover, a Federated Prototype-wise Contrastive Learning (FedPCL) approach is designed that shares knowledge across clients through their class prototypes.", "method": "The authors perform a thorough evaluation of FedPCL in the lightweight framework, measuring and visualizing its ability to fuse various pre-trained models on popular FL datasets.", "outcome": "The evaluation reveals that the proposed strategy can efficiently fuse different pre-trained models and build client-specific representations in a prototype-wise contrastive manner, maintaining shared knowledge in a compact form for efficient communication.", "future_impact": "This proposed lightweight framework and FedPCL approach could foster the improvement and deployment of Federated Learning systems, particularly by addressing computation and communication challenges.", "venue": "NeurIPS", "year": 2022, "title": "Federated Learning from Pre-Trained Models: A Contrastive Learning Approach"} +{"pid": "60bc6b77-f3ba-4ae3-9ede-161bb2852bd3", "context": "Minimizing submodular functions subject to ordering constraints is typically treated as a non-convex optimization problem.", "key_idea": "The authors offer a novel approach to this optimization problem by treating it as a convex optimization problem on a space of uni-dimensional measures, introducing first-order stochastic dominance as the corresponding ordering constraint. They also propose new discretization schemes.", "method": "The authors developed algorithms based on zero-th, first, or higher order oracles, and conducted experiments testing these algorithms and their robustness to outliers in the context of isotonic regression.", "outcome": "The authors' experiments demonstrated that their approach using non-convex loss functions could be more robust to outliers in isotonic regression and still be solved in polynomial time.", "future_impact": "The novel approach and algorithms proposed could potentially lead to more efficient solutions for non-convex isotonic regression and other optimization problems.", "venue": "NeurIPS", "year": 2018, "title": "Efficient Algorithms for Non-convex Isotonic Regression through Submodular Optimization"} +{"pid": "63a413f790e50fcafd6d23b2", "context": "Deep learning technologies have been widely used in Electronic Design Automation (EDA) to assist circuit design, which often encounters heterogeneous and diverse information sources, including the netlist and the design layout. However, handling each information source independently has been found to be sub-optimal.", "key_idea": "The authors propose a novel way to integrate multiple information sources under a unified heterogeneous graph named Circuit Graph, where topological and geometrical information is well integrated. They also propose Circuit GNN to fully utilize the features of vertices, edges as well as heterogeneous information during the message passing process.", "method": "Experiments on the two most representative prediction tasks in EDA, logic synthesis and global placement chip design stages, were conducted to validate the effectiveness of Circuit GNN.", "outcome": "The solution proposed in this paper reaches state-of-the-art performance in both logic synthesis and global placement chip design stages. It also achieves a 10x speed-up on congestion prediction compared to the state-of-the-art model.", "future_impact": "The paper marks the first attempt to design a versatile circuit representation that is compatible across multiple EDA tasks and stages, which has potential for future development and improvements in EDA tasks.", "venue": "NeurIPS", "year": 2022, "title": "Versatile Multi-stage Graph Neural Network for Circuit Representation"} +{"pid": "4a7b03d4-0052-431a-a869-43484547a577", "context": "Elicitation, the study of statistics or properties computable via empirical risk minimization, has been extensively studied to determine which properties are elicitable. However, since all properties can be elicitable by first eliciting the entire distribution or dataset, the premise of this question has been disputed.", "key_idea": "This study proposes a different approach, focusing on discerning 'how elicitable' a property is, and introducing a new concept of 'elicitation complexity'.", "method": "The authors build on existing work to develop a 'calculus of elicitation'. They established methods and techniques to compute upper and lower bounds on elicitation complexity.", "outcome": "The work resulted in several general results and techniques necessary to prove upper and lower bounds on elicitation complexity, which allowed for eliciting the Bayes risk of any loss, a comprehensive class of properties including spectral risk measures and several new properties of interest.", "future_impact": "The work lays the groundwork for the continued exploration of calculus of elicitation and can potentially improve elicitation methods by focusing on the complexity of elicitation, thus serving as a foundation for forthcoming studies on elicitation.", "venue": "NeurIPS", "year": 2015, "title": "On elicitation complexity"} +{"pid": "6295873d5aee126c0fe142e6", "context": "Symmetry is a useful tool in the analysis of complex systems, and has been explored in machine learning models and data.", "key_idea": "This study seeks to link symmetries arising from the architecture of a model family with the symmetry in the family's internal representation of data, through the determination of fundamental symmetry groups termed 'intertwiner groups' of the model.", "method": "The connection between intertwiner groups and a model's internal data representations is explored via experiments that probe the similarities between hidden states across models with identical architecture.", "outcome": "The study finds that a network's symmetries are propagated into the symmetries of the network's data representation, improving understanding of how architecture influences the learning and prediction process.", "future_impact": "The authors speculate that for ReLU networks, the intertwiner groups may justify the prevalent practice of focusing model interpretability exploration on activation basis in hidden layers instead of on arbitrary linear combinations.", "venue": "NeurIPS", "year": 2022, "title": "On the Symmetries of Deep Learning Models and their Internal Representations"} +{"pid": "7e352416-3b15-48d3-b9a2-d0127100ca69", "context": "The maximum mean discrepancy (MMD) is an integral probability metric defined for a reproducing kernel Hilbert space (RKHS), serving as a metric on probability measures for a sufficiently rich RKHS. Despite its utility, its convergence properties and optimization methods were not fully understood.", "key_idea": "The authors propose the construction of a Wasserstein gradient flow of the MMD and investigate its convergence properties, providing conditions for convergence towards a global optimum.", "method": "The study involves the formulation and analysis of the MMD flow with a regularization technique, which involves an injection of noise in the gradient. The implementation is executed using simple closed-form expressions, which are easily estimated with samples.", "outcome": "On an empirical and theoretical basis, the authors found that the proposed regularized MMD flow converges towards a global optimum, and its application could be related to particle transport when optimizing neural networks.", "future_impact": "The algorithmic solution and its practical implementation provide a foundational resource that could better guide future applications of MMD in fields like neural network optimization.", "venue": "NeurIPS", "year": 2019, "title": "Maximum Mean Discrepancy Gradient Flow"} +{"pid": "5f7fdd328de39f0828397bea", "context": "The classic model for the influence maximization problem assumes that all influence relationships in the network are independent.", "key_idea": "The authors propose a distributionally robust model for the influence maximization problem. Unlike the classical model, this model's diffusion process is adversarially adapted to the choice of seed set and aims to maximize the 'worst-case, expected influence'.", "method": "The authors use computations and numerical experiments to contrast their proposed model with the classic independent cascade model.", "outcome": "The authors show that the worst-case influence can be computed efficiently, and though the optimization is NP-hard, a ($1 - 1/e$) approximation guarantee holds. They also provide insights from numerical experiments comparing the adversarial and independent cascade model.", "future_impact": "The paper presents the extent to which the independence assumption may cost optimality in the influence maximization problem, providing future significant insights into adversarial models and influence maximization strategies.", "venue": "NeurIPS", "year": 2020, "title": "Correlation Robust Influence Maximization"} +{"pid": "60517ba0-1b6e-425c-ba06-579569e9677c", "context": "Previous approaches to neural networks do not explicitly represent them as geometric transformations acting on the coordinate representation of the underlying data manifold from which the data is sampled.", "key_idea": "This study proposes to construct a general theory of neural networks in the setting of Riemannian geometry, stating that residual neural networks are finite difference approximations to dynamical systems of first order differential equations, and a closed form solution of the metric tensor on the underlying data manifold can be found by using the neural network.", "method": "Theoretical results are developed and proven for feedforward networks. Toy experiments are run to confirm parts of the proposed theory and provide intuitions on how neural networks operate on data.", "outcome": "The study proves that residual neural networks are the finite differences of first order differential equations, and it's possible to find a closed form solution of the metric tensor on the underlying data manifold by backpropagating the coordinate representations learned by the neural network.", "future_impact": "This study provides a new perspective on understanding neural networks as geometric transformations and might influence the ways neural networks are designed, trained, and understood.", "venue": "NeurIPS", "year": 2017, "title": "Principles of Riemannian Geometry in Neural Networks"} +{"pid": "43c012e1-01a3-44a6-8268-9e8f92af325c", "context": "Stimulus arrays are inevitably presented at different positions on the retina in visual tasks, even those that nominally require fixation. In particular, this applies to many perceptual learning tasks.", "key_idea": "The authors argue that perceptual inference or discrimination in the face of positional variance is structurally different from inference about fixed position stimuli, involving a specific, quadratic, non-linearity rather than a purely linear discrimination.", "method": "The authors demonstrate the superior discrimination performance of a recurrent network, suggesting it as a role for recurrent connections in area VI.", "outcome": "The superior performance of the recurrent network is demonstrated, implying the advantage of taking into account the quadratic non-linearity for discrimination in the face of positional variance.", "future_impact": "The authors propose that learning the feedforward and recurrent neural connections for these tasks corresponds to the fast and slow components of learning and this could potentially impact perceptual learning tasks.", "venue": "NeurIPS", "year": 2000, "title": "Position Variance, Recurrence and Perceptual Learning"} +{"pid": "1621afe6-6af9-402d-b31f-058508b56b5d", "context": "Convolutional neural networks (CNNs) trained for object classification have been used to predict neural responses in human and primate brains. However, due to the complexity of object classification, it is difficult to infer specific aspects of neural information processing using CNN-derived representations.", "key_idea": "This paper proposes encoding models based on learned representations derived from 21 computer vision tasks to predict brain responses from a large dataset of fMRI scans (BOLD5000). The key idea includes construction of a 'task graph' based on the spatial layout of brain areas predicted well by each task.", "method": "The authors used 21 computer vision tasks to develop encoding models for predicting brain responses, tested on the BOLD5000 dataset. The 'task graph' is constructed from result data across all 21 task representations and is compared to task structure derived from transfer learning accuracy.", "outcome": "It was found that models based on task features can predict activity in different brain regions. Features from 3D tasks like keypoint/edge detection explained a larger variance compared to 2D tasks. Tasks with higher transferability made similar predictions for brain responses from different regions.", "future_impact": "These results, enabled by state-of-the-art computer vision methods, help in understanding the task-specific architecture of the human visual system. This understanding might unlock new areas of research in both computer vision and neural information processing.", "venue": "NeurIPS", "year": 2019, "title": "Neural Taskonomy: Inferring the Similarity of Task-Derived Representations from Brain Activity"} +{"pid": "615160026750f8399ecd63ac", "context": "Kernel selection is crucial in determining the performance of Gaussian Process (GP) models, and constructing custom kernel functions for high-dimensional GP regression models is challenging.", "key_idea": "Introduced is a new approach called KITT (Kernel Identification Through Transformers), which uses a transformer-based architecture to rapidly generate kernel recommendations for GP regression models, significantly faster than traditional kernel search algorithms.", "method": "KITT is trained on synthetic data generated from priors over a vocabulary of known kernels. Its performance is then tested on a diverse range of regression benchmarks.", "outcome": "Kernels chosen by KITT demonstrated strong performance on various regression benchmarks.", "future_impact": "As KITT offers rapid kernel selection and yields strong performance on regression benchmarks, it could enhance the practical application of Gaussian Process models in handling complex high-dimensional data.", "venue": "NeurIPS", "year": 2021, "title": "Kernel Identification Through Transformers."} +{"pid": "0b4d09bb-045c-476c-9f48-f3a3c9759f20", "context": "Deep learning can predict quantum-chemical properties quickly and accurately, which aids in the discovery of new molecules and materials. However, existing generative machine learning models, which guide exploration towards desired properties, are limited. Specifically, graph-based models lack spatial information and thus can't recognize spatial isomerism and non-bonded interactions.", "key_idea": "The authors propose a generative neural network model for 3D point sets that respects the rotational invariance of targeted structures and can capture complex relationships between 3D geometry and electronic properties.", "method": "The model is applied to the generation of molecules, and is tested by comparing the generated set to the distribution of equilibrium structures using spatial metrics, as well as established metrics from chemoinformatics.", "outcome": "The model was successful in approximating the distribution of equilibrium structures and showed its ability to capture the complex relationship between 3D geometry and electronic properties. Additionally, the distribution of the generator was biased towards molecules with a small HOMO-LUMO gap, a key property for the design of organic solar cells.", "future_impact": "This novel approach to creating generative models for chemical structures holds the potential to enhance the discovery process of molecules with desired properties, specifically for applications such as the design of organic solar cells.", "venue": "NeurIPS", "year": 2019, "title": "Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules"} +{"pid": "6ddb8fd2-ef41-45f5-9b2a-514dc262f947", "context": "The Stochastic Block Model (SBM) is a commonly used random graph model for networks with communities. Despite much research, the fundamental limits and understandings of recovery in the SBM are still not fully understood.", "key_idea": "The authors dive into the SBM in its full generality, considering no restrictions on the number, sizes of communities or how they grow, as well the connectivity probabilities. By exploiting the tradeoffs among the various parameters, they aim to provide exact recoverability.", "method": "The authors provide guarantees for exact recovery in the heterogeneous SBM via a semidefinite program, and they also provide upper and lower bounds on SBM parameters.", "outcome": "The paper provides novel recovery guarantees for many new interesting SBM configurations, exploiting the tradeoffs among the various parameters of the heterogeneous SBM.", "future_impact": "The results could potentially shed light on further understandings of the stochastic block model and its implications in various community detection efforts.", "venue": "NeurIPS", "year": 2016, "title": "Exploiting tradeoffs for exact recovery in heterogeneous stochastic block models"} +{"pid": "20488a95-bcf5-474d-8399-3cd6e036760e", "context": "Existing models for learning to detect and recognize objects with minimal supervision do not explicitly capture uncertainty in the number of object instances depicted in a given image.", "key_idea": "A hierarchical probabilistic model called the transformed Dirichlet process (TDP) is proposed. This model captures uncertainty in the number of object instances and uses stochastically transformed mixture components which are shared between multiple groups of data.", "method": "The TDP model is validated through learning and inference based on an empirically effective Gibbs sampler. It is applied to a dataset of partially labeled street scenes.", "outcome": "The inclusion of spatial structure in the TDP improves detection performance and flexibly exploits partially labeled training images.", "future_impact": "The TDP, which provides a novel way to model uncertainty in the number of object instances depicted in a given image, has potential applications beyond the field of computer vision.", "venue": "NeurIPS", "year": 2005, "title": "Describing Visual Scenes using Transformed Dirichlet Processes"} +{"pid": "cdf977d6-5a38-4ae1-b213-e9523841a543", "context": "Many measurement modalities like nuclear magnetic resonance (NMR) spectroscopy provide rich data but are still reliant on laborious human interpretation. In NMR spectroscopy, the observed spectrum provides a distinguishing fingerprint of the molecule's bond structure, resulting in an inverse problem of inferring the chemical structure from the molecular formula and spectrum.", "key_idea": "The authors propose to solve this inverse problem by manipulating it as a graph-structured prediction problem, where they intend to predict the edges and edge types given per-vertex information on a subset of the vertices, and treat this as a Markov decision process (MDP).", "method": "The method involves constructing molecules one bond at a time, leveraging a deep neural network trained via imitation learning, where the network learns to imitate a subisomorphic oracle which knows the remaining correct bonds.", "outcome": "The method is found to be fast, accurate and the first among recent chemical-graph generation approaches to use per-vertex information and generate graphs with vertex constraints. It can quickly compute the correct molecular structure for a wide variety of molecules, and can detect with reasonable certainty when it cannot.", "future_impact": "This method can lead towards automation of molecular structure identification and potentially brings future opportunities for active learning in spectroscopy.", "venue": "NeurIPS", "year": 2019, "title": "Deep imitation learning for molecular inverse problems"} +{"pid": "603f6de391e011cacfbda3c4", "context": "Label-imbalanced and group-sensitive classification aim to optimize relevant metrics like balanced error and equal opportunity. Classical methods, such as weighted cross-entropy, struggle in training deep nets to the terminal phase of training (TPT), resulting in a recent influx of heuristic alternatives that promote larger margins for minorities.", "key_idea": "The authors develop a principled analysis for understanding how loss adjustments affect margins during training. They propose a novel approach called the vector-scaling (VS) loss that incorporates different techniques for handling label-imbalance and group-sensitivity in a unified way.", "method": "The authors conduct experiments on state-of-the-art datasets; and perform a generalization analysis on imbalanced Gaussian-mixture data.", "outcome": "The authors' experiments align with their theoretical insights and demonstrate the superior performance of their proposed algorithms. A tradeoff between balanced / standard error and equal opportunity was revealed in the generalization analysis.", "future_impact": "The vector-scaling loss concept and the accompanying insights could improve techniques dealing with label imbalances and group sensitivities, thereby further refining classification algorithms.", "venue": "NeurIPS", "year": 2021, "title": "Label-Imbalanced and Group-Sensitive Classification under Overparameterization."} +{"pid": "63896cd590e50fcafde79e6f", "context": "In deep learning model development, typically the task to be solved is decided and then a fitting model that performs well on it is searched for. However, this approach has generally not considered the reverse scenario, i.e., fixing the model and then searching in the 'task' space.", "key_idea": "The authors propose a task discovery framework that optimizes a generalization-based quantity known as the 'agreement score'. This framework is designed to automatically find tasks on which the pre-decided model generalizes well.", "method": "The authors apply the proposed framework on a set of images to discover many tasks on which neural networks generalize well. They also create adversarial train-test splits based on the discovered tasks.", "outcome": "The study identified many tasks for a predetermined model using the proposed framework. The discovered tasks are a reflection of the inductive biases and statistical patterns present in the data. The results also demonstrated the ability to create adversarial train-test splits which caused the model to fail at test time.", "future_impact": "The tasks discovered by this method provide a meaningful analytical tool for understanding neural networks and their cognitive biases. Furthermore, the human-interpretability of discovered tasks could potentially be a significant area of future research.", "venue": "NeurIPS", "year": 2022, "title": "Task Discovery: Finding the Tasks that Neural Networks Generalize on"} +{"pid": "520611bc-d72e-421c-bd8a-b7f1330378c1", "context": "Symbolic execution is a powerful method of guiding the algorithm design process observed in protocols of human algorithm designers. However, its role in algorithm design has not been earnestly examined, despite work on symbolic execution of complete programs as an aid in the validation of software.", "key_idea": "The authors have developed a data-flow language for representing partially designed algorithms and incorporated a process for the symbolic execution of data-flow configurations that can assist in discovering useful knowledge for guiding design.", "method": "The authors conducted symbolic execution on data-flow configurations of partially designed algorithms using the developed data-flow language as a part of the DESIGNER project.", "outcome": "The techniques of symbolically executing data-flow configurations helped to discover useful knowledge that can guide the algorithm design process.", "future_impact": "The techniques described will contribute significantly towards the goal of building a fully automated algorithm design system.", "venue": "IJCAI", "year": 1985, "title": "Symbolic execution in algorithm design"} +{"pid": "15c3c04e-9a5d-4811-9e76-29f696a618a1", "context": "Saliency detection has been a long-standing problem in computer vision, with most efforts focusing on developing a universal saliency model that applies to all users, regardless of their individual differences in aspects such as gender, race, age, etc.", "key_idea": "This paper introduces the concept that saliency is highly specific rather than universal, due to individuals exhibiting varied gaze patterns even when viewing an identical scene, and therefore proposes a personalized saliency maps (PSMs) approach.", "method": "The authors construct a PSM based on the universal saliency map (USM) shared by different participants and adopt a multitask CNN framework to estimate the discrepancy between PSM and USM.", "outcome": "Experiments show that the new PSM model and prediction scheme are effective and reliable.", "future_impact": "The study produces the first database of personalized saliency maps (PSMs), which may facilitate further research in this domain.", "venue": "IJCAI", "year": 2017, "title": "Beyond Universal Saliency: Personalized Saliency Prediction with Multi-task CNN"} +{"pid": "b81b4c88-1808-481c-b957-a8db7c827ea7", "context": "Over-subscription planning allows users to focus on what they want by associating hard-goal constraints from classical planning with reward values, leaving the decision of the best goals to achieve to the planner. However, the current problem and its goal specification are limited.", "key_idea": "The authors extend the over-subscription planning problem and its goal specification to allow for numeric goals with continuous utility values and for goals with a mix of hard and soft constraints.", "method": "To handle the new goal constraints, the authors extend the Sapaps planner's planning graph-based techniques to identify the most beneficial subset of goals that can include both hard or soft logical and numeric goals. They provide empirical results in several benchmark domains.", "outcome": "The empirical results in several benchmark domains show that the technique proposed by the authors helps to return quality plans.", "future_impact": "The method introduced in this paper significantly extends the modeling power of goal specification, permitting the user to express goal constraints that were not possible before, which could allow greater flexibility and capability in future planning applications.", "venue": "IJCAI", "year": 2005, "title": "Over-subscription planning with numeric goals"} +{"pid": "5eafe7e091e01198d39865dc", "context": "Acquiring commonsense knowledge for artificial intelligence is typically a labor-intensive and costly task requiring extensive human input, making it infeasible on a large scale.", "key_idea": "This paper proposes a method for mining commonsense knowledge from linguistic graphs, aiming to convert relatively low-cost knowledge gleaned from linguistic patterns into more valuable commonsense knowledge. It results in the creation of TransOMCS, a resource comparable to ConceptNet but significantly larger.", "method": "The authors convert ASER, a large-scale selectional preference knowledge resource, into TransOMCS. The effectiveness of this approach is evaluated in terms of quantity, novelty, and quality.", "outcome": "The experiments demonstrate the ability to transfer linguistic knowledge to commonsense knowledge effectively, demonstrating the success of the proposed method in terms of quantity, novelty, and quality of the knowledge acquired.", "future_impact": "The availability of TransOMCS as a publicly available resource could have a significant impact on future artificial intelligence applications by providing a more expansive source of commonsense knowledge.", "venue": "IJCAI", "year": 2020, "title": "TransOMCS: From Linguistic Graphs to Commonsense Knowledge"} +{"pid": "ac68575e-cd57-417c-9cbb-4aa0464d8368", "context": "Emergency managers often need to make rapid, high-quality decisions based on limited and uncertain information, following expert knowledge or past experience. Decisions related to potential release of radioactive material from a nuclear power station like the Guangdong Nuclear Power Station (GNPS) at Daya Bay can have a massive impact on lives and property.", "key_idea": "The authors propose an expert system designed to aid disaster managers in making optimal decisions against a potential nuclear disaster. The expert system assists in determining the activation level of the Daya Bay Contingency Plan (DBCP) and provision of an action checklist, and also provides recommendations on counter-measures.", "method": "A prototype expert system has been developed to cover two major areas of the DBCP plan, which include determination of activation level and provision of an action checklist, and recommendation on counter-measures.", "outcome": "The authors successfully establish a web-based intelligent system for the Daya Bay contingency plan, bringing structure, accuracy, and sufficient expert knowledge to the decision-making process during potential nuclear disasters.", "future_impact": "This system could potentially improve the quality of decisions surrounding emergency management in scenarios involving nuclear disasters, thereby reducing the severity of damage to lives and properties.", "venue": "IJCAI", "year": 2001, "title": "A web-based intelligent system for the Daya Bay contingency plan in Hong Kong"} +{"pid": "ed8c73bb-4d49-405f-bc48-87a94a3dc6a9", "context": "Existing methods for comparing human and agent behavior which are crucial for validating computer models of human behavior, scoring Turing tests, or guiding intelligent tutoring systems.", "key_idea": "The authors introduce behavior bounding, an automated model-based approach for comparing the performances of human and agent behaviors.", "method": "The authors demonstrate this system by applying it to complex agents and comparing the resultant performance with two benchmark methods.", "outcome": "Empirical results demonstrate the efficiency of behavior bounding, requiring minimal human effort and showing superior performance in identifying behavioral problems in certain types of agents compared to two benchmarks.", "future_impact": "Though not explicitly stated, it's implicit that an effective tool like behavior bounding for comparing agent and human behavior can aid in perfecting AI models, improving Turing tests, and enhancing intelligent tutoring systems.", "venue": "IJCAI", "year": 2003, "title": "Behavior bounding: toward effective comparisons of agents & humans"} +{"pid": "331ab2cd-7284-4546-a7b3-54f88cf9ec55", "context": "The paper approaches the problem of creating efficient organizational structures in naturally distributed, homogeneous, cooperative problem solving environments, such as distributed sensor networks.", "key_idea": "The authors propose an approach to statistically describe the performance of any static organization and under what conditions dynamic organizations perform better and worse. They also analyze the impact of variance in agent performance on the uncertainty of a dynamic organization's success compared to a static one and the usefulness of meta-level communication.", "method": "The authors use statistical analysis and simulations to describe the performance of static and dynamic organizations and ascertain the conditions under which these structures perform best.", "outcome": "The authors demonstrate when meta-level communication about the actual state of problem-solving could be useful to agents in constructing a dynamic organizational structure that outperforms a static one.", "future_impact": "The methodology presented in this paper can be used for the systematic analysis and design of distributed problem-solving systems in complex environments.", "venue": "IJCAI", "year": 1993, "title": "An approach to analyzing the need for meta-level communication"} +{"pid": "885e5f85-02f4-4b17-a15e-201fe56898b4", "context": "Current machine translation evaluation paradigms are driven by metrics such as BLEU and WER which are often based on superficial n-gram and string representations. This has led to translations that often produce inaccurate semantic event structures.", "key_idea": "The authors propose a paradigm shift towards a more intelligent evaluation approach that relies on lexical semantic level and the use of PropBank style semantic frame representations.", "method": "The authors use a representational foundation based on AI-oriented lexical semantics to evaluate machine translation quality. This approach is also empirical, achieved by replacing human annotators with automatic semantic role labeling.", "outcome": "The authors show that by using this method, it is possible to surpass the correlation of all currently widespread metrics to human adequacy judgments, including even HTER.", "future_impact": "The proposed evaluation paradigm could offer superior yet low-cost quantitative objective functions for translation quality. It also gives a clear reflection of semantic utility of structural frame-based knowledge and restores representational transparency from an AI perspective.", "venue": "IJCAI", "year": 2011, "title": "SMT versus AI redux: how semantic fames evaluate MT more accurately"} +{"pid": "af1edeb0-2a77-4eb9-89a5-1b1e1d8af09e", "context": "The cross-camera person-matching problem in the person re-identification task remains a challenge.", "key_idea": "An end-to-end deep correspondence structure learning (DCSL) approach is proposed that captures the intrinsic structural information on persons by learning a semantics-aware image representation based on convolutional neural networks. The approach further includes an adaptively learned hierarchical data-driven feature matching function.", "method": "The authors set up a unified end-to-end deep learning scheme to jointly optimize the processes of semantics-aware image representation learning and cross-person correspondence structure learning. They test their approach on several benchmark datasets.", "outcome": "The experimental results show that the proposed approach is effective against the state-of-the-art approaches.", "future_impact": "The newly developed approach has potential to lead to more reliable and robust person re-identification results in complicated scenarios.", "venue": "IJCAI", "year": 2016, "title": "Semantics-aware deep correspondence structure learning for robust person re-identification"} +{"pid": "6216627e-bcc8-40f9-89dc-14460e1b055f", "context": "The original CYPRESS algorithm design system is characterized by its synthesis of divide-and-conquer sorting algorithms from formal specifications.", "key_idea": "The authors reimplement a significant portion of CYPRESS within the Soar problem-solving architecture to create CYPRESS-SOAR, a system that uses heuristic search of problem spaces and learning during algorithm design.", "method": "The key functionalities of CYPRESS-SOAR are tested by synthesizing three divide-and-conquer sorting algorithms and observing the system's performance in terms of search, knowledge acquisition, and learning.", "outcome": "CYPRESS-Soar managed to reproduce most of CYPRESS' behavior, exhibiting significant transfer of learned knowledge, within a single design run and across designs of several different algorithms. This successful reimplementation was achieved with just high-level synthesis control of CYPRESS.", "future_impact": "CYPRESS-Soar could serve as a potent research vehicle for further exploring the interplay between search, knowledge, and learning in the realm of algorithm design.", "venue": "IJCAI", "year": 1987, "title": "CYPRESS-Soar: a case study in search and learning in algorithm design"} +{"pid": "60acc33591e011a83767347e", "context": "Most research on algorithmic decision-making solely focuses on the algorithm's performance and frames the decision-making problems as classification tasks, neglecting the importance of human-machine complementarity across various instances of a given domain.", "key_idea": "The paper proposes a solution for a novel human-machine collaboration problem in a bandit feedback setting with the goal of exploiting human-machine complementarity to maximize decision rewards. This approach is then extended to settings with multiple human decision makers.", "method": "The effectiveness of the proposed methods is demonstrated using both synthetic and real human responses, and a system is designed for personalized routing in the presence of multiple human decision-makers.", "outcome": "The methods implemented outperform both the algorithm and the human when they make decisions independently. It is also shown that personalized routing can further improve the performance of the human-machine team.", "future_impact": "The proposed approach may pave the way for more effective decision-making procedures and systems that optimally leverage both human judgment and algorithmic computation, particularly in situations where multiple human decision-makers are involved.", "venue": "IJCAI", "year": 2021, "title": "Human-AI Collaboration with Bandit Feedback."} +{"pid": "90937b30-894e-4ae8-a82f-a9cb656fbf60", "context": "The bimodal nonmonotonic logic MBNF was suggested as a generalization of a number of nonmonotonic formalisms, but the relationships and equivalences of this logic with nonmodal systems and other formalisms have not been thoroughly explored.", "key_idea": "The paper aims to study and establish equivalences between the bimodal nonmonotonic logic MBNF and nonmodal systems, modal representations that use only one modal operator of belief, and nonmodal default consequence relations.", "method": "The authors evaluate the equivalences by systematically correlating the models, rules and translations in MBNF with its counterparts in the nonmodal system, a single-operator modal representation, and nonmodal default consequence relations.", "outcome": "The authors find that MBNF is equivalent to a specific nonmodal system involving special type of rules, and that this nonmodal system admits a modal representation using only one operator, the operator of belief. It's also found that MBNF is reducible to nonmodal default consequence relations.", "future_impact": "These results illuminate the relationship between different formalizations of nonmonotonic reasoning and could influence how these different formalizations are understood and used in the future.", "venue": "IJCAI", "year": 1995, "title": "On bimodal nonmonotonic logics and their unimodal and nonmodal equivalents"} +{"pid": "2233827a-36e3-4d2f-8392-43a6178e5da4", "context": "In the field of machine learning, constructing an ensemble of classifiers and finding a suitable selection of feature and data points for underdetermined problems is an ongoing challenge.", "key_idea": "The authors present BENCH (Biclustering-driven ENsemble of Classifiers), an algorithm for constructing an ensemble of classifiers that leverages concurrent feature and data point selection guided by unsupervised knowledge from biclustering.", "method": "Experiments were conducted with Bayesian Belief Network (BBN) classifiers as base classifiers in the ensemble on three microarray datasets, using various weighting schemes for combining individual predictions in the ensemble.", "outcome": "The BENCH algorithm was found to improve the prediction accuracy of a single classifier and traditional ensemble of classifiers by up to 15% on three microarray datasets.", "future_impact": "BENCH can be applied to other classification models as well, indicating the potential to improve predictive accuracy across a range of machine learning tasks.", "venue": "IJCAI", "year": 2011, "title": "Biclustering-driven ensemble of Bayesian belief network classifiers for underdetermined problems"} +{"pid": "64671268d68f896efaf14a29", "context": "The learnware paradigm constructs a market of pre-trained models that can be reused, however, existing research focuses on the homogeneous case where models share the same feature space based on Reduced Kernel Mean Embedding (RKME) specification. But in reality, models are typically constructed from different feature spaces. Current methods to handle different feature spaces would require additional auxiliary data which is hard to obtain.", "key_idea": "The authors propose a general framework for accommodating learnwares from heterogeneous feature spaces without requiring additional auxiliary data by utilizing the submitted RKME specifications to establish the relationship between different feature spaces.", "method": "The authors present a matrix factorization-based implementation and propose an overall procedure for constructing and exploiting the heterogeneous learnware market, with experiments conducted on real-world tasks.", "outcome": "Experiments on real-world tasks validate the efficacy of the proposed method.", "future_impact": "The proposed framework will enable the reuse of models built for a particular task even with different feature spaces, advancing the use and effectiveness of the learnware paradigm.", "venue": "IJCAI", "year": 2023, "title": "Handling learnwares developed from heterogeneous feature spaces without auxiliary data"} +{"pid": "6c6994c5-77fd-454d-a418-ae51d1620407", "context": "The task of auditory stream segregation from input sounds is a complex problem in computational auditory scene analysis, with existing solutions not effectively handling segregation of multiple types of sound streams.", "key_idea": "A new model termed 'Residue-Driven Architecture' is proposed, which involves three kinds of agents: an event-detector, a tracer-generator, and tracers. The event-detector calculates a residue for each sound, and when this residue exceeds a certain threshold, a tracer is generated to extract an auditory stream.", "method": "The design and implementation of the Residue-Driven Architecture involves the calculation of a residue by subtracting predicted inputs from actual inputs and subsequent generation of tracers when a threshold value is exceeded.", "outcome": "This approach was found to improve the performance of segregation, being capable of segregating a woman's voiced stream, a man's voiced stream, and a noise stream from a mixture of such sounds.", "future_impact": "The newly proposed model is expected to bring advancements in auditory scene analysis, with possible extensions to binaural segregation systems.", "venue": "IJCAI", "year": 1995, "title": "Residue-driven architecture for computational auditory scene analysis"} +{"pid": "7a55fdfa-78ca-4c2a-98d4-e67500653eca", "context": "Research in abstract argumentation typically pertains to finite argumentation frameworks (AFs). However, actual or potential infinite AFs frequently occur in various practical cases, including nonmonotonic entailment, instantiation-based argumentation, dialogues, n-person-games, or action sequences.", "key_idea": "The authors propose to study the abstract properties like SCC-recursiveness, expressiveness or intertranslatability for unrestricted abstract argumentation frameworks.", "method": "The method is not explicitly stated in the abstract.", "outcome": "The outcome is not explicitly stated in the abstract.", "future_impact": "A profound analysis can yield a better understanding of how the nonmonotonic theory of abstract argumentation works in general.", "venue": "IJCAI", "year": 2017, "title": "A study of unrestricted abstract argumentation frameworks"} +{"pid": "5ef96b048806af6ef27720e9", "context": "Multiagent reinforcement learning scenarios often require independent agents to jointly learn to perform a cooperative task. In many cases, these agents have individual preferences regarding how to accomplish the shared task.", "key_idea": "The authors propose a framework that balances individual preferences against task rewards using a linear mixing scheme. The paper focuses on this framework where agents have individual preferences for the shared task.", "method": "The authors perform a theoretical analysis to establish that agents can reach an equilibrium that leads to optimal shared task reward even with individual preferences that are not fully aligned with the task. Empirical analysis is then conducted to study different mixing schemes.", "outcome": "The authors empirically show that there exist mixing schemes that outperform a purely task-oriented baseline.", "future_impact": "This work could lead to future research on optimizing the proposed mixing scheme in multiagent reinforcement learning scenarios.", "venue": "IJCAI", "year": 2020, "title": "Balancing Individual Preferences and Shared Objectives in Multiagent Reinforcement Learning"} +{"pid": "ef94cc43-43f0-4507-81c6-5c03ccc8bb15", "context": "The Constraint games framework, solving through constraint propagation, was previously established, but the efficiency of the solving technique has been a problem.", "key_idea": "The authors reconsider the solving technique in constraint games, treating players' preferences as global constraints for a better integration in constraint solvers.", "method": "The authors released a new solver, ConGA, in open source and used it to find pure Nash equilibrium for problems with varying number of players, from 200 to 2000, particularly concentrating on telecommunication problems.", "outcome": "The new solver, ConGA, outperforms the previous state-of-the-art in efficiency and effectiveness, as it is able to find all pure Nash equilibrium for some problems with up to 2000 players. The solver was able to tackle real-world telecommunication problems from both centralized and decentralized perspectives.", "future_impact": "The new techniques for improving the resolution process offer a new perspective that can be applied to real-world telecommunications problems and in evaluating the quality of service in multi-user applications.", "venue": "IJCAI", "year": 2017, "title": "Constraint games revisited"} +{"pid": "d7c934dd-ef7b-4fca-a9fe-1dd3f5317092", "context": "Accessing data from a large, distributed data base over a computer network can be a challenging task, with issues relating to appropriate query planning, machine determination, network linking, query monitoring, and error recovery.", "key_idea": "The paper introduces a system that provides real-time access to multiple instances of a database management system (DBMS) over the ARPANET with artificial intelligence. It accepts a broad range of natural language questions, plans suitable queries to answer the question, determines execution machine and network links, monitors query prosecution, and recovers from certain execution errors.", "method": "The researchers developed an implemented system accepting natural language questions, carrying out relevant processes to answer the question including query planning, machine determination, network linking, query monitoring, and error recovery.", "outcome": "The running system was successful in facilitating access to multiple instances of a DBMS over the ARPANET, demonstrating its ability to handle natural language queries and recover from certain errors in execution.", "future_impact": "The successful implementation of the LADDER (Language Access to Distributed Data with Error Recovery) system may prompt higher levels of sophistication in functional components, potentially improving distributed data access capabilities.", "venue": "IJCAI", "year": 1977, "title": "Language access to distributed data with error recovery"} +{"pid": "6e2b1225-d2e5-4be3-b6c9-54e1467f2bc3", "context": "The grid code imposes that a wind power generator, such as one using a double fed induction generator (DFIG), must ride through grid voltage sags and only disconnect when the grid voltage drops under a certain threshold. However, DFIG systems have issues with this rule due to their partial power converter, which can lead to system shutdown if current, power, and reactive power peaks exceed rated values during a voltage sag.", "key_idea": "The paper aims to improve low-voltage ridethrough (LVRT) capability of a DFIG wind power system by analyzing the effects of a crowbar on damping transient currents during voltage sags and supporting the grid by providing reactive power.", "method": "The authors conducted experiments with a wind power simulation system composed of a DC motor, DFIG, 3-phase diode rectifier, converter, crowbar, and a voltage sag generator. The setup is designed to study the dynamic response of the rotor-side converter's inner current loop during voltage sags.", "outcome": "The experimental and simulation results show that using a crowbar in the rotor-side converter using the dq transfer voltage measure method can reduce transient currents in the rotor and stator. Moreover, the DFIG system with a crowbar swiftly recovers from voltage sags and provides reactive power, contributing to the grid support.", "future_impact": "This work emphasizes the role of a crowbar in enhancing the LVRT capability of a DFIG wind power system, which could influence the design and operation of such systems to better accommodate grid codes and potentially enhancing grid stability by providing reactive power during voltage sags.", "venue": "IJCAI", "year": 2009, "title": "Analysis on DFIG Wind Power System Low-Voltage Ridethrough"} +{"pid": "41d918d4-d4a7-40bb-8399-8981e33b77f5", "context": "Recommender systems on E-Commerce platforms use bipartite graphs to track user/item feature and use-item interactions. These systems have started using graph neural networks (GNN), but providing individualized recommendations to billions of users is still extremely challenging.", "key_idea": "This paper proposes a method that clusters users into communities and makes recommendations based on collective user information in those communities. User information in these embeddings can be decomposed into two parts, capturing both community-level generalizations and individualized preferences.", "method": "The authors apply the proposed hierarchical representation method to a public dataset and dataset from a leading E-Commerce company.", "outcome": "The abstract does not include specific outcomes of the application of the proposed method.", "future_impact": "The new method is designed to enhance current GNN methods that do not learn hierarchical representations of graphs, potentially improving the effectiveness of these systems.", "venue": "IJCAI", "year": 2019, "title": "Hierarchical representation learning for bipartite graphs"} +{"pid": "60a4df0b91e011e398b0cf3c", "context": "The computation of acoustic scattering properties of geometric objects has traditionally been challenging, especially in dynamic scenes with objects of arbitrary topologies and deforming models.", "key_idea": "The authors introduce a novel geometric deep learning method that uses a point cloud representation of objects to compute their acoustic scattering properties, which integrates with ray tracing for interactive sound propagation.", "method": "The authors' learning algorithm uses discrete Laplacian-based surface encoders and approximates the neighborhood of each point using a shared multi-layer perceptron. The work presents a neural network that computes the scattering function using spherical harmonics, and is validated on thousands of unseen 3D objects.", "outcome": "The proposed method can process objects with arbitrary topologies and deforming models in less than 1ms per object on a standard GPU. Their accuracy has been demonstrated over other point-based geometric deep learning methods.", "future_impact": "As the first real-time learning algorithm that can accurately approximate the acoustic scattering properties of arbitrary objects, this has potential to greatly improve the interaction of sound propagation in various applications.", "venue": "IJCAI", "year": 2021, "title": "Point-based Acoustic Scattering for Interactive Sound Propagation via Surface Encoding."} +{"pid": "258eb074-fc2f-47df-ac67-07ae65348ea6", "context": "Before this study, systems for automatic assembly that could construct 3-D body models based on visual input were not well-developed or efficient.", "key_idea": "The authors propose a system that uses light stripe input to construct 3-D body models suitable for automatic assembly. The system utilizes special hardware to extract a stripe from a TV picture quickly and then analyze the stripe to develop a body model.", "method": "The authors use a program that forms hypotheses based on clues in the light stripe data and then tries to establish these hypotheses. The results are used to generate body models, which are composed of convex subsets of three-space formed as intersections of (possibly complemented) half-spaces and solid cylinders.", "outcome": "The proposed system can extract a stripe from a TV picture in about 0.2 second.", "future_impact": "The proposed system could be further developed and applied to improve the efficiency and accuracy of automatic assembly operations.", "venue": "IJCAI", "year": 1975, "title": "Forming models of plane-and-cylinder faceted bodies from light stripes"} +{"pid": "cb48836a-1b61-4913-bff8-4e48354de198", "context": "Metaphor is a prevalent feature in natural language discourse, but the systematic interpretation of metaphors remains a challenge.", "key_idea": "This paper proposes a computational approach to interpret metaphor by translating the problem into selecting the relevant inferences to draw from the metaphorical expression.", "method": "The authors illustrate their approach using two examples of metaphors -- a spatial metaphor schema from computer science, and a novel metaphor -- to demonstrate how the interpretation problem can be translated into a selective inferencing problem.", "outcome": "The proposed approach effectively solves the metaphor interpretation problem using the ordinary operations of their system.", "future_impact": "This approach provides new insights on the analogical processes underlying metaphor and helps to explain the power of metaphor, indicating potential value for natural language processing systems.", "venue": "IJCAI", "year": 1981, "title": "Metaphor interpretation as selective inferencing"} +{"pid": "c43a2c4d-3402-49d3-902d-ff2aaab7dc13", "context": "Hierarchical type fuzzy systems are used to represent complex systems in a structured way. However, the construction of these models, especially when dealing with priorities and incomplete information, remains a challenge.", "key_idea": "The authors introduce a model called Hierarchical Prioritized Structure (HPS), which can handle the construction of fuzzy systems where rules are provided by an expert and, importantly, consider the issue of completing incomplete priorities.", "method": "The authors review the structure, operation, and inter-level aggregation algorithm of the HPS, and apply a mathematical programming method for implementing the principle of maximal buoyancy in the context of incomplete priorities.", "outcome": "The paper provides a detailed exploration of the HPS, its construction, and its handling of incomplete priorities via the principle of maximal buoyancy.", "future_impact": "The discussion about tuning hierarchical models suggests that future research could address the optimization of these models, potentially improving their effectiveness in various applications.", "venue": "IJCAI", "year": 1995, "title": "Constructing Prioritized Fuzzy Models"} +{"pid": "9eefcfb5-5579-4c98-8d3d-b61a451b85b5", "context": "Current methods of discovering theorems in two-person finite games in strategic form have limitations and do not fully utilize technology.", "key_idea": "The authors propose a logical framework for using computers to discover classes of two-person games that have unique pure Nash equilibrium payoffs, and all possible classes of games that can be expressed by a conjunction of two binary clauses.", "method": "The authors applied the logical framework to rediscover weakly unilaterally competitive two-person games and identify several other new classes of games that have unique pure Nash equilibrium payoffs.", "outcome": "Using the new framework, the authors were able to rediscover Kats and Thisseu0027s class of weakly unilaterally competitive two-person games and discover several other new classes of games with unique pure Nash equilibrium payoffs, including new classes of strict games.", "future_impact": "The proposed logical framework could be used for further discovery in game theory, particularly in identifying strategic forms of two-person games with unique pure Nash equilibrium payoffs.", "venue": "IJCAI", "year": 2009, "title": "Discovering theorems in game theory: two-person games with unique pure nash equilibrium payoffs"} +{"pid": "498bcf88-5bd5-47d3-b8bd-22fcb5faff26", "context": "Diagnosis of student programming errors in programming learning environments is traditionally implemented manually, which is time-intensive and can be inconsistent.", "key_idea": "The authors propose an automated method to diagnose students' programming errors through an improved dependence graph representation. The student's program is semantically compared to a correct model program, after both have been standardized by program transformations.", "method": "The method is implemented in Smalltalk in SIPLeS-II, an automatic program diagnosis system for Smalltalk programming learning environments, and tested on approximately 330 student programs for various tasks.", "outcome": "Experimental results show the method can identify semantic errors in a student program rigorously and safely, and it can eliminate or accommodate semantics-preserving variations. The system can identify a wide range of errors and produce indications of the corrections needed.", "future_impact": "The proposed method and techniques may prove valuable for the development of programming learning environments and in other research fields including program understanding and software maintenance.", "venue": "IJCAI", "year": 1999, "title": "Automatic diagnosis of student programs in programming learning environments"} +{"pid": "c2217a84-ff24-45ab-9b32-53854c77805e", "context": "The understanding of cognitive aspects of autonomous robots can be improved by developing a mechanism that allows the robot to autonomously analyze physical sensor data and construct a state space. However, a statistically coherent approach for this purpose doesn't exist yet.", "key_idea": "This paper proposes an approach to construct a robot-oriented state space by statistically analyzing sensor patterns and rewards from task executions. Notably, the method includes the creation of sensor pattern classifiers called Empirically Obtained Perceivers (EOPs) that direct attention to select necessary information.", "method": "The state space is obtained with an attention control mechanism using EOPs. The robot tests the construction of state spaces through its vision sensor and executes a navigation task in a simulated world.", "outcome": "The study confirms that the robot could effectively construct state spaces through its vision sensor and successfully execute a navigation task in a complex, simulated world using the obtained state spaces.", "future_impact": "The newly proposed method for constructing state spaces by statistically analyzing sensor patterns and rewards could significantly contribute to advancing the capability of autonomous robots in the future.", "venue": "IJCAI", "year": 1999, "title": "State space construction by attention control"} +{"pid": "5ef96b048806af6ef2772024", "context": "The process of learning to hash in deep learning involves the use of labels for similar and dissimilar items. This conventional method often relies on Hamming distance-based losses and the task of assigning binary codes can get complex.", "key_idea": "The authors propose a novel deep polarized network (DPN) for learning to hash, which uses a differentiable bit-wise hinge-like loss called polarization loss. This loss pushes each channel's output in the network away from zero, removing the need for pairwise labels for similar or dissimilar items.", "method": "The network is reformulated within a generic Hamming Distance Metric Learning framework. The proposed loss function is verified through experiments on CIFAR10, NUS-WIDE, and ImageNet100 datasets.", "outcome": "The implementation of the deep polarized network and the proposed loss function resulted in state-of-the-art performances on CIFAR10, NUS-WIDE, and ImageNet100 datasets, and made the subtask of assigning binary codes extremely simple.", "future_impact": "The introduction of the Deep Polarized Network for learning binary hashing codes can significantly simplify the task of assigning binary codes and has the potential to improve the accuracy of models in the future.", "venue": "IJCAI", "year": 2020, "title": "Deep Polarized Network for Supervised Learning of Accurate Binary Hashing Codes"} +{"pid": "cfd9fc5e-984f-4825-81b8-7272a9b6617b", "context": "Halpern and Shoham's interval logic (HS) has been extensively used for examining Allenu0027s relations. Despite this, there has not been a complete set of inter-definability equations between the modal operators, leading to a lack of complete classification in all possible variants of HS.", "key_idea": "This study establishes a comprehensive set of inter-definability equations to understand the relationships between modal operators in Halpern and Shoham's interval logic (HS), thus providing a complete classification of its variations.", "method": "The researchers used their new set of inter-definability equations and a computer program to examine all possible fragments and variations of Halpern and Shoham's interval logic (HS).", "outcome": "The researchers found and classified 1347 expressively different interval logics based on Halpern and Shoham's (HS) logic over the class of all linear orders.", "future_impact": "This work could provide a basis for future research into interval logic, Allen's relations, and their applications, as it offers a robust means to understand and classify variations within interval logic.", "venue": "IJCAI", "year": 2011, "title": "Expressiveness of the interval logics of Allen's relations on the class of all linear orders: complete classification"} +{"pid": "5f0c30d491e01165a98e19f4", "context": "The process of performing literature survey and generating a summary specific to a research draft is currently manually handled, which can be labor-intensive and time-consuming.", "key_idea": "The authors introduce AutoSurvey, an intelligent system that can automate the generation of literature surveys and summaries for specific research drafts. This is done by employing a neural model for extracting information from abstracts and a novel evolutionary multi-source summarization model for summary generation.", "method": "AutoSurvey uses a neural model for information structure analysis to extract fine-grained information from the abstracts of previous work. It also uses a novel evolutionary multi-source summarization model for generating the summary of related work.", "outcome": "The system, which was developed, AutoSurvey, can automatically generate a literature survey and a summary specific to a research draft.", "future_impact": "AutoSurvey has wide potential for application in both academic and educational purposes, simplifying the task of generating literature reviews and summaries.", "venue": "IJCAI", "year": 2020, "title": "AutoSurvey - Automatic Survey Generation based on a Research Draft."} +{"pid": "142f089a-6555-464a-abd1-41ffe62a920a", "context": "The current safety conditions of foundation excavation need improvements.", "key_idea": "The authors propose a support monitoring system of foundation excavation using virtual reality and virtual instrument technique that integrates data collection, information processing, analytical evaluation, and 3D visualization of the excavation state.", "method": "The authors develop the system using Windows and LabVIEW technologies.", "outcome": "The system is able to collect data automatically, process information, conduct analytical evaluations, and display the state of excavation via 3D visualization technology.", "future_impact": "The developed system is significant in restricting the strain, ensuring the safety of the excavation, and protecting the surrounding environment.", "venue": "IJCAI", "year": 2009, "title": "Monitoring System of Foundation Excavation Based on Virtual Instruments"} +{"pid": "cd3c44d1-c01a-495d-a790-f0a5e211e87f", "context": "Fork-decoupled search is a recent approach to classical planning that uses fork structures, with a simple state-dominance criterion used to prune dominated search states.", "key_idea": "The authors propose several new state-dominance criteria for fork-decoupled search, providing a more powerful tool for pruning dominated search states.", "method": "The authors apply these new state-dominance criteria in fork-decoupled search and run experiments on IPC benchmarks to measure their effectiveness.", "outcome": "The new criteria preserve optimality and can yield exponential reductions, as shown through experiments on IPC benchmarks.", "future_impact": "These new state-dominance criteria could lead to more efficient fork-decoupled search, potentially benefiting practical applications in classical planning.", "venue": "IJCAI", "year": 2016, "title": "On state-dominance criteria in fork-decoupled search"} +{"pid": "1f870913-0ce0-48dc-95af-d9786004d1a2", "context": "Prior to this study, learning systems were approached on a case-by-case basis, with no common functional components identified that span across different techniques and environments.", "key_idea": "The authors introduce a universal model for learning systems, detailing essential functional components that are consistent across different constructions and operating environments. These components include the Performance element, instance selector, critic, earning element, blackboard, and world model.", "method": "The authors test this model by discussing representative AI, pattern recognition, and control systems within its framework.", "outcome": "The discussion and analysis of various AI, pattern recognition, and control systems under this framework support the viability of the proposed model for learning systems.", "future_impact": "The introduction of this model poses the possibility of a more structured and unified approach to learning system design. This will potentially lead to the creation of layered systems, with each layer operating at different levels of abstraction.", "venue": "IJCAI", "year": 1977, "title": "A model for learning systems"} +{"pid": "78bf1174-9cc2-43a6-8ac1-f2cb5ce7ef65", "context": "Interactive decision support in sequential decision problems often assumes the Decision Maker's preferences, which are representable by an additive utility, are initially unknown or imperfectly known. Current methods struggle with determining optimal policies when utilities are imprecisely defined.", "key_idea": "A new adaptive preference elicitation method is proposed that interleaves the elicitation of utilities and backward induction to incrementally determine a near-optimal policy.", "method": "An interactive algorithm with performance guarantees is proposed and numerical tests are conducted to verify the efficacy of the proposed method.", "outcome": "The numerical tests demonstrate the practical efficiency of the proposed approach, showcasing the performance of the introduced algorithm.", "future_impact": "The proposed approach introduces a potentially more efficient and accurate method for adaptive preference elicitation which can significantly improve interactive decision support in sequential decision problems.", "venue": "IJCAI", "year": 2017, "title": "Adaptive Elicitation of Preferences under Uncertainty in Sequential Decision Making Problems"} +{"pid": "883c49a1-f6bb-407c-b9bb-2ffba428e0a7", "context": "Latent Dirichlet allocation (LDA) has been successful in discovering the statistical distribution of the topics over an unstructured text corpus. Recently, more document data, known as semi-structured data, have emerged with rich human-provided tag information due to the evolution of the internet. Modeling such semi-structured documents, which include both unstructured data like plain text and metadata like authorship or tags, is nontrivial.", "key_idea": "The authors propose a new method to model tagged documents using a topic model called Tag-Weighted Topic Model (TWTM). This leverages the tags in each document to infer the topic components for the documents.", "method": "The authors design an efficient variational inference method with an EM algorithm for estimating the model parameters and test their TWTM approach against baseline algorithms over three corpora in document modeling and text classification.", "outcome": "Experimental results show that the TWTM approach outperforms baseline algorithms in document modeling and text classification over three tested corpora.", "future_impact": "The authors anticipate that the TWTM, which can infer the probabilistic weights of tags for each document, not only learns document-topic distributions, but also infers the tag-topic distributions for text mining. It could impact applications such as classification, clustering, and recommendations.", "venue": "IJCAI", "year": 2013, "title": "Tag-weighted topic model for mining semi-structured documents"} +{"pid": "48b11c70-cb8f-468e-8e7e-fd30ea98917f", "context": "Although considerable theoretical research on lifted inference in probabilistic graphical models (PGMs) exists, there are few applications of these methods in combination with top-tier applied algorithms, especially in the field of computer vision (CV).", "key_idea": "The authors propose the applicability of lifted inference for CV, with the idea that a globally optimal (MAP) labeling will likely have the same label for symmetric pixels. They suggest a generic template for Coarse-to-Fine (C2F) inference in CV, which progressively refines an initial coarsely lifted PGM for different quality-time trade-offs.", "method": "The team applies their approach to two near state-of-the-art CV algorithms for stereo vision and interactive image segmentation to demonstrate the performance of C2F inference.", "outcome": "The results show that the lifted versions of the algorithms have superior anytime performance without any loss in final quality, demonstrating an efficient handling of a distinct unary potential on every node (pixel).", "future_impact": "The proposed lifted inference approach and C2F template could extend to a wide class of algorithms that model a computer vision problem through PGM inference, presenting a new direction for improving algorithm performance.", "venue": "IJCAI", "year": 2017, "title": "Coarse-to-Fine Lifted MAP Inference in Computer Vision"} +{"pid": "23ba391a-ebe4-4cb3-9e65-ce7b9f125d97", "context": "The process of concept formation and conjecture generation in pure mathematics typically requires significant manual effort and expertise.", "key_idea": "The authors proposed the HR program that automatically forms mathematical concepts, makes conjectures, and uses theorem prover OTTER and model generator MACE to prove or disprove these conjectures.", "method": "HR measures properties of concepts and assesses the theorems and proofs involving them to estimate each concept\u2019s interestingness, utilizing a best-first search.", "outcome": "The HR program successfully discovered interesting new mathematics and was capable of building theories from only the axioms of finite algebras.", "future_impact": "The approach suggests potential for further automation in mathematics, specifically for concept creation, conjecture making and validation.", "venue": "IJCAI", "year": 1999, "title": "Automatic concept formation in pure mathematics"} +{"pid": "78b6285f-e283-4476-9e22-2846e5e22c18", "context": "Before this study, there was a lack of effective systems that encouraged sharing and collaborative construction of knowledge bases describing physical domains.", "key_idea": "The authors introduce CDME (Collaborative Device Modeling Environment), a compositional modeling system for constructing domain theories of physical systems, composing models of devices, and simulating their behavior, with an emphasis on encouraging knowledge sharing.", "method": "CDME has been implemented as a collection of network services on the World Wide Web, with knowledge represented at three different levels: physical, ontological, and logical.", "outcome": "CDME effectively supports knowledge sharing at the physical, ontological, and logical levels as described by the authors.", "future_impact": "CDME, with its emphasis on sharing and collaborative construction of knowledge bases, could encourage more cooperation and synergistic knowledge advancement in the field of physical systems modeling.", "venue": "IJCAI", "year": 1997, "title": "A Web-Based Compositional Modeling System for Sharing of Physical Knowledge."} +{"pid": "5ef96b048806af6ef2772022", "context": "The Visual Dialogue task involves an agent engaging in a conversation with a human about an image, in which the ability to generate detailed and non-repetitive responses is essential for achieving human-like conversation.", "key_idea": "The authors propose a novel generative decoding architecture, which uses the Deliberation, Abandon and Memory (DAM) module for generating high-quality responses. This design focuses on transparency and flexibility, with word generation broken down into a series of attention-based information selection steps.", "method": "The proposed DAM model is applied to three different encoders and the performance is evaluated on the VisDial v1.0 dataset.", "outcome": "The DAM module provides more detailed and non-repetitive descriptions maintaining semantic accuracy. Experimental results show that the introduced models set a new state-of-the-art performance with high-quality responses.", "future_impact": "The flexibility of DAM to work with existing visual dialogue encoders and adaptability to encoder structures by constraining the information selection mode could stimulate the development of more sophisticated models for visual dialogue tasks.", "venue": "IJCAI", "year": 2020, "title": "DAM: Deliberation, Abandon and Memory Networks for Generating Detailed and Non-repetitive Responses in Visual Dialogue"} +{"pid": "6573cb59-b5f1-434b-bf76-25c00b0c1802", "context": "Temporal reasoning is widely used in Computer Science and A.I., but its underlying complexity in discrete temporal logics has restricted its use to simplified formalisms such as temporal interval algebras or model checking.", "key_idea": "The authors propose the development of tractable sub-classes of propositional linear temporal logic based on the use of XOR fragments of the logic, as a new way to engineer tractable temporal logics by rightly combining XOR fragments.", "method": "The authors use XOR fragments of the logic, and via clausal temporal resolution, they test the decision ability of these fragments. Checks also extend to benefits from combining multiple XOR fragments and establishing the completeness and complexity of the resolution method as applied to such combinations.", "outcome": "The authors demonstrate that tractable sub-classes of propositional linear temporal logic can be developed based on XOR fragments, show that these fragments can be decided tractably, and details how this temporal language can be used in application areas such as verification of multi-agent systems.", "future_impact": "This new approach to temporal reasoning provides a new framework where tractable temporal logics can be engineered by rightly combining appropriate XOR fragments, which may be beneficial and influence future approaches to temporal reasoning.", "venue": "IJCAI", "year": 2007, "title": "Tractable temporal reasoning"} +{"pid": "4c3e3b5a-a92a-44da-bc44-c953b57d5120", "context": "Epistemic logic programming has an enhanced capacity to handle complex incomplete information reasoning and represent agents' epistemic behaviours, but it also involves much higher computational complexity than non-disjunctive and disjunctive answer set programming.", "key_idea": "The key idea of this paper is to investigate important properties of epistemic logic programs, and try to extend Lee and Lifschitz's result on loop formulas for disjunctive logic programs to a special class of epistemic logic programs. They also study the polysize model property for epistemic logic programs.", "method": "The authors extend the result on loop formulas to a special class of epistemic logic programs and study the polysize model property for these programs.", "outcome": "The authors identify two non-trivial classes of epistemic logic programs whose consistency checking complexity is reduced from PSPACE-complete to NP-complete and \u03a32P-complete.", "future_impact": "The paper observes that many important applications on epistemic representation fall into these two classes of epistemic logic programs, indicating potential future impact on the application of such logic programs.", "venue": "IJCAI", "year": 2007, "title": "Epistemic reasoning in logic programs"} +{"pid": "390b77e9-09ed-47b8-b170-e06c927dcd42", "context": "Detecting and measuring objects to represent three-dimensional outdoor scenes is a crucial field of study, especially while using data obtained from stereo vision or scanning laser rangefinders.", "key_idea": "The paper offers a new method for object detection and measurement in three-dimensional outdoor scenes. The method relies on approximating objects using ellipsoids, segmenting them from the background and each other by fitting ellipsoids, clustering points, and adjusting the clusters as per the fit.", "method": "The proposed methodology involves segmentation of objects by finding the ground surface, forming a preliminary segmentation by clustering the points above the ground, fitting ellipsoids to match these clusters and avoid obscuring other points, and adjusting the clusters according to the fits. An example showing implementation of this method is provided using a stereo pair of pictures of Mars from the Viking Lander.", "outcome": "The proposed method is specifically useful for obstacle avoidance and navigation in exploring vehicles, with the example results reflecting its successful application on pictures from the Viking Lander.", "future_impact": "The method may contribute to advancements in navigation and obstacle avoidance in exploring vehicles such as a Mars rover.", "venue": "IJCAI", "year": 1979, "title": "Object detection and measurement using stereo vision"} +{"pid": "4495b4b1-a90b-4953-8723-f42f893f762d", "context": "Metaphorical interpretation and affect detection using context profiles from open-ended text input represent challenging tasks in the field of affective language processing.", "key_idea": "This study explores recognition of certain affective metaphorical phenomena and context-based affect sensing by modeling speakers' improvisational mood and other participants' emotional influence in the context of loosely structured scenarios.", "method": "The updated affect detection module, which includes these new developments, is implemented into an AI agent.", "outcome": "The implementation of the updated module in the AI agent has generally increased the agent's performance in affect sensing tasks.", "future_impact": "This work stays aligned with conference themes relevant to affective dialogue processing, human-agent interaction, and intelligent user interfaces, implying its future usefulness in these domains.", "venue": "IJCAI", "year": 2011, "title": "Affect sensing in metaphorical phenomena and dramatic interaction context"} +{"pid": "13b76692-269c-4a52-b9d8-04b572569076", "context": "Coalitional games are widely studied, but they traditionally assume full information about each agent's capabilities or types, which does not effectively capture real-world scenarios where agents have private beliefs about others.", "key_idea": "The paper introduces coalitional games with beliefs (CGBs), a generalization of coalitional games where agents have private beliefs about the capabilities or types of other agents.", "method": "The authors study coalitional stability in this new setting, introducing a notion of the core for CGBs with and without coalition structures. A characterization of the core is provided for simple CGBs without coalition structures, along with a polynomial-time algorithm to check core nonemptiness. The study also explores complexity implications in games with coalition structures.", "outcome": "The study finds that in games without coalition structures, the core of CGBs can be analyzed in a way that matches the full information case, and core nonemptiness can be checked in polynomial time. However, when coalitions structures are introduced, allowing for agents' beliefs increases the complexity of stability-related problems.", "future_impact": "The introduced model could be used to represent more complex real-world cases in coalitional games and could establish connections with other classes of coalitional games. In particular, the work on weighted voting games with beliefs could be of independent interest.", "venue": "IJCAI", "year": 2009, "title": "Simple coalitional games with beliefs"} +{"pid": "ef829e79-f13b-4f1d-bd89-4e25d0cbe5d4", "context": "Inferencing for SROEL(\u00d7), a description logic that includes the main features of OWL EL, has been a topic of recent research but does not yet have efficient rule-based deduction systems for various reasoning tasks and logical sublanguages.", "key_idea": "The authors have developed rule-based deduction systems for various reasoning tasks and logical sublanguages within SROEL(\u00d7), a subset of OWL EL.", "method": "The authors review recent results on inferencing for SROEL(\u00d7) and evaluate the efficiency of the developed systems.", "outcome": "The study found that certain feature combinations lead to increased space upper bounds for materialisation, suggesting that efficient implementations are easier to obtain for suitable fragments of OWL EL.", "future_impact": "The discovery about feature combinations could guide the development of more efficient implementations for OWL EL.", "venue": "IJCAI", "year": 2011, "title": "Efficient rule-based inferencing for OWL EL"} +{"pid": "60da8fc20abde95dc965f8a3", "context": "The current strategies deployed to solve dynamic games focus on equilibrium refinement but are limited in their scalability and applicability.", "key_idea": "The authors propose Temporal-Induced Self-Play (TISP), a novel reinforcement learning-based algorithm that uses belief-space representation, backward induction, policy learning, and non-parametric approximation to find strategies with good performances from any point in the game onwards.", "method": "The TISP-based policy-gradient-based algorithm, TISP-PG, is tested in various games, including finite repeated security games and a grid-world game.", "outcome": "The authors found that TISP-PG is more scalable than existing mathematical programming-based methods and significantly outperforms other learning-based methods.", "future_impact": "The TISP-based algorithms can find approximate Perfect Bayesian Equilibrium in zero-sum one-sided stochastic Bayesian games with finite horizon, suggesting their potential impact in constructing improved solutions for dynamic games.", "venue": "IJCAI", "year": 2021, "title": "Temporal Induced Self-Play for Stochastic Bayesian Games."} +{"pid": "65b87c5c-36f9-46c6-9980-9827f2a9c2b7", "context": "Previous investigations into contingent planning have explored the use of various state representations, but the relative effectiveness of Conjunctive Normal Form (CNF) and Disjunctive Normal Form (DNF) representations remains unclear.", "key_idea": "This study develops a new contingent planner, CNFct, which extends a previous representation of CNF for conformant planning to handle nondeterministic and sensing actions for contingent planning, and also proposes a new heuristic function for both CNF and DNF based contingent planners.", "method": "The researchers conduct experiments using the new planner, CNFct, along with DNFct. Then, they analyze their performance on a series of benchmarks.", "outcome": "The study found that both CNFct and DNFct offer very competitive performance across a range of benchmarks, though neither definitively outperforms the other. The paper identifies certain properties of the representation schemes that can influence their performance on diverse problems.", "future_impact": "The insights gained on the effect of representation schemes on planner performance could guide the design of more effective contingent planning systems in the future.", "venue": "IJCAI", "year": 2011, "title": "On the effectiveness of CNF and DNF representations in contingent planning"} +{"pid": "60a79d0a91e011f90a51de81", "context": "Modern machine learning algorithms, which rely on several design decisions, render the problem of Hyperparameter Optimization (HPO) crucial, necessitating an efficient approach to handle it.", "key_idea": "The authors propose DEHB, a novel HPO method that combines the advantages of Hyperband and Differential Evolution.", "method": "The authors conduct comprehensive experiments on a wide range of HPO problems and tabular benchmarks from neural architecture search to assess the efficacy of DEHB.", "outcome": "DEHB outperforms previous HPO methods, especially for high-dimensional problems with discrete input dimensions, and is up to 1000 times faster than random search.", "future_impact": "DEHB's robust performance, efficiency, simplicity, and easy implementation positions it well to become a new default HPO method in machine learning.", "venue": "IJCAI", "year": 2021, "title": "DEHB - Evolutionary Hyberband for Scalable, Robust and Efficient Hyperparameter Optimization."} +{"pid": "834cb2c3-c4a8-419d-9ce9-efbfd504d3f3", "context": "Most research on narrative text summarisation has been conducted within the paradigm of experimental psychology, but recent language processing research in artificial intelligence suggests that the predominant theory of text summarisation requires further examination.", "key_idea": "This paper interrogates whether the story grammar model, stemming from psychology, or the plot unit model, developed within AI, offers a better framework for text summarisation.", "method": "Summary data from 72 subjects is presented and analyzed in terms of two summarization models: the story grammar model of psychology and the plot unit model developed in artificial intelligence. The predictions of the two models are compared for short term summarization.", "outcome": "The authors identify two complicating factors that have a significant impact on summarisation behavior, indicating that neither model alone fully explains the experimental data.", "future_impact": "The findings call for further exploration of the impacts of the two identified factors and refinement of existing models to better understand text summarization.", "venue": "IJCAI", "year": 1981, "title": "Summarizing narratives"} +{"pid": "b4e36e89-e531-40cb-8dfd-c7f281b0e991", "context": "The concept of composing tuples from two relations to derive additional tuples of one of these relations is an existing approach, but the determination of when the composition is plausible and for which relation the new tuples are derived is not clearly defined.", "key_idea": "The authors introduce an extended composition approach to inferencing, that involves a formal definition of composition and an extension to the existing concept of composition.", "method": "The authors define conditions on the domains and ranges of the relations that are necessary for extended composition to occur, and propose an algebra that combines primitives to determine plausible composition. The method for extended composition is then applied to a representative group of semantic relations.", "outcome": "The authors successfully developed a methodology for determining plausible composition in the context of the composition of tuples from two relations. This method was evaluated by applying it to a representative group of semantic relations but the specific results are not mentioned.", "future_impact": "The authors anticipate that their method for extended composition may influence future inferencing approaches involving the composition of tuples from two relations, though they do not explicitly state future studies or impacts.", "venue": "IJCAI", "year": 1989, "title": "Plausible inferencing using extended composition"} +{"pid": "5ef96b048806af6ef277204d", "context": "The research work addresses the challenging task of video-language segmentation, where the goal is to segment the object or actor described by a sentence in video frames. The description often refers to multiple attributes, including nearby objects with spatial relations, making the task complex.", "key_idea": "The paper proposes a novel mechanism, Polar Relative Positional Encoding (PRPE), that represents spatial relations in terms of direction and range, allowing for a more direct interaction between sentence features and positional embeddings.", "method": "The authors employ the PRPE to design a Polar Attention Module (PAM), a basic module for vision-language fusion, and test it on the A2D Sentences dataset and the J-HMDB Sentences dataset.", "outcome": "The proposed method outperforms the previous best method by a large margin, with an 11.4% absolute improvement in Mean Average Precision (mAP) on the A2D Sentences dataset. It also achieves competitive performances on the J-HMDB Sentences dataset.", "future_impact": "While not explicitly mentioned by the authors, the significant improvement in video-language segmentation obtained by the proposed method could potentially influence future research and applications in the field of video processing and vision-language tasks.", "venue": "IJCAI", "year": 2020, "title": "Polar Relative Positional Encoding for Video-Language Segmentation"} +{"pid": "60913f7091e01105f877fd20", "context": "Despite the success of Graph Neural Networks (GNNs) in graph representation learning, adapting GNN-variants to hypergraphs, which allow for better modelling of higher-order correlations among entities, remains a challenge due to the lack of a suitable framework.", "key_idea": "This paper proposes UniGNN - a unified framework which interprets the message passing process in both graph and hypergraph neural networks, enabling the generalization of established GNN models to hypergraphs.", "method": "The authors conduct extensive experiments on multiple real-world datasets, and also compare the distinguishing capacities of UniGNN models to the 1-dimensional Generalized Weisfeiler-Leman (1-GWL) algorithm.", "outcome": "The UniGNN framework, applied to multiple datasets, outperforms state-of-the-art approaches, particularly increasing accuracy from 77.4% to 88.8% in semi-supervised hypernode classification on the DBLP dataset. It is further demonstrated that UniGNN models based on message passing are at most as powerful as the 1-GWL algorithm in distinguishing non-isomorphic hypergraphs.", "future_impact": "This work provides a modality for better understanding and adapting GNN models to hypergraphs; the extension of GNNs to hypergraphs could be instrumental in more efficiently addressing problems requiring the modeling of higher-order correlations among entities.", "venue": "IJCAI", "year": 2021, "title": "UniGNN - a Unified Framework for Graph and Hypergraph Neural Networks."} +{"pid": "60ab853291e011e7dcc857ac", "context": "Random Forests (RFs) are widely used in Machine Learning but are not interpretable, and there are no dedicated non-heuristic approaches for explaining RFs. Additionally, while recent work has proposed polynomial algorithms for explaining other ML models, it is unclear if explanations for RFs can be solved in polynomial time.", "key_idea": "The thesis of this paper is twofold; firstly, it proves that finding a PI-explanation for RFs is D^P-complete, thereby making it non-trivial. Secondly, it proposes a propositional encoding method aimed at finding PI-explanations with a SAT solver, presenting an alternative to previous methods that rely on SMT/MILP encodings.", "method": "The authors use a SAT solver alongside a propositional encoding to find explanations for RFs. They carry out experiments on a variety of publicly available data sets to assess the scalability and effectiveness of their proposed approach.", "outcome": "The SAT-based method was shown to scale well with commonly used RF sizes and significantly outperform existing heuristic methods in most considered examples.", "future_impact": "The results of this study indicate that the SAT-based approach could improve the interpretability of RFs in practical applications.", "venue": "IJCAI", "year": 2021, "title": "On Explaining Random Forests with SAT."} +{"pid": "bb8c5935-1c1e-49c3-bab5-e8a28a94f14e", "context": "The doctrine of double effect (DDE) is a long-studied ethical principle that governs when actions that have both positive and negative effects are to be allowed. However, the automation of this doctrine remained unaddressed.", "key_idea": "The authors present an attempt to automate the DDE using a first-order modal logic, the deontic cognitive event calculus. They extend this to stronger versions of the principle and aim to use this approach for building DDE-compliant autonomous systems or verifying DDE compliance in existing AI systems.", "method": "The authors formalize the DDE and present its formalizations in the form of increasingly stronger versions, including the doctrine of triple effect. They then simulate scenarios that have been used to test for the presence of the principle in human subjects and sketch initial work on applying their DDE layer to the STRIPS-style planning model and a modified POMDP model.", "outcome": "The authors successfully use their framework to simulate scenarios to test for DDE presence and provide initial application sketches of the DDE layer to STRIPS-style planning model and a modified POMDP model, which show the feasibility of the intended use.", "future_impact": "The presented work can be useful for other researchers in incorporating DDE in their own frameworks, ultimately allowing the creation of AI systems that comply with DDE or verifying DDE compliance in existing AI systems.", "venue": "IJCAI", "year": 2017, "title": "On automating the doctrine of double effect"} +{"pid": "0ac4f44e-523c-43a0-a370-a01db6868c04", "context": "Online stochastic multiple vehicle routing with time windows, where requests arrive dynamically and the goal is to maximize the number of serviced customers, has typically involved moving vehicles only to known customers.", "key_idea": "This paper proposes investigating waiting and relocation strategies for vehicle routing, in which vehicles may wait at their current location or relocate to arbitrary sites, as opposed to only moving to known customers.", "method": "The paper uses an experimental approach to test the efficacy of these proposed waiting and relocation strategies in improving customer service in highly dynamic problems with many late requests.", "outcome": "Experimental results show that waiting and relocation strategies may dramatically improve customer service, especially for problems that are highly dynamic and contain many late requests.", "future_impact": "The proposed waiting and relocation strategies could set a foundation for developing more effective online algorithms for vehicle routing, without the need to exploit any problem-specific features.", "venue": "IJCAI", "year": 2007, "title": "Waiting and relocation strategies in online stochastic vehicle routing"} +{"pid": "fe16466e-fcc1-47ad-8ea1-d7fc322dc347", "context": "Prior knowledge acquisition systems typically relied on rule-based representations, which may not be optimal for capturing domain-familiar concepts or facilitating easily-understood debugging techniques.", "key_idea": "The authors introduce TEST (Troubleshooting Expert System Tool), an architecture that utilizes a schematic representation, leading to a knowledge base characterized by the use of domain-familiar concepts.", "method": "The methodology of this study mainly involves the development of TDE (TEST Development Environment), which is characterized by features such as facilitating knowledge base development through interviews and direct icon manipulation, multiple knowledge-based browsing strategies, heuristic error analysis, and easily-understood debugging techniques.", "outcome": "TEST has led to a more efficient development of TDE, exhibited by its multiple features, such as direct interviews, heuristic error analysis, and debugging techniques.", "future_impact": "The novel schematic representation in the TEST architecture can facilitate more user-friendly knowledge acquisition systems development and analysis techniques in the future.", "venue": "IJCAI", "year": 1987, "title": "From application shell to knowledge acquisition system"} +{"pid": "8dea7c35-28b0-4539-93b0-fae47bc526b0", "context": "Learning for kernel machines with differentiable loss functions is still a time-consuming process, and the optimization problem conditioning is not ideal, while existing techniques do not provide satisfactory performance.", "key_idea": "The authors propose the Kernel Conjugate Gradient (KCG), a novel variant of the conjugate gradient algorithm that aims to both streamline learning for kernel machines with differentiable loss functions and condition the optimization problem better.", "method": "The effectiveness of KCG is evaluated by applying it to various differentiable kernel learning problems and comparing its performance with existing techniques.", "outcome": "It is observed that Kernel Conjugate Gradient consistently and significantly outperforms existing techniques in differentiable kernel learning problems.", "future_impact": "Data-structure techniques recently used to speed up kernel machine approaches are compatible with this algorithm, offering potential for future efficiency improvements in training, particularly in function evaluation and RKHS inner product computation.", "venue": "IJCAI", "year": 2007, "title": "Kernel conjugate gradient for fast kernel machines"} +{"pid": "5ef5d473-9cb7-41c5-ba15-4e7bd1e0e4b3", "context": "The Knuth-Bendix completion procedure was introduced as a means of deriving canonical term-rewriting systems to serve as decision procedures for given equational theories.", "key_idea": "The authors propose using the Knuth-Bendix completion procedure to synthesize both logic programs and functional programs from specifications and domain knowledge expressed as equivalence-preserving rewrite rules.", "method": "The authors plan to implement their proposed technique on specifications and domain knowledge provided in the form of equivalence-preserving rewrite rules.", "outcome": "An implementation of the proposed idea has begun, but no specific outcomes or results are reported.", "future_impact": "The research may establish a novel way to use term-rewriting systems for programming tasks, although this potential is not explicitly addressed in the abstract.", "venue": "IJCAI", "year": 1985, "title": "Synthesis by completion"} +{"pid": "938e819a-a766-4a2e-b6c8-1acd24faf502", "context": "Testing embedded software systems on control units of vehicles is an essential and safety-relevant task, however, developing test suites for these tests is a time-consuming process.", "key_idea": "The authors propose an automation process for generating tests for control software based on a specification of requirements in terms of finite state machines. They describe converting these state machine representations into relational models for the application of test generation algorithms.", "method": "The authors apply the test generation algorithm to a real example of control software for vehicle control units and analyze the results.", "outcome": "The authors successfully present the transformation of a finite state machine representation into a relational model and the application of the test generation algorithm to a real-world example. However, the abstract does not provide specific numerical results.", "future_impact": "The paper hints at specific challenges regarding software testing, indicating that further research may focus on addressing these challenges.", "venue": "IJCAI", "year": 2007, "title": "Fault-model-based test generation for embedded software"} +{"pid": "12a0782f-811d-4962-92d2-2621650434b8", "context": "The focus on algorithmic recommendation process for children in both research and industries is growing, and a diverse community around this topic has started to form.", "key_idea": "The purpose of the 2nd Workshop of the International and Interdisciplinary Perspectives on Children & Recommender Systems (KidRec 2018) was to gather a diverse group of participants to continue to build community around this important topic and to facilitate discussions on salient research questions and the next steps for this area.", "method": "The workshop invited a diverse group of participants from K-12 teachers, educators, computer science researchers, and industry participants, and organized discussions.", "outcome": "The workshop resulted in building community around the topic, further discussing and outlining the salient research questions and identifying the next steps for promoting more research in this area.", "future_impact": "The community\u2019s efforts will promote more research in the area of children's recommendation systems.", "venue": "SIGIR", "year": 2019, "title": "Building Community: Report on the 2nd International and Interdisciplinary Perspectives on Children & Recommender Systems (KidRec) at IDC 2018"} +{"pid": "3fc30a55-7cc5-4c0b-98b6-72254664b9d1", "context": "Image retrieval by text query is a challenging problem and current methods might not be providing satisfactory performance.", "key_idea": "This paper formulates image retrieval by text query as a vector space classification problem, creating a high-dimensional visual vocabulary to represent image documents in great detail.", "method": "The authors apply well-known text retrieval techniques such as Rocchio tf-idf and naive Bayes to the semantic image retrieval problem, and test these methods on a subset of Corel images.", "outcome": "The proposed approach achieved state-of-the-art retrieval performance on a subset of Corel images.", "future_impact": "The proposed methods have the potential to advance the field of image retrieval using text queries.", "venue": "SIGIR", "year": 2007, "title": "High-dimensional visual vocabularies for image retrieval"} +{"pid": "fa3bde89-e25e-485f-aa20-543a3473d63d", "context": "The problem of systematically identifying anomalies in music datasets in an unsupervised manner remains unresolved.", "key_idea": "The authors present an unsupervised method for anomaly detection in music datasets integrating categorical regression and robust estimation techniques designed to infer anomalous scores in music clips.", "method": "The authors apply the method to a music genre recognition dataset to test its ability to detect corrupted, distorted, or mislabeled audio samples based on commonly used features in music information retrieval.", "outcome": "Evaluation results show the algorithm surpasses other anomaly detection methods and is capable of finding problematic samples that were identified by human experts.", "future_impact": "The proposed method provides a preliminary framework for anomaly detection in music data, potentially serving as a useful tool for improving data integrity in future.", "venue": "SIGIR", "year": 2016, "title": "An Unsupervised Approach to Anomaly Detection in Music Datasets"} +{"pid": "60b9a59de4510cd7c8fd28a8", "context": "Accurate skill retrieval is a key factor for the success of modern conversational AI agents, but the major challenges lie in the ambiguity of human spoken language and the wide spectrum of candidate skills.", "key_idea": "The authors propose a self-adaptive dialogue system (AdaDial) for conversational AI agents that implements a user feedback enhanced reranking strategy.", "method": "The authors use a scalable schema for user feedback estimation and a feedback enhanced reranking model with customized feature encoding, target attention based feature assembling, and multi-task learning.", "outcome": "Online experimental results indicate that AdaDial can retrieve desired skills for different users in different scenarios and correct its regular strategy according to negative feedback. AdaDial is deployed on a large-scale conversational AI agents with tens of millions daily queries, enhancing user experience.", "future_impact": "The method brings continued positive impacts on user experience of large-scale conversational AI agents, and potentially further studies could improve the self-adaptive capability.", "venue": "SIGIR", "year": 2021, "title": "User Feedback and Ranking in-a-Loop: Towards Self-Adaptive Dialogue Systems"} +{"pid": "187c5ed2-c2fb-41c3-9b93-23c6cfdb1bb6", "context": "Currently, fulltext indexing is extensively used in information retrieval systems.", "key_idea": "The authors propose the use of generic summaries for indexing in information retrieval, indicating that they can be as effective as fulltext indexing for precision-oriented search of highly relevant documents.", "method": "The study investigates the effectiveness of a summary index with or without pseudo-relevance feedback and compare it with the corresponding fulltext index. Additionally, it investigates the combination of using a summary index at the initial search and a fulltext index at the final search.", "outcome": "The authors find that a summary index, ideally using a compression ratio of 10-30%, can be as effective as fulltext indexing for precision-oriented searches. Moreover, the combination of a summary index at initial search and a fulltext index at final search proves effective, likely better than just using the summary index.", "future_impact": "This research suggests that the strategy of combining summary and fulltext indexing could potentially improve precision-oriented search outcomes in information retrieval.", "venue": "SIGIR", "year": 2001, "title": "Generic summaries for indexing in information retrieval"} +{"pid": "13c1a948-61e0-4ed7-9f0b-806507062cea", "context": "The development of entity-relationship (E-R) search techniques has been hindered because of a shortage of test collections, particularly for queries that involve multiple entities and relationships.", "key_idea": "The authors present a method for creating E-R test queries from content available in tabular form where columns symbolize entity types and the structure implies one or more relationships among the entities, and the creation of the RELink(test collection).", "method": "Queries and relevance judgments are formulated from table contents. Editorial labor involves the creation of natural language queries based on relations represented by table entries. The authors create the RELink Framework that includes Apache Lucene indexing and search specifically tailored to E-R retrieval.", "outcome": "A new test collection, called RELink, comprising 600 queries and relevance judgements was created from a sample of Wikipedia List-of-lists-of-lists tables. The authors also developed the open-source RELink Framework, which provides specific Apache Lucene indexing and search for E-R retrieval.", "future_impact": "The creation of the 'RELink' framework can expedite entity-relationship research by providing ready-to-use search resources and a comprehensive test collection.", "venue": "SIGIR", "year": 2017, "title": "RELink: A Research Framework and Test Collection for Entity-Relationship Retrieval"} +{"pid": "9f7bb24f-a3fc-4b84-afbe-451c72c904ea", "context": "ECIR, the European Conference on Information Retrieval, is a main forum for the presentation of research results in information retrieval and related areas. Since 1998, it has alternated between the UK and continental Europe, but this rule was removed starting from 2012.", "key_idea": "The 35th ECIR conference took place in Moscow, Russia, marking the first time it was held in the east, introducing new features like a two-tier program committee, a redesigned mentoring program, and a session for poster presentations.", "method": "The conference received submissions of full papers, posters, and demonstrations. All submissions were reviewed by members of an international two-tier Program Committee. Selection for oral presentation and poster/short presentation was also conducted.", "outcome": "ECIR 2013 attracted a high number of participants and accepted papers, leading to a 900-page proceedings volume. It received a total of 287 submissions across three categories (191 full papers, 78 posters, and 18 demonstrations) and accepted 30 for oral presentation, 25 for poster/short presentation, 38 posters and 10 demonstrations.", "future_impact": "The successful organization and increased number of participants of ECIR 2013, as well as the introduction of new features in the conference, signals a potential for further development and expansion in the future.", "venue": "SIGIR", "year": 2013, "title": "ECIR 2013: 35th european conference on information retrieval"} +{"pid": "7047a008-d356-4e35-8fa3-13a59b445986", "context": "The explosion of information on the World Wide Web has resulted in an ongoing challenge of efficiently accessing this vast amount of data. Information Retrieval (IR) tackles this challenge by exploring approaches to obtain relevant information items from a large collection of textual data. Recently, the potential contribution of entities extracted from large volumes of free text to retrieval effectiveness has become an active area of exploration in the research community.", "key_idea": "This thesis investigates leveraging entities to improve retrieval in several ways. It proposes a new paradigm of finding entities with certain semantic relations to directly meet the user's information needs, unlike the traditional document retrieval. The thesis also introduces a novel probabilistic framework that uses supporting documents to model the relevance between queries and entities.", "method": "The research includes retrieval experiments such as entity-centric query expansion, where related entities of a query are identified and query expansion is performed using the names and relations of these entities. It also includes entity-centric relevance modeling, where the Latent Entity Space (LES) retrieval approach is proposed that uses entity profiles to represent semantic content of documents and queries.", "outcome": "The study shows significant improvements over several state-of-the-art feedback models on multiple data collections in entity-centric query expansion. In entity-centric relevance modeling, LES effectively captures latent semantic content and improves the search accuracy of several state-of-the-art retrieval models for entity-bearing queries.", "future_impact": "The thesis reveals the potential benefits of using entities to improve retrieval effectiveness. With the fast curation of high-quality knowledge bases, it is expected that more information about entities could be easily integrated into retrieval models and applied to more applications. This work could serve as guidance for future research leveraging entities to improve information retrieval.", "venue": "SIGIR", "year": 2016, "title": "Entity Centric Information Retrieval"} +{"pid": "b20664d0-1e45-4382-a6e5-ecf8faf9d338", "context": "The process of remembering and entering query terms for information retrieval from TV watching activities can be burdensome.", "key_idea": "The researchers conduct a study involving a one-button search interface named Video Bookmarking Search, which utilizes closed captions from watching TV to estimate query terms through named-entity recognition and sentence labeling techniques, alleviating the need to manually enter queries.", "method": "The researchers conducted an empirical evaluation with 1,138 search queries from 206 bookmarks using seven actual TV shows on city life, travel, health, and cuisine.", "outcome": "The results demonstrated that the one-button search resulted in a wider variety of queries and search results, even though the number of directly relevant results was not high.", "future_impact": "While not directly stated, this research may inform future development of more user-friendly information retrieval interfaces that can interpret and respond to user interests beyond specific query terms.", "venue": "SIGIR", "year": 2008, "title": "One-button search extracts wider interests: an empirical study with video bookmarking search"} +{"pid": "fcebb341-79d8-4a3b-be45-7024a7fec6a1", "context": "The goal of search result diversification is to select a subset of documents from the candidate set to satisfy as many different subtopics as possible. Existing methods often handle this challenging NP-hard problem by ranking the documents with greedy sequential document selection, often leading to suboptimal rankings.", "key_idea": "The authors propose Monte Carlo tree search (MCTS) enhanced Markov decision process (MDP), referred to as M$^2$Div, which formalizes diverse ranking construction as an MDP, and uses a recurrent neural network to guide document selection and predict ranking quality.", "method": "The proposed M$^2$Div approach, including the use of a recurrent neural network and MCTS for document selection and predicting ranking quality, is tested using the TREC benchmarks.", "outcome": "Experimental results using TREC benchmarks showed that the proposed M$^2$Div approach can significantly outperform the state-of-the-art baselines based on greedy sequential document selection.", "future_impact": "The demonstrated effectiveness of the exploratory decision-making mechanism in M$^2$Div could inspire new approaches in the subset selection and document ranking problem space.", "venue": "SIGIR", "year": 2018, "title": "From Greedy Selection to Exploratory Decision-Making: Diverse Ranking with Policy-Value Networks"} +{"pid": "a6d0168e-b6b1-420f-b9a0-5b9de7a688da", "context": "Entity disambiguation, which is the task of mapping ambiguous terms in natural-language text to its entities in a knowledge base, is crucial in applications like extraction of structured data from textual documents and in facilitating AI applications such as Semantic Search, Reasoning and Question & Answering.", "key_idea": "The authors propose a novel collective, graph-based disambiguation algorithm utilizing semantic entity and document embeddings for robust entity disambiguation, which is able to achieve better results over a wide range of data sets.", "method": "The authors evaluate their proposed algorithm on 9 different datasets, demonstrating its performance without dataset-specific tuning.", "outcome": "Their evaluation indicates that their approach achieves significantly (>5%) better results than all other publicly available disambiguation algorithms on 7 of 9 datasets.", "future_impact": "The authors discuss the influence of the quality of the knowledge base on the disambiguation accuracy, suggesting that the proposed algorithm has potential to achieve better results than non-publicly available state-of-the-art algorithms as the quality of the knowledge base improves.", "venue": "SIGIR", "year": 2016, "title": "Robust and Collective Entity Disambiguation through Semantic Embeddings"} +{"pid": "62de85685aee126c0f974c5f", "context": "Neural approaches that use pre-trained language models are effective at various ranking tasks. However, there is a gap in comparing the effectiveness of these neural approaches and feature-based Learning-to-Rank (LtR) methods, due to differences in how these methods are evaluated: LtR benchmarks contain feature vectors but no raw text, while neural models are evaluated on text but no feature vectors.", "key_idea": "The paper introduces Istella22, a novel dataset that bridges this gap by including both query-document text and query-document feature vectors, enabling a fair comparison of both traditional LtR and transfer ranking techniques on the same data.", "method": "The authors conducted preliminary experiments on Istella22, a dataset consisting of a comprehensive corpus of 8.4M web documents, relevance judgments, and a set of 2,198 textual queries used for testing purposes.", "outcome": "Preliminary experiments on Istella22 found that neural re-ranking approaches were less effective than LtR models, but that LtR models considered the scores from neural models as strong signals.", "future_impact": "Istella22, with its combination of query-document text and feature vectors, provides a platform to more fairly evaluate traditional learning-to-rank and transfer ranking techniques, and further our understanding of their strengths and weaknesses.", "venue": "SIGIR", "year": 2022, "title": "The Istella22 Dataset: Bridging Traditional and Neural Learning to Rank Evaluation"} +{"pid": "5875218c-a285-478f-a9fe-094b81acde5c", "context": "In evidence-based medicine, PICO (Patient, Intervention, Comparator, and Outcome) annotations in medical literature are used for automatic relevant document filtering. However, defining boundaries for PICO elements in text is challenging due to diverse interpretations.", "key_idea": "This paper focuses on analyzing the agreement of PICO annotations made by multiple human annotators, both experts and non-experts, using a standard span agreement method (matching both labels and boundaries of text spans) and two types of relaxed span agreement methods (matching labels without guaranteeing matching boundaries).", "method": "The authors performed an analysis on the agreement of PICO span annotations made by different human annotators using standard and relaxed span agreement methods.", "outcome": "The study found that boundaries of PICO span annotations by individual annotators are very diverse, and despite the disagreement in span boundaries, general areas of the span annotations are broadly agreed by annotators. The results also revealed that using standard agreement alone may undermine the agreement of PICO spans.", "future_impact": "This study suggests that adopting both standard and relaxed agreements may be more suitable for PICO span evaluation, which can potentially improve the quality and reliability of PICO annotations in evidence-based medicine.", "venue": "SIGIR", "year": 2019, "title": "A Study on Agreement in PICO Span Annotations"} +{"pid": "c7ecf58d-f881-4a5b-9a49-4c24ae6e104f", "context": "Most major application (app) marketplaces rank and recommend apps based on search relevance features or marketplace popularity statistics, such as similarity between app descriptions and user search queries, or the number of downloads and user ratings. However, these rankings do not capture the dynamics of the app ecosystem, including the dependencies between apps.", "key_idea": "In this paper, the authors propose the concept of interoperability ranking for mobile applications. It ranks apps based on their importance to other apps in the ecosystem and can answer questions such as which app is more likely to be launched in a particular user context and whether one app provides complementary functionality to another.", "method": "The authors outline a set of methods for computing interoperability ranks, and validate these approaches using real data from the Windows Phone app marketplace.", "outcome": "The paper does not provide specific outcomes or results of the analysis performed on the real data from the Windows Phone app marketplace.", "future_impact": "The authors anticipate that the notion of interoperability ranking could contribute to solving problems attracting the attention of both researchers and industry, such as building personalized real-time app recommender systems or intelligent mobile agents.", "venue": "SIGIR", "year": 2013, "title": "Interoperability ranking for mobile applications"} +{"pid": "33fc2f76-613c-4ba4-99b3-6574333ced9b", "context": "The current signature file organizations employ either key-based or bit-sliced partitioning strategies, respective performance of which is quite bad for handling queries having various number of query terms.", "key_idea": "A new signature file organization is proposed that combines key-based and bit-sliced partitioning strategies for a more efficient search performance for queries with an arbitrary number of terms.", "method": "The design of the new organization is based on theoretical analysis of key-based and bit-sliced approaches, taking into account an abstract storage structure model.", "outcome": "The proposed approach achieves high search performance for queries containing an arbitrary number of terms, implying an improvement over the limitations of bit-sliced or key-based organization individually.", "future_impact": "Discussion of update performance and a generalization of the method show potential for future research to further adjust the trade-off between search efficiency and maintenance costs.", "venue": "SIGIR", "year": 1995, "title": "Key-based partitioned bit-sliced signature file"} +{"pid": "8d690d4c-7e5e-4b9c-9626-e8abe020d0b8", "context": "Web search engines, which millions of people use daily, employ ranking algorithms that greatly influence which pages users visit. However, there is limited understanding of how these algorithms favor different sites or points of view.", "key_idea": "The authors present PAWS, a platform to analyze differences among web search engines by measuring content emphasis i.e., the correlation between differences in search engine rankings and features of the ranked content (like point of view and advertisements). They also propose a novel technique to identify orientations in search results at scale, minimizing the expected number of human judgments required.", "method": "PAWS was applied to analyze news searches on Google and Bing, assessing the engines' emphasis on results expressing different orientations and containing different types of advertisements.", "outcome": "The authors found no evidence that Google or Bing emphasize results expressing positive orientation towards the engine company's products. However, they found that these engines do emphasize particular news sites and favor pages containing their company's advertisements rather than competitor advertisements.", "future_impact": "PAWS could help users gain more transparent access to the Web's information by informing them when different engines favor certain sites or points of view.", "venue": "SIGIR", "year": 2014, "title": "Analyzing the content emphasis of web search engines"} +{"pid": "62de85525aee126c0f97432b", "context": "Customers find useful tips extracted from product reviews helpful for informed purchase decisions and better product usage. However, the validity of these tips has not been thoroughly examined based on the degree of support or opposition they receive from all product reviews.", "key_idea": "The authors propose a classifier to determine the degree to which product tips are supported or contradicted by review sentences, creating aggregated global support and contradiction scores that reflect the overall support-level of all reviews for a specific tip.", "method": "By analyzing a large set of tips extracted from product reviews, the classifier assigns them to one of the proposed categories: highly-supported, highly-contradicted, controversial (both supported and contradicted), and anecdotal (neither supported nor contradicted).", "outcome": "The system's classification provides a global support score and a global contradiction score, improving customer confidence in the validity of a given tip.", "future_impact": "With this method, customers could make more informed purchase decisions and better use of products while potentially influencing future e-commerce platforms to incorporate similar features.", "venue": "SIGIR", "year": 2022, "title": "Analyzing the Support Level for Tips Extracted from Product Reviews"} +{"pid": "5f1d5a5a9fced0a24b59be02", "context": "While the World Wide Web provides a large amount of text in many languages, cross-lingual parallel data is difficult to obtain. It is however crucial for a variety of tasks in natural language processing with applications in machine translation, cross-lingual information retrieval, and document classification, as well as learning cross-lingual representations.", "key_idea": "The authors propose the end-to-end process of searching the web for parallel cross-lingual text. They motivate obtaining parallel text as a retrieval problem and introduce techniques for searching for cross-lingual parallel data based on language, content, and other metadata.", "method": "The authors introduce multilingual sentence embeddings as a core tool, demonstrating techniques and models that leverage them for identifying parallel documents and sentences as well as techniques for retrieving and filtering this data.", "outcome": "The authors describe several large-scale datasets curated using these techniques and show how training on sentences extracted from parallel or comparable documents mined from the Web can improve machine translation models and facilitate cross-lingual NLP.", "future_impact": "This process of retrieving and curating cross-lingual parallel data may impact the performance of future machine translation models and cross-lingual NLP applications.", "venue": "SIGIR", "year": 2020, "title": "Searching the Web for Cross-lingual Parallel Data"} +{"pid": "60b9a581e4510cd7c8fd0b6c", "context": "Previous studies mainly extract set names for a group of entities from knowledge bases which offer high-quality entity relations but struggle with a limited range of entities and set names representing wide semantic classes.", "key_idea": "The authors propose AutoName, an unsupervised framework that generates set names for a set of query entities from a large-scale text corpus. It generates candidate labels by examining a pre-trained language model and the entities' context in documents.", "method": "Candidate phrases are clustered to find those describing common concepts among entities. AutoName then ranks the refined phrases based on the co-occurrences of their words with the query entities and the conceptual integrity of their clusters. A new benchmark dataset consisting of 130 entity sets with name labels was also built for testing purposes.", "outcome": "The experimental results reveal that AutoName can create coherent and meaningful set names, significantly surpassing all baseline methods.", "future_impact": "AutoName would be beneficial in various tasks related to natural language processing and information retrieval such as session-based and conversational information seeking.", "venue": "SIGIR", "year": 2021, "title": "AutoName: A Corpus-Based Set Naming Framework"} +{"pid": "62de85775aee126c0f975205", "context": "Internet memes, known for their non-standard writing and network vocabulary, have become a communicating medium for users worldwide. As most existing research largely ignores the rich metaphorical information in memes, this poses a tricky challenge in natural language processing.", "key_idea": "This paper introduces MET-Meme, a novel multimodal meme dataset specifically enriched with metaphorical features to facilitate better analysis of Internet memes when it comes to sentiment analysis and semantic understanding tasks.", "method": "MET-Meme, consisting of 10045 text-image pairs, is manually annotated with the occurrence of metaphors, sentiment categories, intentions, and degree of offensiveness. The authors propose a range of strong baseline models to showcase the importance of integrating metaphorical features for meme sentiment analysis and semantic understanding.", "outcome": "The authors have publicly released the MET-Meme dataset and its corresponding code for research purposes.", "future_impact": "The introduction of the MET-Meme dataset could significantly enhance research in the area of meme analysis, especially in the context of sentiment analysis and semantic understanding tasks.", "venue": "SIGIR", "year": 2022, "title": "MET-Meme: A Multimodal Meme Dataset Rich in Metaphors"} +{"pid": "7a52ad5d-d68b-4ef8-b166-12c527db33c3", "context": "Query suggestion mechanisms are widely used by search engines, but there is a lack of a commonly accepted evaluation methodology and metrics for this area. The metrics used often lack proper justification and it is unclear if reported improvements result in an actual improvement in the user experience.", "key_idea": "The paper addresses the evaluation issue by creating a new family of user models based on users' interaction with search engines' query log. They also propose two novel metrics, pSaved and eSaved, for evaluating query suggestions based on a query log-based evaluation approach.", "method": "The authors propose to study the users' behaviour from a search engine's query log to develop their new user models and metrics. The effectiveness of pSaved and eSaved was evaluated by comparing with pre-existing metrics on a dataset of 6.1M sessions.", "outcome": "The experimental results showed that the newly proposed metrics, pSaved and eSaved, show the best alignment with the users' satisfaction compared to other metrics considered.", "future_impact": "The new family of user models and the proposed metrics could establish a more reliable standard for evaluating query suggestion mechanisms, potentially leading to more meaningful improvements in users' experience.", "venue": "SIGIR", "year": 2013, "title": "User model-based metrics for offline query suggestion evaluation"} +{"pid": "927b8174-01e4-41f8-914b-b9841664e713", "context": "Metasearch is a process that involves combining the ranked lists of documents returned by multiple search engines for a given query to optimize the performance of the combination. Existing metasearch strategies don't always achieve best performance.", "key_idea": "The paper presents three contributions to metasearch: a model based on the Borda Count, a model based on Bayesian inference, and a model for obtaining upper bounds on the performance of metasearch algorithms.", "method": "The authors investigated these new metasearch models with rigorous experiments, comparing them with the performance of the best input system and existing metasearch strategies.", "outcome": "The experimental results showed that metasearch algorithms based on the proposed models (Borda count and Bayesian inference) frequently outperform the best input system and are competitive with existing metasearch strategies.", "future_impact": "The initial upper bounds demonstrate that there is much to be learned about the limits of metasearch performance, possibly leading to better future strategies or methods.", "venue": "SIGIR", "year": 2001, "title": "Models for metasearch"} +{"pid": "234f1bd9-3525-476f-8d94-769c2ee57b86", "context": "Current user models for evaluating metrics for search engine results fail to account for time spent by the user and various user behaviors when scanning a ranked result list.", "key_idea": "The authors propose a family of new evaluation measures, called Markov Precision (MP), that utilize continuous-time and discrete-time Markov chains to inject user models into precision and account for both time spent and different behaviors in scanning a ranked result list.", "method": "Through the use of Markov models, the authors recreate average precision using specific user models. The reliability of MP is tested with a thorough experimental evaluation on standard TREC collections and an example of calibration of MP's time parameters based on click logs from Yandex.", "outcome": "MP is found to be as reliable as other measures in the TREC collections experiment, and an example of successful calibration of time parameters was demonstrated with the Yandex click logs data.", "future_impact": "The flexibility of the proposed Markov Precision can lead to more realistic user models in evaluating the effectiveness of search engines, improving precision considering different user behaviors and time spent on search results.", "venue": "SIGIR", "year": 2014, "title": "Injecting user models and time into precision via Markov chains"} +{"pid": "37734c20-1a76-49fb-b86b-6f2b258ecda1", "context": "Despite the growth of the Web, user bases, and query traffic, current research lacks in quantifying the performance and quality gains that can be achieved by distributed search engines based on geographical partitioning of a central Web index.", "key_idea": "The paper proposes various cost models to evaluate the performance benefits of a geographically distributed search engine architecture based on partial index replication and query forwarding.", "method": "The authors focus on performance gains in query processing and Web crawling processes in a distributed environment, and examine whether improved response time can be used to enhance search relevance. They also evaluate if distributed Web crawling leads to better Web coverage and enhances search quality. The claim is validated through simulations on large, real-life datasets.", "outcome": "The findings show that response time gain achieved by distributed query processing can be utilized to improve search relevance, and distributed Web crawling can lead to better Web coverage.", "future_impact": "The application of this study can help enable the use of complex but more accurate algorithms for document ranking in distributed search engine architectures.", "venue": "SIGIR", "year": 2009, "title": "Quantifying performance and quality gains in distributed web search engines"} +{"pid": "4e1e5234-48d0-43ff-b142-8eff8628e2ae", "context": "Query expansion, either as pseudo-relevance feedback or relevance feedback, is widely used for enhancing retrieval performance. However, most of the previous approaches overlooked the role of features and the significance of modelling term dependencies.", "key_idea": "The authors propose a robust query expansion technique called 'latent concept expansion', which is based on the Markov random field model for information retrieval. This technique allows for modeling term dependencies during the expansion phase.", "method": "The new technique is evaluated against relevance models, an existing advanced language modeling query expansion technique. The comparative evaluation is performed across multiple TREC datasets.", "outcome": "The proposed model shows consistent and significant improvements in retrieval performance across several TREC datasets.", "future_impact": "Besides boosting retrieval effectiveness, the proposed technique can also aid in generating meaningful multi-term concepts, which is useful for tasks such as query suggestion or reformulation.", "venue": "SIGIR", "year": 2007, "title": "Latent concept expansion using markov random fields"} +{"pid": "4e58e686-a546-4011-a4a9-6f58da58f8d6", "context": "Despite significant improvements in Web search over the last decade, finding required answers for many user needs is still a challenge, as they can be spread across multiple documents or may not exist on the Web. Yet, many of these needs could be addressed by leveraging Community Question Answering (CQA) services, such as Baidu Knows, Quora, or Yahoo! Answers.", "key_idea": "The authors carry out a large-scale analysis of how searchers on a major web search engine turn into askers by posting questions on a popular CQA site when their web search fails to yield satisfactory results.", "method": "The study involves analyzing search engine logs, tracing the transformation of numerous unsuccessful searches into questions posted on a CQA site. The analysis focuses on query characteristics, search behavior patterns preceding the posting of a question, the relationship between the attempted searches and the posted questions, and the subsequent actions the users took on the CQA site.", "outcome": "This research provides novel insights into searcher intent and behavior regarding the transition to posing questions to a community, when web search fails.", "future_impact": "The findings of this paper provide a foundation for more effectively integrating automated web search and social information seeking in the future.", "venue": "SIGIR", "year": 2012, "title": "When web search fails, searchers become askers: understanding the transition"} +{"pid": "5c8222ab-6bd2-4eff-bf66-d41775c78f69", "context": "Previous studies involving information retrieval have not fully considered the cognitive abilities of end-users.", "key_idea": "The authors studied the effects of various cognitive abilities on the performance of university students searching for references using a standard computerized index.", "method": "The cognitive abilities of fifty university students were assessed using eight tests from the Kit of Factor-Referenced Cognitive Tests, and then they were asked to search for references using a standard computerized index. The performance in the searches was analysed using various measures, and effects for demographic characteristics and knowledge were also considered.", "outcome": "It was found that perceptual speed affected the quality of searches, and logical reasoning, verbal comprehension, and spatial scanning abilities influenced search tactics.", "future_impact": "This finding suggests that information retrieval systems could be improved to accommodate users with different levels of cognitive abilities, which can assist users to scan lists of terms, choose appropriate vocabulary for searching, and select useful references.", "venue": "SIGIR", "year": 1992, "title": "Cognitive differences in end user searching of a CD-ROM index"} +{"pid": "5d79a4f43a55ac5af95ae25b", "context": "In the training of large-scale recommender systems with dynamic sparse features, conventional distributed training frameworks are plagued by high communication costs and slow convergence due to the need to transmit large sparse weights across the network.", "key_idea": "The authors propose Distributed Equivalent Substitution (DES) training, which implements fully synchronous training in large-scale recommendation systems for the first time and reduces communication overhead by substitifying weight-rich operators with computationally equivalent sub-operators, and aggregating partial results instead of transmitting huge sparse weights directly.", "method": "The DES system was applied to multiple popular Deep Learning Recommendation Models (DLRMs) and its performance compared with state-of-the-art PS-based training frameworks in terms of communication efficiency and throughput.", "outcome": "Experiments demonstrated that DES training outperforms the state-of-the-art PS-based training framework, achieving up to 68.7% communication savings and delivering higher throughput compared to other PS-based recommender systems.", "future_impact": "The DES training framework will likely streamline the training of commercial recommender systems, enabling them to converge faster and achieve better click-through rates.", "venue": "SIGIR", "year": 2020, "title": "Distributed Equivalent Substitution Training for Large-Scale Recommender Systems"} +{"pid": "d3e1bf26-9ca1-478f-b717-a395604d24e4", "context": "Collecting, organizing and utilizing language specific resources from the Web is a challenging task, especially for minority languages.", "key_idea": "The authors propose CorpusBuilder, an automatic method for generating Web-search queries to collect documents in a minority language. It uses an automatic language classifier to label retrieved documents, and a subset of these documents is used to generate new queries.", "method": "The authors experiment with various query-generation methods and query-lengths to find inclusion/exclusion terms and use the odds-ratio scores calculated over the documents acquired so far to generate queries. They also substitute initial documents with a handful of words elicited from a user to compare the performance.", "outcome": "Using odds-ratio scores calculated over the documents acquired so far was found to be one of the most consistently accurate query-generation methods. The authors also found that the performance of the system is similar even when they use a handful of words elicited from a user instead of initial documents.", "future_impact": "The generalizability of the proposed system to multiple languages suggests its potential to aid in processing and performing tasks in various language contexts.", "venue": "SIGIR", "year": 2001, "title": "Automatic web search query generation to create minority language corpora"} +{"pid": "0b45b897-fb2f-4530-8124-c614ca5ddd3f", "context": "Search engines typically present results in ranked lists, but due to the unprecedented scale of the Web and diversity of search results, this approach has become increasingly inadequate. Clustering, or dividing search results into topic-coherent clusters, is seen as a promising alternative.", "key_idea": "The authors propose ClusteringWiki, the first prototype and framework for personalized clustering that allows users to directly edit clustering results for a more personalized presentation of search results.", "method": "The authors developed ClusteringWiki with a Wiki interface, enabling users to edit and annotate the membership, structure and labels of clusters for personalized presentations.", "outcome": "The authors did not provide specific quantitative outcomes in the abstract.", "future_impact": "The authors suggest that users' edits and annotations can be shared among users to collaboratively improve the organization of search results and overall utility of search engines.", "venue": "SIGIR", "year": 2011, "title": "ClusteringWiki: personalized and collaborative clustering of search results"} +{"pid": "62de88a15aee126c0f98fba1", "context": "Query-Focused Summarization (QFS) aims to extract essential information from long documents and organize it into a summary that can answer a query. Transformer-based summarization models in QFS have been widely used, but they can't directly use the relationships between distant words and information from a query.", "key_idea": "The proposed QSG Transformer is a novel QFS model leveraging structure information on Query-attentive Semantic Graph (QSG) to address the issue of distant word and query information utilization.", "method": "A query-attentive graph attention network is proposed for QSG node representation advancement, dispersing the query node's information into QSG using Personalized PageRank, which is then used to generate a more accurate summary. The proposed method is evaluated on two QFS datasets.", "outcome": "The QSG Transformer model demonstrates superior performance over the state-of-the-art models on the two evaluated QFS datasets.", "future_impact": "This new transformer model could potentially enhance the quality of the Query-Focused Summarization task by onboarding the structure information of Query-attentive Semantic Graph.", "venue": "SIGIR", "year": 2022, "title": "QSG Transformer: Transformer with Query-Attentive Semantic Graph for Query-Focused Summarization"} +{"pid": "c3f458aa-1cde-45a4-9bab-e80eaa0775d2", "context": "The problem of finding nearest neighbors to a query in a document collection is a special case of associative retrieval, in which searches are performed using more than one key. Existing solutions involve sequential search and clustered search, with the latter failing to guarantee the retrieval of the absolute nearest neighbors.", "key_idea": "The paper proposes a novel associative retrieval algorithm suitable for document retrieval using similarity matching. The algorithm operates based on a binary tree structure, where the selection of the most promising branch is determined by testing a set of keys (concepts) at each node.", "method": "The paper evaluates the proposed retrieval algorithm using a series of experiments with a small collection and compares the results with those from sequential search and clustered search algorithms.", "outcome": "The results show that under certain conditions, the proposed algorithm achieves a search time of log2N * k, where N is the number of documents and k is a system-dependent parameter (approximated as 4 in this context). The average search time of this algorithm for large collections is less than that of a sequential search but greater than that of a clustered search.", "future_impact": "The proposed algorithm could be utilized and further optimized for nearest neighbor searching in document retrieval systems, offering a balance between search time and result accuracy.", "venue": "SIGIR", "year": 1978, "title": "A tree algorithm for nearest neighbor searching in document retrieval systems"} +{"pid": "8aa5ddc5-e842-45cd-a245-1bc176fb3e85", "context": "On social media, users send millions of microblogs on varied topics and predicting which topics a user will join in the future can provide insights into prospective popular topics and the types of users they may attract.", "key_idea": "This study introduces a novel deep convolutional neural network with external neural memory and attention mechanism to predict whether a user will join a topic based on their posting history.", "method": "The authors use the proposed deep convolutional neural network to model user posting history and topics with an external neural memory architecture. Relations between users and topics are constructed using convolutional neural network based matching methods, and decisions are made based on these matching results. The method is trained and tested using a large-scale Twitter dataset.", "outcome": "The experimental results show that the proposed method outperforms other methods, achieving an 18.2% improvement in F1-score and 28.9% improvement in MAP@10 compared to state-of-the-art deep neural networks.", "future_impact": "The proposed method for predicting the topic engagement of users could potentially inform interest prediction applications, and help in the identification of rising popular topics on social media.", "venue": "SIGIR", "year": 2017, "title": "Predicting Which Topics You Will Join in the Future on Social Media"} +{"pid": "3a0a9898-135e-4899-b750-9aede6fa6378", "context": "Graphic processors are high-parallel, high-performance shared-memory architectures originally designed for graphical problems. While these architectures are being broadly repurposed to solve a range of issues, sorting is a key but challenging problem in computer science that still can benefit from novel technologies.", "key_idea": "The authors present a new fast sorting algorithm implementing an efficient bitonic sorting network that is highly suitable for information retrieval applications.", "method": "The authors design their algorithm to exploit graphic processors as co-processors, using a memory-efficient data access pattern to limit unecessary memory access. They introduce an efficient instruction dispatch mechanism to improve overall sorting speed while also presenting a cache-based computational model for graphics processors.", "outcome": "Experimental results show marked improvements over both previous CPU-based methods and GPU-based sorting algorithms, indicating faster sorting speed.", "future_impact": "While not explicitly mentioned, it can be inferred that the proposed method can potentially improve the efficiency of information retrieval applications and other use cases reliant on quick sorting of large volumes of data.", "venue": "SIGIR", "year": 2009, "title": "Sorting using BItonic netwoRk wIth CUDA"} +{"pid": "3bb34a92-20ee-4a60-8c91-c3edd7aa4c32", "context": "Most research in full-text information retrieval focuses on automated and semi-automated methods of query expansion to improve the effectiveness of user queries. However, constraints in the computer troubleshooting domain, where user queries are short and contain a mix of natural language terms with technical sublanguages terminology, present unique challenges.", "key_idea": "The authors consider the specific challenges of computer troubleshooting and propose using several heuristic techniques for domain knowledge acquisition that exploit both query log data and corpus analysis.", "method": "The authors discuss and design heuristic techniques for domain knowledge acquisition by exploiting query log data and corpus analysis in the computer troubleshooting domain.", "outcome": "The paper identifies the necessity of domain-specific adaptation of full-text information retrieval system, and concludes that conventional devices for more purely \u201cnatural language\u201d domains may be inadequate for computer troubleshooting.", "future_impact": "The study emphasizes the need for implementing domain-specific adaptations to improve query expansion tools and document indexing, offering potential directions for future research in improving full-text information retrieval system.", "venue": "SIGIR", "year": 1994, "title": "Adapting a full-text information retrieval system to the computer troubleshooting domain"} +{"pid": "43c5bc15-5e74-4cb5-b0e3-a6f02f99e37f", "context": "LinkedIn is the world's largest professional network where one of the primary activities is people search. The role network distance plays in search behavior wasn't previously clear.", "key_idea": "This paper unveils insights about people search behavior on LinkedIn, particularly the role that network distance plays in name searches and non-name searches.", "method": "The study conducts a log analysis and a user study to examine the role of network distance in LinkedIn search behavior.", "outcome": "For name searches, users primarily click on one of the results, and closer network distances lead to higher click-through rates. As for non-name searches, users are likely to click on multiple results that are not in their existing connections, but with whom they have shared connections.", "future_impact": "The revelation that network distance plays a significant role in LinkedIn search engagement, albeit varying with the type of search query, could guide the development of improved search algorithms for social and professional networks.", "venue": "SIGIR", "year": 2014, "title": "The role of network distance in linkedin people search"} +{"pid": "e53dd2d8-a53e-42c8-b57b-084c54a5a203", "context": "Learning to rank for document retrieval can be evaluated with performance measures such as MAP and NDCG, which existing methods only loosely related to due to their focus on minimizing loss functions on instance pairs.", "key_idea": "AdaRank has been proposed as a boosting algorithm to address this issue; it is designed to minimize a loss function directly defined on the performance measures.", "method": "AdaRank constructs 'weak rankers' repeatedly based on reweighted training data and finally linearly combines the weak rankers to make ranking predictions. Experiments were performed on four benchmark datasets to compare AdaRank's performance with baseline methods such as BM25, Ranking SVM, and RankBoost.", "outcome": "AdaRank significantly outperforms baseline methods of BM25, Ranking SVM, and RankBoost on tested benchmark datasets.", "future_impact": "The AdaRank learning algorithm could potentially help improve document retrieval by better optimizing performance measures during the training process.", "venue": "SIGIR", "year": 2007, "title": "AdaRank: a boosting algorithm for information retrieval"} +{"pid": "c7ab61bf-08b4-4a85-86eb-2cea014ccca0", "context": "Information Retrieval (IR) practitioners involved in TREC-like campaigns have to use multiple different software tools and scripts for tasks like building test collections, evaluating systems, or analysing results from empirical experiments, often needing to implement their own ad-hoc scripts.", "key_idea": "TrecTools is introduced as an open-source Python library providing a unified environment for performing common activities related to TREC-like campaigns, aiming to lower the entry barrier and unify existing tools and activities.", "method": "TrecTools is implemented in Python, providing an object-oriented, easily extensible library. It seeks to unify existing tools, frameworks, and activities under one common umbrella while also being distributed as an open source library.", "outcome": "The authors have developed and published TrecTools as an open-source Python library for IR practitioners.", "future_impact": "The adoption of TrecTools as a centralised solution for developing, evaluating, and analysing TREC-like campaigns could ease the burden on organisers and provide participants and users with a standard environment for common IR experimental activities.", "venue": "SIGIR", "year": 2019, "title": "TrecTools: an Open-source Python Library for Information Retrieval Practitioners Involved in TREC-like Campaigns"} +{"pid": "60f2b0bc91e01191f18a7deb", "context": "Cold-start is the most challenging and time-consuming phase when building a question answering based chatbot for a new business scenario due to the need to collect sufficient training data.", "key_idea": "This paper presents AliMe DA, a practical data augmentation (DA) framework consisting of data production, denoising and consumption aimed at addressing the cold-start problem in introducing a new chatbot.", "method": "The authors demonstrate how their DA approach enhances annotation productivity and improves downstream model performance, and they provide best practices for each stage of data augmentation.", "outcome": "The framework improves both the quality (denoising) and quantity (data production) of the training data, as well as the performance of the downstream models (consumption).", "future_impact": "The authors share their observation on the potential uses of data augmentation in the era of pre-trained language models, thus providing guidelines for future applications and developments in this domain.", "venue": "SIGIR", "year": 2021, "title": "AliMe DA: A Data Augmentation Framework for Question Answering in Cold-start Scenarios"} +{"pid": "97c90d9c-5882-43a7-9852-ec3407508d7c", "context": "A key step in conducting systematic reviews in evidence-based medicine is identifying relevant documents from a large pool of potential documents. This often involves manually screening around 2000 candidate documents to find about 20 relevant documents, which is time-consuming and laborious.", "key_idea": "This paper proposes a new model, Seed-Driven Document Ranking (SDR), which optimizes the screening process by ranking documents based on one known relevant document, or 'seed' document. The SDR algorithm uses a bag of clinical terms, not just simple word bag models, and assigns importance to these terms based on their distribution in candidate documents.", "method": "The authors test the performance of the SDR model on a benchmark dataset from the CLEF eHealth Task 2. The model ranking is also combined with word embedding representation of documents to examine the complementarity.", "outcome": "The proposed SDR method outperformed state-of-the-art solutions on the benchmark dataset. Combining relevances estimated by SDR and by word embedding achieved the best performance.", "future_impact": "The SDR model could dramatically improve the manual screening process in systematic reviews, making it more efficient and less laborious for reviewers.", "venue": "SIGIR", "year": 2018, "title": "Seed-driven Document Ranking for Systematic Reviews in Evidence-Based Medicine"} +{"pid": "7b7fc004-9ff9-4485-b8be-695ae1d7c82d", "context": "High precision Information Retrieval (IR) is often challenging due to the large number of morphological variants for any given term. To address mismatches between different word forms used in queries and the relevant documents, existing research proposed the use of various stemming algorithms to reduce terms to a common base form.", "key_idea": "The paper proposes that stemming can assist in specific contexts and suggests an empirical investigation into the relationship between stemming performance and retrieval performance. The study extends previous work with a novel, dictionary-based perfect stemmer that can be parametrized for different accuracy and coverage levels.", "method": "The study conducts an empirical evaluation of the stemming accuracy for three stemming algorithms, including the widely used Porter algorithm, and discusses their use in IR. The researchers measure changes in IR performance corresponding to changes in stemming performance on a given dataset.", "outcome": "The authors present experimental evidence of their stemming algorithms' relative coverage and accuracy, and findings from the usage of these stemmers for IR.\n", "future_impact": "The study's results could open the door to future research on improving stemming and IR algorithms, and their interconnection, especially in specific contexts where stemming may be beneficial.", "venue": "SIGIR", "year": 2000, "title": "Stemming and its effects on TFIDF ranking."} +{"pid": "e95cd93b-2433-4fea-a881-a8a5d691028e", "context": "Medical researchers heavily deliberate over the final query when conducting systematic reviews, as choosing the best query among many variations is difficult. This has led to a new type of query performance prediction (QPP) task, where the aim is to predict the performance of a set of query variations for a particular topic.", "key_idea": "The study investigates the effectiveness of different QPP methods for the new task of Query Variation Performance Prediction (QVPP). Query variations involve reductions, expansions, and modifications of a seed query under the hypothesis that there exist variations that could improve retrieval effectiveness over the original query.", "method": "Using the CLEF 2017 TAR Collection, sixteen pre and post-retrieval predictors for QVPP are evaluated, primarily through correlations between prediction and actual performance.", "outcome": "The findings show that IDF-based QPPs exhibit the strongest correlations with performance. However, despite the existence of better-performing query variations, little improvement over the original query is seen when using QPPs to select the best query.", "future_impact": "These findings highlight the difficulty in identifying effective queries within this new context, thereby motivating further research to develop more accurate methods for assisting systematic review researchers in the query selection process.", "venue": "SIGIR", "year": 2018, "title": "Query Variation Performance Prediction for Systematic Reviews"} +{"pid": "6091398491e01105f877fcbb", "context": "Recent work has proposed stochastic Plackett-Luce (PL) ranking models for optimizing relevance and fairness metrics. But, the computation of gradient to optimize ranking metrics via stochastic gradient descent is infeasible in practice because it requires iterating over all possible permutations of items, resulting in the use of approximating techniques.", "key_idea": "The paper introduces a novel algorithm, PL-Rank, to estimate the gradient of a PL ranking model with respect to both relevance and fairness metrics, leveraging the specific structure of Plackett-Luce models and ranking metrics.", "method": "The authors conduct an experimental analysis comparing PL-Rank with existing policy gradients in terms of sample-efficiency and computational cost.", "outcome": "The experimental analysis reveals that PL-Rank is more sample-efficient and computationally less costly than existing policy gradients, leading to faster convergence at higher performance.", "future_impact": "PL-Rank may enable industry applications to utilize PL models for more relevant and fair real-world ranking systems.", "venue": "SIGIR", "year": 2021, "title": "Computationally Efficient Optimization of Plackett-Luce Ranking Models for Relevance and Fairness"} +{"pid": "01b0caf2-c118-4ef1-b7ac-cb8f1ab6e1fe", "context": "Researchers in information filtering have mainly concentrated on either content-based methods or collaborative methods since a principled unifying framework is still lacking.", "key_idea": "The authors suggest that both content-based and collaborative filtering methods can be combined under a hierarchical Bayesian framework using a nonparametric common prior, a sample generated from a Dirichlet process that serves as a hyper prior.", "method": "The authors describe the steps to learn this nonparametric distribution and apply it to learn users' information needs. Existing information filtering techniques are interpreted within this framework.", "outcome": "The result is a simple and understandable algorithm that offers a principled way to combine content-based filtering and collaborative filtering. The authors demonstrate the empirical success of the proposed information filtering methods.", "future_impact": "This work could potentially influence future developments in information filtering by providing a unifying view of various existing techniques.", "venue": "SIGIR", "year": 2004, "title": "A nonparametric hierarchical bayesian framework for information filtering"} +{"pid": "775fd04f-d3db-4b5d-917e-ef05b9cc5716", "context": "The structure of a document is usually a tree, with ordered nodes. Unstructured documents, however, have a single-level ordered tree structure. There are not established algebras for handling unstructured documents.", "key_idea": "The authors present a many-sorted algebra model for handling unstructured documents, where documents are represented by relations. The model can be extended to handle documents of various types and combined with the relational algebra for a common algebra system.", "method": "The authors demonstrate the capabilities and versatility of their model by using it to address various document handling and information retrieval problems.", "outcome": "The authors show through examples that the proposed model can handle a wide range of document handling and information retrieval problems.", "future_impact": "The model proposed in this paper can be as a part of a general document model, potentially extending its usage to a broader context involving document handling and data management.", "venue": "SIGIR", "year": 1987, "title": "A relational model for unstructured documents"} +{"pid": "01be850d-1281-4de9-881b-2f706b2b97aa", "context": "The field of information retrieval (IR) is typically defined as concerned with retrieval of documents that satisfy an information need. However, these definitions fail to capture major threads of activity in IR research and often neglect the human element in the retrieval process.", "key_idea": "The author proposes a new definition for information retrieval centered on the purpose of IR, which is support of cognition, arguing that existing definitions are inaccurate and miss major components of IR research.", "method": "The author formulates newer definitions, taking into account different perspectives in IR, and illustrates a definition centered around the purpose of IR.", "outcome": "The author presents a new definition of the field of IR, which is centered on 'support of cognition', rather than just the retrieval of documents satisfying an information need.", "future_impact": "This revised definition could potentially enable a richer connection with other disciplines and provides a vision of what IR research can accomplish, shaping future research within this area.", "venue": "SIGIR", "year": 2018, "title": "What We Talk About When We Talk About Information Retrieval"} +{"pid": "6e752dfd-e161-4d35-9764-1c0c9b903e1f", "context": "Automatic image annotation is the task of automatically assigning words to an image to describe its content. This is typically achieved using machine learning approaches to model the relationship between words and images in an annotated set of images.", "key_idea": "This paper proposes the use of a hierarchy defined on the annotation words, derived from a text ontology, to improve automatic image annotation and retrieval. This hierarchy is used in generating a visual vocabulary for representing images and as a framework for a hierarchical classification approach for automatic image annotation.", "method": "The proposed approach exploits the hierarchy in creating a visual vocabulary, and its effectiveness is demonstrated by measuring improvements in the annotation performance of translation models.", "outcome": "The use of the proposed hierarchy in generating the visual vocabulary results in improved annotation performance of translation models.", "future_impact": "The hierarchical classification approach provides a basis for the potential construction of multimedia ontologies, which may lead to advancements in the field of automatic image annotation and retrieval.", "venue": "SIGIR", "year": 2005, "title": "Exploiting ontologies for automatic image annotation"} +{"pid": "b514675a-0d45-4358-913a-ee33b28aa5b2", "context": "The influence of domain expertise on how people search information on the web is a relevant aspect but hasn't been sufficiently investigated, particularly in relation to medical expertise.", "key_idea": "This study investigates how medical domain experts search the Web for information related to their expertise, compared with non-experts, considering differences in sites visited, query vocabulary, and search behavior.", "method": "A log-based study is conducted to observe and compare the search behavior, sites visited, and query vocabulary of medical domain experts and non-experts.", "outcome": "The study finds noticeable differences in the sites visited, query vocabulary, and search behavior between medical domain experts and non-experts.", "future_impact": "These findings have implications for the automatic identification of domain experts from interaction logs, and for utilizing domain knowledge in applications like query suggestion or page recommendation to assist non-experts.", "venue": "SIGIR", "year": 2008, "title": "How medical expertise influences web search interaction"} +{"pid": "60b9a4abe4510cd7c8fc2534", "context": "In the education area, especially examination paper analysis such as Paper Difficulty Prediction and Finding Similar Papers, the existing works mainly focus on representation learning of each test item and do not account for the hierarchical document structure inherent in examination papers.", "key_idea": "This paper introduces a novel Examination Organization Encoder (EOE), which recovers the hierarchical document structure of examination papers and uses it for robust representation learning.", "method": "The authors develop a syntax parser to interpret an examination paper into an Examination Organization Tree (EOT). They utilize a two-layer GRU-based module for the representation of each leaf node (or test item) and a subtree encoder module to aggregate this representation and compute an embedding for each EOT layer. All layer embeddings are then fed into an output module to obtain the final representation.", "outcome": "Results from experiments on real-world examination paper data demonstrate the increased effectiveness and interpretability of the proposed Examination Organization Encoder method.", "future_impact": "The new representation learning approach with hierarchical document structure awareness can be potentially useful for various downstream tasks in examination paper analysis.", "venue": "SIGIR", "year": 2021, "title": "Enhanced Representation Learning for Examination Papers with Hierarchical Document Structure"} +{"pid": "f14314ae-8259-41fe-b984-9fc4f70efa4e", "context": "Many techniques in information retrieval produce counts from a sample, and it is common to analyse these counts as proportions of the whole. Proportions carry only relative information and are not free to vary independently of one another. These characteristics mark compositional data.", "key_idea": "The paper introduces the application of Compositional Data Analysis (CoDA) to the field of information retrieval, specifically for analyzing compositional data.", "method": "The authors explore compositional data in information retrieval through the lens of distance measures and compare common naive measures to composition-aware measures.", "outcome": "The authors demonstrate that common measures naive to compositions have some undesirable properties which can be avoided with composition-aware measures and show that these measures improve clustering.", "future_impact": "The utilization of CoDA, previously unexplored in the field of information retrieval, can potentially change the way compositional data is analyzed in the field.", "venue": "SIGIR", "year": 2014, "title": "Compositional data analysis (CoDA) approaches to distance in information retrieval"} +{"pid": "5fac54ea-8159-416b-9475-5e72e9a88a39", "context": "Previously, searches requested of the Oncology Information Service at the University of Leeds were compared to the searching expertise of a MEDLARS indexer, demonstrating the need for a more efficient searching system for the cancer therapy literature on MEDLINE.", "key_idea": "The authors propose a rule-based approach, using a modified subset of Medical Subject Headings (MeSH) and stored on a microcomputer, for an end-user searching tool that exhibits detailed knowledge regarding on the subject matter to enhance retrieval performance.", "method": "The authors test their system using previously-requested searches from the Oncology Information Service at the University of Leeds. The search term and frame selection process uses a 'blackboard' philosophy and is programmed using PROLOG.", "outcome": "The implementation of the rule-based approach using PROLOG was found to be efficient in search term and frame selection, suggesting its superior retrieval performances in comparison to alternative approaches in end-user searching.", "future_impact": "The development of this rule-based approach signifies progress in creating an effective expert system for end-user searching of cancer therapy literature, potentially improving medical research efficiency.", "venue": "SIGIR", "year": 1983, "title": "End user touch searching for cancer therapy literature: a rule based approach"} +{"pid": "097b0711-9def-4de3-b11f-cf5467f74bcb", "context": "Existing evaluation methods for Personal Search may not fully capture the complexity of user- and context- variables and their impact on such aspects as query characteristics including length, field submitted to and use of named entities.", "key_idea": "This paper explores the idea of using findings from a lab-based user study of email re-finding behavior to seed simulated evaluations of various retrieval models for personal search, enhancing existing evaluation methods.", "method": "A lab-based user study (n=21) of email re-finding behavior was performed, examining how query characteristics change in different situations and developing logistic regression models on the query data to explore relationships.", "outcome": "Several interesting trends were revealed and used to seed a simulated evaluation which determined that different models are more effective in different situations.", "future_impact": "The study's findings may have implications both for the design of email search tools and for the manner in which algorithms for Personal Search are evaluated, suggesting a more context-sensitive approach to these tasks.", "venue": "SIGIR", "year": 2011, "title": "Seeding simulated queries with user-study data for personal search evaluation"} +{"pid": "5f0277e911dc830562231e79", "context": "In modern applications, rankings are vital and often determined by machine learning-based algorithms, which may exacerbate harmful biases contained in the data. Existing methodologies for fair ranking usually imply that the protected groups affected by the biases are known beforehand, which may not always be the case in realistic scenarios.", "key_idea": "The authors aim to design fair ranking algorithms without knowing in advance the groupings that will be used to assess their fairness. Their approach involves relying on a pre-selected set of groupings when creating the ranked lists.", "method": "The authors propose a two-step greedy brute-force method to embed their strategy and validate it using the TREC 2019 Fair Ranking track dataset and settings.", "outcome": "The paper focuses on the design of the fair ranking strategy and methodology, detailed outcomes are not presented in the abstract.", "future_impact": "The proposed approach to fair ranking could potentially mitigate the reproduction and exacerbation of harmful biases in machine learning-based ranking systems even when the protected groups aren't known in advance.", "venue": "SIGIR", "year": 2020, "title": "Multi-grouping Robust Fair Ranking"} +{"pid": "d19af3bb-802f-4908-8e08-f756ad0a6288", "context": "In scientific literature search, a large portion of queries are entity-set queries, which reflect users' need for finding documents that contain multiple entities and inter-entity relationships. However, these queries are usually sparse, which challenges existing search algorithms that model each entity separately and makes many supervised ranking models that rely heavily on associated click history ineffective.", "key_idea": "The authors introduce SetRank, an unsupervised ranking framework, designed to model inter-entity relationships and capture entity type information, coupled with a novel unsupervised model selection algorithm based on the technique of weighted rank aggregation.", "method": "The authors test and evaluate the proposed unsupervised approach using datasets from TREC Genomics Tracks and Semantic Scholar's query log.", "outcome": "The experiments demonstrate that SetRank significantly exceeds the performance of baseline unsupervised models, especially on entity-set queries, and the model selection algorithm effectively selects suitable parameter settings.", "future_impact": "The proposed model can potentially enhance scientific literature search, particularly for those inquiries involving entity-set queries, and opens up possibilities for queries with sparse data.", "venue": "SIGIR", "year": 2018, "title": "Entity Set Search of Scientific Literature: An Unsupervised Ranking Approach"} +{"pid": "32b384ec-cbe7-4dda-a2ff-b64f2c827623", "context": "Building test collections based on nuggets is a useful evaluation method for systems that return documents, answers, or summaries. However, nugget construction is time-consuming and does not scale well for large query sets.", "key_idea": "To enhance the efficiency and scalability of nugget-based evaluation, the authors explore the applicability of semi-automatic nugget extraction in the context of the NTCIR One Click Access (1CLICK) task.", "method": "The authors compare manually-extracted and semi-automatically-extracted Japanese nuggets to assess the coverage and efficiency of semi-automatic nugget extraction.", "outcome": "The findings show that the manual nugget extraction can be replaced with a direct adaptation of the English semi-automatic nugget extraction system, particularly for queries looking for broad answers from free-form text.", "future_impact": "The study promises a path toward efficient and scalable nugget-based evaluation, which could have broad implications for the development and testing of information retrieval systems.", "venue": "SIGIR", "year": 2013, "title": "Exploring semi-automatic nugget extraction for Japanese one click access evaluation"} +{"pid": "1a894153-4319-4055-912b-20ed8fd1f051", "context": "A growing amount of XML-documents are being stored in Peer-to-Peer (P2P) networks, and current P2P search engines use Information Retrieval (IR) techniques for content-based search. However, these techniques do not consider the structural features of documents and thus, retrieval quality in terms of recall and precision can suffer substantially in these distributed environments.", "key_idea": "The key idea is to use the document structure as an extra source of information to improve the retrieval quality of XML-documents in a P2P environment. A search engine is proposed that exploits structural hints expressed by the user or the self-describing structure of XML-documents and provides more focused results by offering ranked retrieval units.", "method": "The proposed search engine uses XML IR techniques to select peers participating in the retrieval process and compute the relevance of documents. The indexing approach includes both content and structural info, using fixed-sized posting lists. A re-ordering on the posting list is performed for a specific query, taking into account the structural similarity between key and query. The final ranking is computed in parallel on selected peers based on an extension of the vector space model.", "outcome": "A first prototype for P2P Information Retrieval of XML-documents named SPIRIX is under implementation. It is expected to improve retrieval quality of XML-documents in a P2P environment.", "future_impact": "Once evaluated and tested on the INEX Wikipedia Collection, the proposed approach in the thesis and the SPIRIX system could potentially change the way XML-documents are retrieved in P2P systems, improving result precision, recall, and specificity while minimizing bandwidth consumption and considering performance aspects.", "venue": "SIGIR", "year": 2008, "title": "Exploiting XML structure to improve information retrieval in peer-to-peer systems"} +{"pid": "49ebc1de-bc48-465b-b663-426496db6295", "context": "Query Auto-Completion (QAC) is a highly used feature in web search engines but there is little published research on how users interact with such services.", "key_idea": "This paper carries out the first large-scale study of user interactions with auto-completion based on query logs of Bing, a commercial search engine.", "method": "The authors use query logs from Bing, a commercial search engine, to analyze user interactions with auto-completion systems.", "outcome": "The results show that lower-ranked auto-completion suggestions receive substantially less engagement than those ranked higher. Users engage with auto-completion most likely after typing about half of the query, particularly at word boundaries. The likelihood of using auto-completion varies with the distance of query characters on the keyboard.", "future_impact": "The insights from this study are valuable for understanding user engagement with auto-completion, and are likely to inform the design of more effective QAC systems.", "venue": "SIGIR", "year": 2014, "title": "On user interactions with query auto-completion"} +{"pid": "02923c94-8ead-4fb6-af2e-4a788a785d47", "context": "The precision and recall of information retrieval systems is not very satisfactory in existing solutions.", "key_idea": "The authors propose the development of concept-oriented databases using AI knowledge representation schemes, with an example being the augmentation of a 238,000 citation database, Chemical Abstracts (CA) Volume 105, by adding detailed conceptual information in the form of frames and hierarchies.", "method": "The process involves parsing the initial text data using natural language processing (NLP) techniques to create frames describing the semantics of the index entries in the database and creating links into a large semantic network of conceptual objects.", "outcome": "The authors proposed the development of a database which will result in a knowledge base that can be examined using a simple hypertext system, with the conceptual information serving as pathways to connect related citations.", "future_impact": "The creation of such concept-oriented databases can contribute to improving the precision and recall of information retrieval systems.", "venue": "SIGIR", "year": 1988, "title": "Development of a large, concept-oriented database for information retrieval"} +{"pid": "00b7d041-4fdb-422a-8586-6577986c2d3f", "context": "Current search technology, while widely used for learning-oriented information needs, is optimized mainly for generic relevance, not effective learning outcomes. This leads to inefficient information trails that a user must follow while searching to achieve a learning goal, which could involve unnecessarily easy or difficult content, or material irrelevant to actual learning progress.", "key_idea": "The authors address this problem by introducing a theoretical framework, algorithms, and an information retrieval model that is optimized for learning outcomes instead of generic relevance. They incorporate a cognitive learning model into a retrieval objective to find the search results that represent the most efficient 'training set' for a human learner.", "method": "The authors present an algorithm for an efficient approximate solution and conduct a crowdsourced user study involving a vocabulary learning task, comparing the effectiveness and efficiency of their retrieval framework against the results from a commercial search engine (Google).", "outcome": "The authors demonstrate that the personalized results generated by their model were effective in improving word learning outcomes in their user study.", "future_impact": "The authors' work represents a new way of thinking about search algorithms, with the potential to significantly improve the efficiency and effectiveness of online learning.", "venue": "SIGIR", "year": 2017, "title": "Retrieval Algorithms Optimized for Human Learning"} +{"pid": "60b99c43e4510cd7c8ef064c", "context": "Learning the preferences of new users quickly and reliably is a key challenge in the design of recommender systems. Existing methods struggle to deal with intra-cluster variability between users, which can add noise to user ratings.", "key_idea": "This paper introduces cluster-based bandits, an online learning algorithm, that groups users into clusters based on the similarity of their preferences, and identifies the cluster a new user belongs to in order to quickly learn their preferences.", "method": "The authors propose the idea of distinguisher items that are highly informative for deciding the cluster a new user belongs to, even in the presence of rating noise. The cluster-based bandit algorithm is used to efficiently adapt to user responses.", "outcome": "The cluster-based bandit algorithm is able to rapidly learn the correct cluster to assign to a new user, despite the noise in rating.", "future_impact": "The proposed cluster-based bandits can speed up the process of understanding new user preferences in recommender systems and potentially improve recommendation accuracy.", "venue": "SIGIR", "year": 2021, "title": "Cluster-Based Bandits: Fast Cold-Start for Recommender System New Users"} +{"pid": "825b1da4-6b82-40f9-a2bd-abcf80da1088", "context": "Advancements in location-acquisition technologies have resulted in an abundance of spatial trajectories. However, these trajectories are often generated at inconsistent or low frequencies due to application characteristics or energy saving, creating uncertain trajectories (routes between two consecutive points of a single trajectory).", "key_idea": "The authors present the Route Inference framework based on Collective Knowledge (RICK), which aims to construct popular routes from uncertain trajectories. The framework can construct top-k routes that pass through specified locations within a given time span, by aggregating such uncertain trajectories in a mutual reinforcement way.", "method": "The method involves two components: the construction of a routable graph by collaborative learning among the uncertain trajectories, and a routing algorithm to construct top-k routes based on a user-specified query. The performance of the RICK framework is evaluated using two real datasets, consisting of Foursquare check-in datasets and taxi trajectories.", "outcome": "The results from the extensive experiments demonstrate that the RICK framework is both effective and efficient in inferring routes from uncertain trajectories.", "future_impact": "The paper suggests that the proposed RICK framework could be useful in applications such as trip planning, traffic management, and studies of animal movement.", "venue": "SIGKDD", "year": 2012, "title": "Constructing popular routes from uncertain trajectories"} +{"pid": "2d798ef2-1938-48bd-8419-13864dacf6df", "context": "Blogs have become a prominent social media on the Internet that enable users to share content, including personal thoughts and professional opinions, forming a fast-growing social network. However, summarizing and understanding the 'wisdom of crowds' from the complex temporal graphs formed by blog entries and hyperlinks presents a new domain for social media analysis.", "key_idea": "The study presents three graph analysis and mining technologies, namely, evolutionary graph clustering to discover blog communities, information flow ranking to identify influential bloggers, and temporal graph visualization to understand blogger dynamics.", "method": "Clustering techniques are applied to temporal blog data, with an emphasis on temporal smoothness in evolutionary spectral clustering. Two graph ranking algorithms leverage information flow to identify influential nodes. Visualization tools have been developed to understand the behaviours and dynamics of the bloggers.", "outcome": "The author's methods of incorporating temporal smoothness in evolutionary spectral clustering, identifying influential bloggers using information flow ranking, and utilizing temporal graph visualizations led to the development of tools that enable the summarization of the blogosphere.", "future_impact": "These methodologies for analysing and visualizing blog data can facilitate future researchers to better understand patterns and behaviours in social networks, promoting the exploration and understanding of the 'wisdom of crowds' in the blogosphere.", "venue": "SIGKDD", "year": 2007, "title": "Blog analysis and mining technologies to summarize the wisdom of crowds"} +{"pid": "1cbfab57-db03-454e-ad03-7c71fe420e19", "context": "Multi-agent models for simulating urban population mobility require significant input data to reliably match real-world behavior. Traditional methods for generating synthetic populations based on real-world observations use costly travel diary surveys with a focus on trip choice. This leads to unrealistic assumptions regarding activity chains or the independence of various trips during a day.", "key_idea": "The paper proposes utilizing 'motifs', which are individual daily movement patterns observed in Call Detail Records (CDRs), to generate more realistic synthetic populations for multi-agent models. Motifs appear stably in various cities, offering a more comprehensive view of daily activities.", "method": "The research investigates the relationship between these motifs and other mobility-related quantities (like travel distances and times, and mode choice). It also discusses transition probabilities for motifs and mode transitions.", "outcome": "The study found that while some characteristics appear unrelated to motifs, others, such as mode choice, show strong correlations. This can enhance the creation of synthetic populations for multi-agent models.", "future_impact": "The findings are seen as an advancement towards creating more realistic synthetic populations for multi-agent models, which could have implications for analyzing the performance of multimodal transportation systems or disease spread in urban areas.", "venue": "SIGKDD", "year": 2013, "title": "Daily travel behavior: lessons from a week-long survey for the extraction of human mobility motifs related information"} +{"pid": "6209c8265aee126c0f1e7f51", "context": "Twitter forms a heterogeneous information network (HIN) where various interactions among entities encode valuable information about them not fully captured by a single relation.", "key_idea": "The authors propose learning and using knowledge-graph embeddings of entities in the Twitter HIN (TwHIN) to capture and utilize richer information of Twitter social network entities.", "method": "The authors evaluate their approach through offline and online experiments on a variety of downstream tasks: personalized ads rankings, account follow-recommendation, offensive content detection, and search ranking.", "outcome": "They show that the learned embeddings yield significant offine and online improvement for various recommendation and classification tasks.", "future_impact": "The authors discuss the practical challenges in deploying industry-scale HIN embeddings, including compressing embeddings for reduced model latency and managing parameter drift across versions, suggesting future research directions.", "venue": "SIGKDD", "year": 2022, "title": "TwHIN: Embedding the Twitter Heterogeneous Information Network for Personalized Recommendation"} +{"pid": "c1002498-def9-4c85-ad85-d52dc35a01e0", "context": "Traditional itinerary planning involves users pre-determining their destinations, which can be inconvenient especially for strangers in unfamiliar environments. Existing solutions rely on the rank and classification of Points of Interest (POI) from location based social networks (LBSN) to deduce user intentions, and mining physical vehicle trajectories for actual traffic information.", "key_idea": "The authors propose a categorization-based itinerary recommendation framework that merges physical trajectory data with LBSNs, using a Voronoi graph-based GPS trajectory analysis method to build traffic information networks and implementing an ant colony algorithm for multi-object optimization to discover suitable itineraries.", "method": "The authors validate their concept through experiments on datasets from FourSquare and the GeoLife project, also carrying out a satisfaction test on the recommended items.", "outcome": "The proposed framework's recommendation satisfaction rate averaged 80% in trials.", "future_impact": "This paper suggests a new avenue for further research and development in itinerary planning and recommendation, specifically those that integrate physical transportation data with online social network data for the purpose of generating more precise, user-intentionally oriented suggestions.", "venue": "SIGKDD", "year": 2012, "title": "Intention oriented itinerary recommendation by bridging physical trajectories and online social networks"} +{"pid": "63391c60-2556-436c-bbd5-bdc7ae01e459", "context": "Developing ensemble models has been an area of active research in large-scale data mining. Model uncertainty has been addressed by bagging i.e ensemble of models, and plurality voting method has been commonly chosen due to its simplistic implementation despite its ad-hoc reconciliation inefficiencies.", "key_idea": "The paper proposes exploring various voting methods such as anti-plurality method, plurality method with elimination, Borda count method, and Condorcet's method of pairwise comparisons for model reconciliation in place of the commonly used plurality method.", "method": "The authors empirically compare the performance of these alternative voting methods when applied in bagging.", "outcome": "The empirical results suggest that plurality is not always the best voting method, the number of classes and the degree of data set noise can affect the performance of voting methods.", "future_impact": "While definitive conclusions about specific voting methods are withheld, the study's results open the door to the application of voting theory in classification theory, raising interesting research questions for future investigation.", "venue": "SIGKDD", "year": 2003, "title": "Empirical comparisons of various voting methods in bagging"} +{"pid": "aa20d541-4e75-444a-8944-10e61e988579", "context": "The implicit interconnections or explicit relationships among individual entities in a dataset, often modeled as a graph, hold the value proposition of that dataset. Visualization of such graphs can uncover insights that reveal this value, and sometimes, these relationships are represented as numerical attributes associated with nodes or edges on the graph.", "key_idea": "The authors propose a visualization method for exploring attributed graphs, focusing on those with numerical attributes. The strategy aims to uncover the relationship between attribute values and graph topology by transforming the network into a terrain map, which represents components of interest and their relationships, sensitive to the attributes over the graph.", "method": "The proposed method transforms the network to generate a terrain map that reveals the overall distribution of components-of-interest and the relationships among them while being sensitive to attribute values over the graph.", "outcome": "The visualization method successfully demonstrates efficacy across various real-world data science tasks, with the potential to scale to large graphs with millions of nodes.", "future_impact": "This visualization method can significantly improve analysis efficiency and reveal insights from large, numerically attributed graphs, propelling further research and improvements in this field.", "venue": "SIGKDD", "year": 2017, "title": "Visualizing Attributed Graphs via Terrain Metaphor"} +{"pid": "9302f892-3bd7-479e-a72b-8565f147c7cd", "context": "Online information services have grown too large for users to navigate without automated tools such as collaborative filtering. Many similarity measures have been proposed and individually evaluated, but they have not been evaluated relative to each other in a large real-world environment.", "key_idea": "The authors carry out an extensive empirical comparison of six distinct measures of similarity for recommending online communities to members of the Orkut social network.", "method": "The authors measure the usefulness of the different recommendations by actually measuring users' propensity to visit and join recommended communities, and examine how the ordering of recommendations influenced user selection.", "outcome": "The paper provides insights into how different similarity measures perform relative to each other in recommending communities, including how the order of recommendations influences user choices, within a real social network.", "future_impact": "The findings could be important for the development of recommendation systems in social networks, potentially influencing how communities and interaction features are designed and implemented.", "venue": "SIGKDD", "year": 2005, "title": "Evaluating similarity measures: a large-scale study in the orkut social network"} +{"pid": "af7962b4-b1c9-4594-a765-7245d9855e93", "context": "There has been no previous work on segmenting densely packed aggregates based on image classification.", "key_idea": "The authors propose a new methodology for delineating densely packed aggregate particles based on image classification which is novel in the field.", "method": "The proposed method has been tested experimentally on different kinds of densely packed aggregate images that are difficult for a normal edge detector to process.", "outcome": "The studied algorithm was found applicable to diverse use-cases beyond its original application in processing densely packed aggregate images.", "future_impact": "The authors anticipate that the proposed algorithm can be applied to a variety of other applications.", "venue": "SIGKDD", "year": 2007, "title": "Image classification and segmentation for densely packed aggregates"} +{"pid": "60c1657c91e0112cf43c1ed4", "context": "Several cross domain sequential recommendation models have leveraged information from a source domain to improve CTR predictions in a target domain, but they have not considered bidirectional latent relations of user preferences across source-target domain pairs, which prevents them from enhancing cross-domain CTR predictions for both domains simultaneously.", "key_idea": "This paper proposes a novel approach to cross-domain sequential recommendations based on the dual learning mechanism, called Dual Attentive Sequential Learning (DASL), that simultaneously transfers information between two related domains in an iterative manner till the learning process stabilizes.", "method": "The DASL model is examined through offline experiments on three real-world datasets and an online A/B test at a major video streaming platform Alibaba-Youku.", "outcome": "The DASL model consistently outperforms several state-of-the-art baselines across all experimental settings, and it significantly improves business performance over the latest production system in the company based on results from Alibaba-Youku.", "future_impact": "The novel approach proposed can help improve the cross-domain click-through-rate prediction in online commerce platforms that deal with a multiplicity of product domains.", "venue": "SIGKDD", "year": 2021, "title": "Dual Attentive Sequential Learning for Cross-Domain Click-Through Rate Prediction"} +{"pid": "aa91b2fe-9abd-4f79-963c-e30700d75cd8", "context": "Recent studies on social network evolution propose generative models which capture only the statistical properties related to node-to-node link formation, but they do not explain group formation.", "key_idea": "The authors propose a novel model capturing the co-evolution of social and affiliation networks, revealing insights into group formation that users often join groups for reasons other than their friends being there.", "method": "The model was applied to several real-world networks, in experiments that captured both newly observed and previously studied network properties.", "outcome": "The model successfully captured both the newly observed and previously studied network properties, providing insights into group formation in social and affiliation networks.", "future_impact": "The proposed model can facilitate controlled experiments to study the effect of actor's behavior on the evolution of affiliation networks and it allows the generation of realistic synthetic datasets.", "venue": "SIGKDD", "year": 2009, "title": "Co-evolution of social and affiliation networks"} +{"pid": "e587fc75-0bd3-417e-8069-45f53ad7a32a", "context": "Generating accurate models for network intrusion detection in data-flow environments, where the data is unstructured and time-dependent, is an unresolved issue. Applying models to real-time data brings additional complexity due to varying computational costs at different levels of data.", "key_idea": "The paper proposes an approach that creates features from different levels of data, each having its own computational costs. It leverages 'necessary conditions' associated with low-cost features to decide if certain high-cost features need computation and if the corresponding classification rules need checking.", "method": "The authors build network intrusion detection models and apply their tools to this task, testing them using the network data provided in the 1998 DARPA Intrusion Detection Evaluation program.", "outcome": "The authors do not provide specific outcomes from the application of their method to the task of network intrusion detection in the abstract.", "future_impact": "The authors additionally discuss the potential for using the models created through their approach in real-time network intrusion detection systems but do not specify potential future impacts in the abstract.", "venue": "SIGKDD", "year": 1999, "title": "Mining in a data-flow environment: experience in network intrusion detection"} +{"pid": "290fddef-2d68-4ea1-88fc-7aad5481072c", "context": "Mining large datasets recording human activities for individual data understanding is key for new knowledge-based personalized services. Transactional data is a pervasive type of information collected by many services involving large user bases. The problem is clustering individual transactional data for masses of users.", "key_idea": "The paper proposes txmeans, a parameter-free clustering algorithm designed for partitioning transactional data automatically and efficiently, especially when clustering must be applied to a massive number of different datasets such as individual transaction histories.", "method": "Txmeans is tested on both real and synthetic datasets, demonstrating its application in a personal cart assistant.", "outcome": "The experiment results show the practical effectiveness of txmeans for mass clustering of different personal datasets, outperforming existing methods in terms of quality and efficiency.", "future_impact": "The results suggest potential use cases of txmeans in applications like a personal cart assistant which could lead to further advancements in knowledge-based personalized services.", "venue": "SIGKDD", "year": 2017, "title": "Clustering Individual Transactional Data for Masses of Users"} +{"pid": "60d996c80abde95dc965f698", "context": "Existing large-scale media search systems lack modern multimedia search experience that blends text understanding, media understanding, and multimodal technologies.", "key_idea": "The authors present VisRel, a deployed large-scale media search system that leverages text understanding, media understanding, and multimodal technologies for improved multimedia search experiences.", "method": "The authors develop image and video understanding models for content retrieval and train efficient media-to-query relevance models, refining both online and offline metrics. They summarize learnings gathered from hundreds of A/B test experiments.", "outcome": "The use of VisRel led to a 34% absolute improvement in media-to-query relevance and a 10% improvement in user engagement.", "future_impact": "This work can provide practical solutions and insights for engineers who are interested in applying media understanding technologies to improve multimedia search systems, especially at large scales like Facebook.", "venue": "SIGKDD", "year": 2021, "title": "VisRel: Media Search at Scale"} +{"pid": "d3c84380-7beb-4e05-8f63-18dec7ff69a9", "context": "Understanding how people use public transport is crucial for the effective operation and future planning of the transport networks. Current methods for gaining insights into crowd behavior are not specified.", "key_idea": "The authors have developed and deployed a traffic measurement system, consisting of an efficient algorithm that detects individual subway trips from anonymized real-time data generated by a location-based system.", "method": "The system has been testing in the real-world for several months, reporting subway station crowdedness, passenger flows, and expected travel times at hourly intervals. The results of these measurements are validated by an independent market research company's manual passenger counts at a selected station on three distinct days.", "outcome": "The system revealed a strong correlation between its measurements and the independent company's manual counts, and found that travelers do not always opt for the route with the shortest travel time.", "future_impact": "The insights gained from the traffic measurement system have led to the development of a mobile app that allows users to plan their commute based on the average travel time between stations.", "venue": "SIGKDD", "year": 2015, "title": "Traffic Measurement and Route Recommendation System for Mass Rapid Transit (MRT)"} +{"pid": "63034f3590e50fcafd74feeb", "context": "Survival analysis, which predicts the risk of events such as death from cancer, has been shown to unintentionally bias against protected attributes like age, race, and gender. Current techniques often lead to biased survival predictions for different groups when censoring is assumed to be unrelated to prognosis and covariates, a situation typically violated in real datasets.", "key_idea": "The authors introduce newly defined fairness based on survival functions and censoring. They propose novel, fair, and interpretable survival models that utilise pseudo valued-based objective functions, with fairness definitions as constraints for predicting subject-specific survival probabilities.", "method": "The proposed models are tested on three real-world survival datasets. The study also investigates the interplay between interpretability and fairness and examines how fairness and censoring impact survival predictions for different protected attributes.", "outcome": "Experiments demonstrate that the proposed fair survival models significantly improve over existing survival techniques in terms of accuracy and fairness measures. The models provide fair predictions for protected attributes under different types and amounts of censoring.", "future_impact": "Further investigation into the interplay between interpretability, fairness, and how fairness and censoring impact survival predictions for different protected attributes may enhance the field of survival analysis and lead to more accurate and unbiased predictions.", "venue": "SIGKDD", "year": 2022, "title": "Fair and Interpretable Models for Survival Analysis"} +{"pid": "261cfe96-cb89-47eb-9388-b4d9c92e5b53", "context": "Existing studies aiming to understand the spreads of information flow within social media have limited understanding to fanning-out or narrow tree-like geometries, little is known about the geometry of these spreading patterns and how they are formed.", "key_idea": "In contrast to existing methodologies, this study explores 432 million information flow patterns extracted from a large scale online social media dataset and discovers three key ingredients that govern the formation of complex geometric patterns of information flow. The study proposes a new stochastic process model that incorporates these ingredients.", "method": "The authors use a data driven approach, analyzing a dataset of 432 million information flow patterns extracted from large-scale online social media to discern patterns and validate their proposed model.", "outcome": "The authors discovered a wide range of complex geometric patterns characterized by a three-dimensional metric space and demonstrated that their proposed stochastic process model can successfully reproduce these diverse geometries discovered from the empirical spreading patterns.", "future_impact": "The discoveries made in this study provide a theoretical foundation for the microscopic mechanisms of information flow, which could have wide implications for prediction, control, and policy decisions in social media.", "venue": "SIGKDD", "year": 2019, "title": "Uncovering Pattern Formation of Information Flow"} +{"pid": "60d996c80abde95dc965f5d0", "context": "Knowledge graph completion aims to predict missing relations between entities in a knowledge graph. Existing methods primarily rely on entity embedding.", "key_idea": "The authors propose a relational message passing method for knowledge graph completion. It's explicitly different because it only considers edge features (i.e., relation types) without entity IDs and aggregates this neighborhood information by passing relational messages among edges iteratively.", "method": "The authors model two kinds of neighborhood topology under the relational message passing framework: relational context and relational paths. They combine the two message passing modules together for relation prediction and validate this approach on knowledge graph benchmarks and their newly proposed dataset.", "outcome": "The proposed method, PathCon, outperforms state-of-the-art knowledge graph completion methods by a large margin. It is also applicable to inductive settings where entities are not seen in the training stage and provides interpretable explanations for the predicted results.", "future_impact": "The proposed method has the potential to provide more accurate and interpretable knowledge graph completions, even for graphs with entities unseen during training.", "venue": "SIGKDD", "year": 2021, "title": "Relational Message Passing for Knowledge Graph Completion"} +{"pid": "6bbc09b1-bfcd-4ab4-b047-8f88d5508910", "context": "Large datasets have typically been handled by throwing resources at the problem, using massive fault-tolerant storage architectures, and parallel and graphical computation models such as MapReduce, Pregel, and Giraph. However, this is not possible in all environments, and also not necessary for all queries, hence the interest in sampling as a more cost-effective and efficient solution.", "key_idea": "This paper discusses the use of sampling to create summary datasets that can rapidly answer queries and extend the functional lifespan of data in storage, balancing resource constraints, data characteristics, and required query accuracy.", "method": "The paper reviews progress in sample design for large datasets, including streaming and graph-structured data.", "outcome": "The tutorial provides an overview of state-of-the-art sampling methods far beyond simple uniform selection of elements to maximize the usefulness of the resulting sample.", "future_impact": "Applications of these sampling methods in various domains including network traffic and social networks are discussed. The proposed methods have the potential to change how large datasets are managed, particularly when exact returns are not required.", "venue": "SIGKDD", "year": 2014, "title": "Sampling for big data: a tutorial"} +{"pid": "7cead77c-c27e-4cfa-8f27-4f008be5fc40", "context": "In real-time bidding (RTB) for online display advertising, publishers traditionally set a reserve price for ad impressions. Estimating the failure rate, or probability that a reserve price fails to be outbid, is difficult due to most publishers' lack of knowledge regarding historical highest bidding prices offered by RTB advertisers.", "key_idea": "The authors aim to address the issue of estimating the failure rate of a reserve price in RTB. They propose a parametric survival model for prediction, which incorporates user and page interactions, as well as header bidding information.", "method": "To validate their solution, the authors model the prediction problem using a parametric survival model, considering user and page interaction factors, along with header bidding information.", "outcome": "The experimental results, as stated in the abstract, show the effectiveness of the proposed method for predicting the failure rate of reserve prices in RTB, although specific metrics aren't provided.", "future_impact": "The solutions proposed in this work are anticipated to have practical applications for publishers in the optimization of reserve prices to manage risks and improve expected revenue in RTB, a dominant technique in the online display advertising field.", "venue": "SIGKDD", "year": 2019, "title": "Reserve Price Failure Rate Prediction with Header Bidding in Display Advertising"} +{"pid": "a3343eb9-4dcc-4071-95b8-5bac9f4113ec", "context": "Various industries are turning to social media to identify key influencers on topics of interest, including the All England Lawn Tennis and Croquet Club (AELTC) which wants to analyze the social pulse around the Wimbledon Championships.", "key_idea": "The authors propose a multi-dimensional definition of influence that satisfies a set of functional and usability criteria defined in the study, stressing on both all-time influencers and recent influencers, and tracking of user influences over multiple time-scales.", "method": "The authors investigate an approach that uses an aspect hierarchy to annotate tweets with topics or aspects before analyzing them for influence.", "outcome": "The authors successfully define a number of criteria that social influence scores should satisfy, and described interesting insights and lessons they learned from the engagement.", "future_impact": "This work can influence how organizations analyze social influence and might push for aspect-specific influence analysis.", "venue": "SIGKDD", "year": 2015, "title": "Online Topic-based Social Influence Analysis for the Wimbledon Championships"} +{"pid": "fdcdac66-bf8c-43e6-b264-1be35b05f467", "context": "Crowd management at large events is a critical research area due to its high societal impact. It becomes especially challenging during large-scale events like earthquakes or festivals because human behavior deviates greatly from routine, complicating crowd dynamics prediction at a citywide level.", "key_idea": "The study proposes DeepUrbanEvent, an online system that uses a unique deep learning architecture built with recurrent neural networks to predict citywide crowd dynamics based on data from the current hour and predict the trend for the next hour.", "method": "The deep learning model is used to handle complex sequential data similar to video prediction tasks. The prototype system is tested with multiple large real-world events.", "outcome": "Experimental results show that DeepUrbanEvent outperforms existing approaches in crowd dynamics prediction. The system was successfully applied to multiple real-world events demonstrating its efficacy as an online crowd management tool.", "future_impact": "The DeepUrbanEvent system has potential to be deployed widely as an online crowd management system, demonstrating its practical applicability and probable societal impact.", "venue": "SIGKDD", "year": 2019, "title": "DeepUrbanEvent: A System for Predicting Citywide Crowd Dynamics at Big Events"} +{"pid": "658739ed-6f96-4b25-9903-f5e125c8f29c", "context": "The need for automated fact-checking in political discourses is becoming increasingly important due to the volume and speed of information distribution.", "key_idea": "ClaimBuster, a fact-checking platform, uses natural language processing and supervised learning to identify important factual claims in political discourses.", "method": "The claim spotting model of ClaimBuster is built using a human-labeled dataset of check-worthy factual claims from U.S. general election debate transcripts. A case study is presented where ClaimBuster is used to live cover the U.S. Presidential Election debates of 2016 and to monitor social media and Australian Hansard.", "outcome": "The paper provides details on the architecture, components, and evaluation of the model. However, specific evaluation results are not provided in the abstract.", "future_impact": "The authors outline the current state and future goals of ClaimBuster's continued improvement and expansion.", "venue": "SIGKDD", "year": 2017, "title": "Toward Automated Fact-Checking: Detecting Check-worthy Factual Claims by ClaimBuster"} +{"pid": "5f03f3b611dc8305622320cf", "context": "Approximately 100 million families across the world depend on cotton farming, which is especially susceptible to pest attacks. This leads to overuse of pesticides, lost income for farmers, and in some instances, farmer suicides.", "key_idea": "The authors propose a new solution for pesticide management in cotton farming, utilizing deep learning, smartphone cameras, inexpensive pest traps, existing digital pipelines, and agricultural extension-worker programs, specifically aimed at smallholder farmers in the developing world.", "method": "The study developed and deployed a solution for pesticide management in cotton farming using deep learning and various digital technologies, taking into account unique constraints such as data diversity, annotation challenges, issues with traditional evaluation metrics, computing on low-resource devices, and deployment through intermediaries.", "outcome": "The paper summarizes the key lessons learned from developing and deploying the proposed solution.", "future_impact": "The lessons from this study may assist other teams interested in developing AI solutions for global development.", "venue": "SIGKDD", "year": 2020, "title": "Pest management in cotton farms: an AI-system case study from the global South"} +{"pid": "d098935c-eb82-4fbe-a132-5d6877d294af", "context": "The Sylvester equation is a powerful tool for graph mining tasks. However, it has high computational complexity, requiring a complexity that is at least quadratic in the number of nodes of graphs, even with approximations.", "key_idea": "The authors propose a family of Krylov subspace-based algorithms (FASTEN) to speed up the computation of Sylvester equation for graph mining, by projecting the original equivalent linear system onto a Kronecker Krylov subspace.", "method": "The authors exploit the implicit representation of the solution matrix as well as the associated computation, and the decomposition of the original Sylvester equation into a set of inter-correlated Sylvester equations of smaller size.", "outcome": "Experimental evaluations on a diverse set of real networks demonstrate that the proposed algorithms are up to 10,000 times faster against the Conjugate Gradient method, the best known competitor that outputs the exact solution, and can scale up to million-node graphs.", "future_impact": "The proposed methods provide exact solutions without any approximation error and significantly reduce the time and space complexity for solving Sylvester equation, with potential implications for large-scale graph mining tasks.", "venue": "SIGKDD", "year": 2018, "title": "FASTEN: Fast Sylvester Equation Solver for Graph Mining"} +{"pid": "4872e4c7-4bed-4c11-8815-ec5acdce6e27", "context": "Lots of private information are collected and spread without proper protection due to privacy leak behavior in various malwares and suspicious applications, collectively referred to as privacy leak software (PLS).", "key_idea": "This paper presents a model called Privacy Petri Net (PPN) for analyzing privacy leaks in software, offering quantitative analysis indicators such as possibility, severity, crypticity, and manipulability.", "method": "The PPN model for privacy leaks is applied on real-world PLS for quantitative analysis of privacy leak behavior.", "outcome": "The case study demonstrates the model's ability to not only identify tested software as PLS, but also to calculate the severity, crypticity, and manipulability of each software, including those simply treated as benign by antivirus software.", "future_impact": "This method may lead to a more accurate and comprehensive identification and analysis of privacy leak behaviour in applications.", "venue": "SIGKDD", "year": 2012, "title": "Quantitative analysis for privacy leak software with privacy Petri net"} +{"pid": "f1c1bdd8-77db-4981-b988-2e93f71310ee", "context": "While itemset mining and graph mining have received attention in the field of data mining, most studies focus on either of the two. Only a few studies have addressed a combination of both, leaving a gap in the exploration of interconnected applications.", "key_idea": "The paper introduces a new problem known as 'itemset-sharing subgraph (ISS) set enumeration', which finds sets of subgraphs with common itemsets in a large graph, where every vertex has an associated itemset.", "method": "The authors propose an algorithm called ROBIN that enumerates connected subgraphs with common itemsets and finds their combinations. They conduct experiments using a synthetic network, a real biological network, and a citation network.", "outcome": "The proposed ROBIN algorithm proved to be efficient in processing networks with over one million edges. It was able to find biologically interesting patterns in a real biological network and identify successful collaborative research works in a citation network.", "future_impact": "The proposed problem and algorithm have potential applications in areas like side-effect analysis in drug discovery, and word-of-mouth influence analysis in marketing within social networks.", "venue": "SIGKDD", "year": 2010, "title": "Finding itemset-sharing patterns in a large itemset-associated graph"} +{"pid": "8dcb6579-2fe4-4dd5-a699-c1c0a5a5f303", "context": "Onion sites on the dark web use the Tor Hidden Service (HS) protocol to hide their locations on the internet and facilitate the hosting of malicious and illegal content. Identifying and monitoring such malicious sites is crucial to Computer Security and Law Enforcement communities, but it remains a challenging task due to the scale and complexity of the dark web.", "key_idea": "The paper presents an automated infrastructure that crawls and indexes content from onion sites into a data repository named the LIGHTS, and introduces an analysis service called Automated Tool for Onion Labeling (ATOL), which consists of a keyword discovery mechanism (ATOLKeyword), a classification framework (ATOLClassify), and a clustering framework (ATOLCluster) to conduct a thematic assessment of the content of onion sites.", "method": "ATOL was put to the test with onion datasets derived from the LIGHTS repository and its algorithms were benchmarked on the publicly available 20 Newsgroups dataset to further validate its capability. It uses information from multiple external heterogeneous knowledge sources, ranging from domain expertise to Bitcoin transaction data, to categorize onion content in the absence of sufficient supervised data.", "outcome": "On the LIGHTS dataset, ATOLClassify gives a 12% performance gain over an analyst-provided baseline, while ATOLCluster gives a 7% improvement over state-of-the-art semi-supervised clustering algorithms. ATOL has been successfully deployed as part of the LIGHTS system.", "future_impact": "The deployment indicates potential for adoption within the Computer Security and Law Enforcement communities for identification and monitoring of illicit sites on the dark web. The performance gains suggested can lead to further research and development in the field.", "venue": "SIGKDD", "year": 2017, "title": "Automated Categorization of Onion Sites for Analyzing the Darkweb Ecosystem"} +{"pid": "64af9a053fda6d7f065a6b50", "context": "Large pre-trained code generation models, such as OpenAI Codex, are making significant improvements in programmer productivity by generating syntax-and function-correct code.", "key_idea": "The authors introduce CodeGeeX, a multilingual model to generate code which is pre-trained on tokens of 23 programming languages.", "method": "The authors evaluated CodeGeeX based on two tasks (code generation and translation) in an extensive experimental set-up. They also developed the HumanEval-X benchmark to evaluate multilingual models and conducted a user study.", "outcome": "CodeGeeX outperforms multilingual code models of similar scale in both tasks of code generation and translation on HumanEval-X. The user study showed that CodeGeeX increased coding efficiency for 83.4% of its users. Further, CodeGeeX-based extensions generated 8 billion tokens for tens of thousands of active users per week.", "future_impact": "As CodeGeeX is publicly accessible and its code, model weights, API, extensions, and HumanEval-X are open-sourced, the community is well-equipped to build upon and further improve it.", "venue": "SIGKDD", "year": 2023, "title": "CodeGeeX: A Pre-Trained Model for Code Generation with Multilingual Benchmarking on HumanEval-X."} +{"pid": "4c073236-9b4d-4a41-908b-6a2bd6dfad66", "context": "Sparse vectors are commonly used but handling them efficiently, especially for l1 programming problems, presents challenges due to the costs associated with memory and speed, typically making preprocessing necessary.", "key_idea": "The authors present a novel data structure for sparse vectors using Cuckoo hashing which is memory-efficient and provides nearly the same random access rates as dense vector operations.", "method": "The authors implement their Cuckoo hashing-based data structure and apply it to exact l1 programming problem-solving.", "outcome": "The proposed data structure enables solving of sparse l1 programming problems without preprocessing, at a cost identical to dense linear algebra in terms of both memory and speed.", "future_impact": "The proposed data structure provides a feasible alternative to the hash kernel, particularly beneficial where exact solutions are required, like feature selection.", "venue": "SIGKDD", "year": 2015, "title": "Cuckoo Linear Algebra"} +{"pid": "8073ead0-1d13-45fa-afa6-e48e3cb41415", "context": "The emergence of the Internet as a communication channel and the maturation of data management technologies have opened up opportunities for improving customer relationship management (CRM) efforts.", "key_idea": "The authors propose the utilization of data analytics to enhance various CRM functions such as customer segmentation, communication targeting, retention, and loyalty.", "method": "The authors provide a detailed description of the key technologies needed to implement analytical CRM, highlighting the organizational issues that must be carefully managed to realize effective CRM.", "outcome": "The paper successfully illustrates issues with current CRM efforts and demonstrates how data analytics techniques can address these problems.", "future_impact": "The authors express their hope to attract the data mining community to consider this important application domain, which can lead to future research and improvements in CRM.", "venue": "SIGKDD", "year": 2002, "title": "A Case for Analytical Customer Relationship Management"} +{"pid": "586ec96f-43a1-4e92-9b3a-8b188445b0e9", "context": "Efficient discovery of time series motifs, repeated subsequences, in static offline databases has seen significant research effort in recent years in light of their utility in higher-level data mining algorithms. Nonetheless, many domains require online discovery and maintenance of these motifs because time series data is inherently streaming.", "key_idea": "The authors developed the first algorithm that exactly discovers and maintains motifs in real time over the most recent history of a stream.", "method": "The proposed algorithm, which is extendible to maintain more complex pattern structures, is pitted against current offline algorithms that either require significant time to update or involve very costly pre-processing steps.", "outcome": "The new algorithm showed a worst-case update time that is linear to the window size and provided useful extensions to deal with arbitrary data rates and discover multidimensional motifs.", "future_impact": "The authors present potential applications of their algorithm by demonstrating its utility in several domains such as robotics, acoustic monitoring, and online compression.", "venue": "SIGKDD", "year": 2010, "title": "Online discovery and maintenance of time series motifs"} +{"pid": "85fdb861-4e89-4cfd-a334-fe4e5fbb7009", "context": "Current state-of-the-art correlation clustering approaches are sensitive to the initial set of seeds chosen and do not yield the optimal result in the presence of noise, which is a challenge when finding correlation clusters in the arbitrary subspaces of high-dimensional data.", "key_idea": "The authors propose RObust SEedless Correlation Clustering (ROSECC), an algorithm which does not require the selection of the initial set of seeds, and incrementally partitions the data, applying PCA to each partition independently.", "method": "They validate their approach on both synthetic and real-world datasets, and test the robustness of their method in the presence of significant noise levels in the data.", "outcome": "Experimental results demonstrated that the proposed method was effective and showed robustness in the presence of significant noise levels in the data.", "future_impact": "ROSECC \u2014 having no need for initial seed selection, an incremental partitioning approach, and automatic determination of cluster dimensionality \u2014 could address limitations of current correlation clustering approaches in handling high-dimensional data and noise.", "venue": "SIGKDD", "year": 2010, "title": "A robust seedless algorithm for correlation clustering"} +{"pid": "024a9ba5-7809-41b8-9e50-44790a808c1c", "context": "The Great East Japan Earthquake and the Fukushima nuclear accident resulted in large human population movements and evacuations. Understanding and predicting these movements have been a challenge for planning effective humanitarian relief, disaster management, and long-term societal reconstruction.", "key_idea": "The authors propose constructing a large human mobility database using GPS records from around 1.6 million people's mobile devices in Japan, and develop a probabilistic model that can be effectively trained by the discovered evacuations via machine learning techniques.", "method": "The authors used a large dataset of GPS records from 1.6 million people throughout Japan and mined this data to understand short-term and long-term evacuation behaviors during the disaster. They then used this data to train their probabilistic model.", "outcome": "With the proposed probabilistic model, population mobility in various cities impacted by the disasters throughout Japan could be automatically simulated or predicted. Also, new features or population mobility patterns were discovered after the recent severe earthquake, tsunami, and release of radioactivity in Japan.", "future_impact": "The discovered new features and population mobility patterns are likely to play a vital role in future disaster relief and management worldwide.", "venue": "SIGKDD", "year": 2013, "title": "Modeling and probabilistic reasoning of population evacuation during large-scale disaster"} +{"pid": "5f4f6e669fced0a24bec82ae", "context": "Graph represents a natural way of encoding data sample features and their relationships among them. Many techniques have been proposed in the past. However, recent years have seen the rapid development of data mining and knowledge discovery leading to novel graph analytics algorithms, which have been applied in various areas, including healthcare.", "key_idea": "This tutorial proposes to summarize the contemporary graph analytics algorithms and their application in healthcare. It covers deep network embedding techniques, graph neural networks, knowledge graph construction and inference, graph generative models and graph neural ordinary differential equation models.", "method": "The authors introduce and discuss various technical advances in graph analytics and how these methods can be applied in predictive modeling of clinical risks, disease subtyping, knowledge discovery from biomedical literature, and pharmaceutical research and development.", "outcome": "The tutorial summarizes existing research and methodologies in advanced graph analytics algorithms and their application in healthcare, including their use for understanding the mechanism, transmission, treatment, and prevention of COVID-19.", "future_impact": "The tutorial concludes with a discussion of potential issues and challenges such as interpretability, fairness, and security in utilizing graph analytics in healthcare. It also identifies available resources and potential opportunities for future research, indicating the continuing scope of graph analytics in addressing healthcare challenges.", "venue": "SIGKDD", "year": 2020, "title": "Recent Advances on Graph Analytics and Its Applications in Healthcare"} +{"pid": "6303517b90e50fcafd7875e3", "context": "DNA-stabilized silver nanoclusters (AgN-DNAs) show promise as biosensors and fluorophores due to their size, compatibility with DNA, and bright fluorescence, but researchers face challenges in understanding and designing these due to complex DNA-silver interactions and limitations on DNA sequences that produce high fluorescence.", "key_idea": "The authors propose using variational autoencoders (VAEs) as generative models in the design of AgN-DNAs, with the latent space of an LSTM-based \u03b2-VAE regularized to correlate with AgN-DNA properties such as color and brightness.", "method": "The authors use their model to design AgN-DNAs in the near-infrared (NIR) band, a range where few AgN-DNAs have been observed to date. The results of these designs are validated through wet lab experiments.", "outcome": "Through wet lab experiments, the model significantly shifts the distribution of AgN-DNA colors towards the NIR while also achieving bright fluorescence.", "future_impact": "This methodology of using VAE-based generative models has potential to advance the prospective applications of these nanomaterials for bioimaging, biosensing, and other technologies.", "venue": "SIGKDD", "year": 2022, "title": "DNA-Stabilized Silver Nanocluster Design via Regularized Variational Autoencoders"} +{"pid": "d4ac271c-2f0d-49ae-b0ff-39afcc1fb990", "context": "Power quality (PQ) monitors installed on the electricity network collect large amounts of event data, which can be difficult to manage and classify.", "key_idea": "The authors propose a k-means-based clustering method to handle the large size of power quality event data and help in classifying the events.", "method": "Power quality monitors, developed through the National PQ Project, detect PQ events such as voltage sags, swells, and interruptions as per the IEC Standard 61000-4-30, generating the raw data for method application.", "outcome": "The proposed method manages event data effectively, providing PQ assessments for specific measurement points and comparisons of varying points in terms of electricity network quality.", "future_impact": "This method can benefit future management of power quality events, assisting in generating assessments and comparisons of power quality at various points in an electrical network.", "venue": "SIGKDD", "year": 2009, "title": "Clustering of power quality event data collected via monitoring systems installed on the electricity network"} +{"pid": "5c090b94-04d6-41f6-a652-4075f550cab5", "context": "Association Rule Mining is a technique originally proposed for market basket data, and has potential applications in diverse fields, including Remote Sensed Imagery (RSI) data. However, the large size of the image data usually makes it challenging to be mined using existing algorithms in a reasonable amount of time.", "key_idea": "The authors propose an efficient association rule mining method for RSI data using Peano Count Tree (P-tree) structure, a lossless and compressed data representation structure that they had previously proposed.", "method": "The authors create an algorithm, P-ARM, based on P-trees for fast support calculation and provides significant pruning techniques. Experiments to test the efficiency of P-ARM are conducted by comparing its performance with FP-growth and Apriori algorithms.", "outcome": "The experiments reveal that the P-ARM algorithm is more superior for association rule mining on RSI spatial data compared to FP-growth and Apriori algorithms.", "future_impact": "The proposed approach has the potential to improve the mining of data in areas like precision agriculture, community planning, resource discovery and other areas where RSI is relevant.", "venue": "SIGKDD", "year": 2002, "title": "Association Rule Mining on Remotely Sensed Images Using P-trees"} +{"pid": "96a04ce4-3632-40b8-98f8-341baa42b529", "context": "Studies on Electronic Health Records (EHR) have enabled a broad range of applications in healthcare. However, each study is often independently commissioned, and data is obtained through long and often complex processes. This is followed by a cycle of analysis, model building, and insight generation which can take between 1 - 3 years.", "key_idea": "The authors propose an end-to-end machine learning SaaS system for analyzing a massive EHR dataset, comprising approximately 55 million patient lives and over 20 billion data points in the US. The system uses a ridge regularized Survival Support Vector Machine (SSVM) with a clinical kernel and Chi-square distance-based feature selection to discover relevant risk factors, leveraging weak correlations in EHR data.", "method": "The proposed system was proven in multiple real use cases, where it identified relevant factors effectively without expert supervision.", "outcome": "The machine learning methods used by the system are interpretable, key for acceptance in the target user base. The studies were completed within a time frame of 3-4 weeks, significantly faster than the industry standard of 12-36 months, indicating the system's robustness and efficiency.", "future_impact": "The proposed system may lead to accelerated analysis and discovery, better return on investment due to reduced investments, and quicker turnaround of studies.", "venue": "SIGKDD", "year": 2019, "title": "A Robust Framework for Accelerated Outcome-driven Risk Factor Identification from EHR"} +{"pid": "52c0b141-be60-414d-b9a6-88b941a8789e", "context": "Optimization tools are essential in data analysis and learning, and the rich collection of problems in these fields is providing new perspectives on optimization algorithms.", "key_idea": "The authors discuss the importance of optimization algorithms in several areas such as signal reconstruction, manifold learning, and regression/classification, with a particular focus on applications of asynchronous parallel algorithms for optimization and linear algebra in data analysis and learning.", "method": "The authors are discussing and analyzing recent research where optimization algorithms have been developed and applied successfully in various fields.", "outcome": "The authors showcase the successful application of optimization and asynchronous parallel algorithms in signal reconstruction, manifold learning, and regression/classification.", "future_impact": "This study could potentially drive new fundamental research in optimization algorithms for data analysis and learning.", "venue": "SIGKDD", "year": 2013, "title": "Optimization in learning and data analysis"} +{"pid": "65ace326-f1cf-4cef-944e-69b7efd89228", "context": "Dimension attributes in data warehouses are typically hierarchical, and OLAP tools are used to summarize measure attributes along a dimension hierarchy and to characterize changes in a hierarchical summary over time.", "key_idea": "The paper proposes a model of parsimonious explanation for changes in hierarchical summaries. This model uses a composition of node weights along the root-to-leaf paths in a dimension hierarchy, allowing for maximal generalization.", "method": "The authors demonstrate optimal parsimonious explanations on arbitrary rooted one dimensional tree hierarchies. They also ran experiments using real data sets, comparing the parsimony of their model against alternatives.", "outcome": "The study demonstrated that optimal parsimonious explanations can be computed efficiently in time essentially proportional to the number of leaves and the depth of the hierarchy. The method outperformed alternatives in parsimony and proved robust across different data sets.", "future_impact": "The proposed method's utility and its ability to produce parsimonious explanations from the output of any statistical model that provides predictions and confidence intervals, would make it widely applicable in future studies.", "venue": "SIGKDD", "year": 2007, "title": "Efficient and effective explanation of change in hierarchical summaries"} +{"pid": "946ad9fc-7436-4cc3-a2a6-6e6b12e51c30", "context": "With the rapid proliferation of video cameras in public places, the ability to identify and track people and other objects creates tremendous opportunities for business and security applications.", "key_idea": "This paper presents the Multiple Camera Indoor Surveillance project, which uses multiple cameras, agent-based technology and knowledge-based techniques to identify, track people, and summarize their activities, and a people localization system to identify and localize people in an indoor environment using low-level color features and a Bayesian decision-making approach.", "method": "The authors conduct a pilot experiment using 32 hours of data and augment the system with domain knowledge such as the location of working places in cubicles, doors, and passages.", "outcome": "The pilot experiment showed average recall and precision values of 68 and 59% respectively. Augmenting the system with domain knowledge increased the average recall to 87% and precision to 73%.", "future_impact": "The presented approach to people localization and indoor surveillance leverages multiple cameras and knowledge-based techniques, and has potential to enhance business and security applications.", "venue": "SIGKDD", "year": 2006, "title": "Multiple-camera people localization in an indoor environment"} +{"pid": "ce76e652-1a8c-4f19-b3b1-26636568b7b4", "context": "Coresets are compact representations of data sets such that models trained on a coreset are provably competitive with models trained on the full data set. They have been successfully used to scale up clustering models to massive data sets. However, existing approaches generally only allow for multiplicative approximation errors.", "key_idea": "The authors propose a novel notion of lightweight coresets that allow for both multiplicative and additive errors. They also provide a single algorithm to construct lightweight coresets for k-means clustering as well as soft and hard Bregman clustering.", "method": "The proposed algorithm is evaluated for scalability and efficiency in constructing lightweight coresets, and is tested on its applications to k-means clustering, soft and hard Bregman clustering, statistical k-means clustering, and empirical risk minimization.", "outcome": "The proposed algorithm is substantially faster than existing coreset constructions, embarrassingly parallel, and results in smaller coresets. In extensive experiments, the proposed algorithm outperforms existing data summarization strategies in practice.", "future_impact": "The proposed approach to lightweight coresets can potentially be utilized to compute smaller summaries for empirical risk minimization, extending its utility and impact in the field of data clustering and large-scale data analyses.", "venue": "SIGKDD", "year": 2018, "title": "Scalable k -Means Clustering via Lightweight Coresets"} +{"pid": "d4bf7b5a-5e76-4414-8d3e-642b33d149c7", "context": "Network embedding (NE) aims to learn low-dimensional representations for nodes in networks. Previous NE methods often separate content and structure representations of nodes, which requires a post-processing combination step. Existing methods also generally consider only short and fixed neighborhood scopes, which poses issues when dealing with complex networks.", "key_idea": "The paper proposes a sequence-to-sequence model-based NE framework called Self-Translation Network Embedding (STNE). This model translates each sequence from the content sequence to the node sequence using random walks on a network. This allows for seamless fusion of content and structure information and flexible learning of high-order proximity.", "method": "STNE uses a bi-directional LSTM encoder to fuse content and structure information from raw input. Experiments are performed on three real-world datasets to validate its performance in comparison to existing NE approaches.", "outcome": "The STNE model outperforms state-of-the-art NE approaches in categorizing nodes based on the experimental results.", "future_impact": "The proposed STNE can offer improved performance in network analysis problems. The methodologies and the dataset used in this study have been made publicly available for reproduction and further study.", "venue": "SIGKDD", "year": 2018, "title": "Content to Node: Self-Translation Network Embedding"} +{"pid": "1c62ffb3-2973-4762-afe3-e203ff4d08e6", "context": "Many machine-learning techniques, both supervised and unsupervised, assume data in an attribute-value form. However, this form is insufficient for many applications, leading to a shift towards first-order learning systems. These systems have high computational complexity. Most current databases are designed according to an entity-relationship formalism and implemented on a relational database management system, which is less complex but more expressive.", "key_idea": "This study introduces a clustering system designed to discover useful groupings in structured databases. This system is based on the COBWEB algorithm, extended with the ability to cluster structured objects.", "method": "The authors incorporate the entity-relationship formalism into the COBWEB algorithm, thereby enabling the algorithm to cluster structured data.", "outcome": "The paper presents a clustering system capable of discovering useful groupings in structured databases, extending the capabilities of the traditional COBWEB algorithm.", "future_impact": "This work potentially allows more expressive clustering in machine learning, offering a practical approach to conceptual clustering in structured databases.", "venue": "SIGKDD", "year": 1995, "title": "Conceptual clustering in structured databases: a practical approach"} +{"pid": "5f4f6e669fced0a24bec82f0", "context": "The abundance of text data from various sources poses a challenge for data mining to extract actionable structures and knowledge without requiring extensive and corpus-specific human annotations. The goal is to satisfy diverse applications and needs for comprehending and making good use of large-scale corpora.", "key_idea": "The authors introduce recent advances in text embeddings and their use in various text mining tasks for multi-dimensional analysis of massive text corpora. Specifically, the authors introduce weakly-supervised, domain-independent, language-agnostic and scalable text mining techniques that use text embeddings for discovering structured knowledge from texts.", "method": "The authors present their methods for text mining, including discussing state-of-the-art context-free embeddings and pre-trained language models, presenting several embedding-driven text mining techniques and demonstrating how these techniques can be used in text classification, retrieval, and summarization on real-world datasets, including political news articles and scientific publications related to the coronavirus.", "outcome": "Text embedding methods and their applications to a wide range of text mining tasks that facilitate multi-dimensional analysis of massive text corpora have been demonstrated. The results show that the topics and taxonomies discovered could form a multi-dimensional TextCube structure, which enhances text exploration and analysis for various important applications.", "future_impact": "The advancements and techniques presented in this paper could have broad applications in text classification, retrieval, and summarization, and help facilitate more efficient comprehension and utilization of massive text corpora.", "venue": "SIGKDD", "year": 2020, "title": "Embedding-Driven Multi-Dimensional Topic Mining and Text Analysis"} +{"pid": "1149bcae-b091-43af-895c-27a0d08f4d2e", "context": "Location based services are gaining importance to provide value-added services in various businesses due to the increased adoption of digital inclusion.", "key_idea": "The authors propose a computer vision-based system that tracks customer locations in shopping malls by recognizing individual shopping carts using an efficient two-stage approach - cart detection and recognition.", "method": "A binary pattern is placed between two pre-defined color markers and attached to each cart. The system processes live video feeds from cameras mounted on the mall aisles in real-time, using color segmentation, feature extraction and classification for cart detection, and spatial image processing techniques to decode the cart identification number in the recognition stage.", "outcome": "The abstract doesn't include any factual statements about actual study outcomes or results obtained with this approach.", "future_impact": "The abstract doesn't explicitly anticipate potential future impacts, developments or improvements based on this work.", "venue": "SIGKDD", "year": 2011, "title": "Video analytics solution for tracking customer locations in retail shopping malls"} +{"pid": "60c31f669e795e9243fd1670", "context": "Heterogeneous graph neural networks (HGNNs) have been widely developed in recent years, but the unique data processing and evaluation setups used by each work obstruct a full understanding of their advancements.", "key_idea": "The authors present a reproduction of 12 recent HGNNs and stress the importance of appropriate settings by demonstrating that simple homogeneous GNNs such as GCN and GAT are largely underestimated and can outperform all existing HGNNs with proper inputs. They also propose a new strong baseline, Simple-HGN.", "method": "The authors use the official codes, datasets, settings, and hyperparameters of 12 recent HGNNs for systematic reproduction, and construct a new Heterogeneous Graph Benchmark (HGB) consisting of 11 diverse datasets with three tasks.", "outcome": "The authors find that with proper settings, GAT potentially matches or outperforms all existing HGNNs across various scenarios. Their proposed baseline, Simple-HGN, significantly outperforms all previous models on HGB.", "future_impact": "Their work in refining the approach to HGNNs and the introduction of HGB is anticipated to standardize the process of heterogeneous graph data splits, feature processing, and performance evaluation, potentially accelerating the advancement of HGNNs research in future.", "venue": "SIGKDD", "year": 2021, "title": "Are we really making much progress?: Revisiting, benchmarking and refining heterogeneous graph neural networks"} +{"pid": "6245fe54-d1fe-46fb-9a11-fd64b5fc29d4", "context": "The concept of understanding a graph using a dense subgraph has been formalized in numerous ways that balance between approximating actual interestingness and computational efficiency. The issue is that while the computational cost of an algorithm is clearly defined, the interestingness of a pattern is fundamentally subjective, and usually treated informally or neglected. Most often, some form of density is used as a stand-in.", "key_idea": "This study introduces a framework for formalizing the subjective interestingness of a dense subgraph pattern, which depends on a user's prior beliefs about the graph. Specifically, two cases are considered: one where the user only has a belief about the overall density of the graph, and another where the user has prior beliefs about the degrees of the vertices.", "method": "The researchers propose and implement exact and approximate algorithms for mining the most interesting dense subgraph according to the new measure. The efficiency and validity of the proposed method are assessed by comparative empirical evaluation.", "outcome": "The empirical evaluation highlights the unique properties of this new interestingness measure in regards to different prior belief sets, demonstrating that the method can find interesting subgraphs that other methods are unable to locate.", "future_impact": "Since the proposed approach is conducive to iterative dense subgraph discovery and allows for overlapping of subsequently found patterns, it could reshape current practices and inspire novel research directions in dense subgraph discovery.", "venue": "SIGKDD", "year": 2016, "title": "Subjective interestingness of subgraph patterns"} +{"pid": "80ea14a1-88a2-45b9-8817-0c6ceecba00f", "context": "Supervised learning has been traditionally focused on error rate or model storage as performance metrics, but the dynamic nature of some realms, where class distributions change over time, necessitates more suitable measures.", "key_idea": "The authors introduce the concept of 'usefulness', arguing it can enhance classification rule performance and cut storage requirements, while also being applicable in dynamic settings where class distribution changes with time.", "method": "The concept of usefulness is incorporated into three algorithms: a dynamic nearest neighbour classifier, a dynamic version of Learning Vector Quantization, and a Radial Basis Function network. All these methods take into account the age of observations as well as the record of usefulness. These are applied to credit industry data sets.", "outcome": "The exact outcomes are not explicitly stated in the abstract.", "future_impact": "The use of usefulness as a measure could have implications in any data-driven industry, particularly those dealing with real-time, dynamically changing data sets, such as financial or trading industries.", "venue": "SIGKDD", "year": 1998, "title": "Evaluating usefulness for dynamic classification"} +{"pid": "de6d4b85-548a-456d-a571-869e21569723", "context": "The information stored in newsgroups is not currently organized based on Usenet's community structure, and there is a lack of detailed data on Usenet cross-postings.", "key_idea": "The authors propose to organize and retrieve newsgroup information based on the community structure of Usenet, focusing specifically on the network formed by cross-posts.", "method": "The authors collected and analyzed detailed data on Usenet cross-postings, examined the network's characteristics, and used a spectral algorithm to cluster newsgroups based on the cross-post matrix.", "outcome": "The network analysis showed that Usenet's cross-posting network is a small-world network with significant clustering. The spectral algorithm provided a topical classification of newsgroups, revealing many significant relationships that semantic clustering methods would miss.", "future_impact": "The proposed way of organizing and retrieving information stored in newsgroups using community structure could potentially improve the effectiveness of information retrieval from such platforms.", "venue": "SIGKDD", "year": 2004, "title": "Exploring the community structure of newsgroups"} +{"pid": "10817337-902d-4d11-b571-a303d8a067ba", "context": "Sequential data is often partitioned in numerous ways, either inherently in the data or as a result of sequence segmentation algorithms, leading to numerous different segmentations of the same underlying data points.", "key_idea": "The authors propose creating an aggregate partition - a segmentation that agrees as much as possible with the input segmentations.", "method": "The authors define each partition as a set of continuous non-overlapping segments of the timeline and solve the aggregation problem using dynamic programming and propose faster greedy heuristics. They experiment with their algorithms using mobile-phone user behavior data and results of different segmentation algorithms on genomic sequences.", "outcome": "The authors show that the problem of creating an aggregate partition can be solved optimally in polynomial time using dynamic programming.", "future_impact": "Their approach to aggregate partitions can be potentially applied to various fields where sequential data analysis is crucial, such as mobile-phone users' behavior analysis and genomic sequences segmentation.", "venue": "SIGKDD", "year": 2006, "title": "Aggregating time partitions"} +{"pid": "dbbae737-4938-42a3-a92c-a380e405ff9b", "context": "Existing methods for web user profiling based on user navigation paths are generally manual or non-optimal.", "key_idea": "The authors propose an approach that automatically extracts web user profiling based on user navigation paths, using Markov models (MM) to represent behaviors, and overcoming the high-dimensionality and sparsity of Markov models by clustering web documents based on their content before applying Markov analysis.", "method": "The approach is built around three notions - applying probabilistic exploration using Markov models, clustering web documents to manage high-dimensionality and sparsity of Markov models, and clustering Markov models to extract their gravity centers.", "outcome": "The approach results in the prediction of future states to be visited in k steps and allows for navigation sessions monitoring based on content and traversed paths.", "future_impact": "The approach can be used for the exploitation of multimedia archives with a focus on the Copyright Deposit which preserves French WWW documents, and could potentially be used as an exploitation tool for any website.", "venue": "SIGKDD", "year": 2003, "title": "A Markovian approach for web user profiling and clustering"} +{"pid": "60c19bdc91e0112cf43c2120", "context": "In E-commerce, voucher redemption rate prediction is crucial and often achieved through User-item Click-Through-Rate (CTR) models. However, these models overlook the complex relations among users, items and vouchers, as well as users' historical behavior in a voucher collection activity.", "key_idea": "This paper introduces the Deep Multi-Behavior Graph Networks (DMBGN), a model that captures the complex structural user-voucher-item relationships and considers user behavior happening before and after voucher collection.", "method": "The authors developed a User-Behavior Voucher Graph (UVG) for user behavior representation, used Higher-order Graph Neural Networks for high-level representation extraction, and built an attention network on top of a sequence of UVGs. The effectiveness of the DMBGN model was then evaluated through extensive experiments on three large-scale production datasets.", "outcome": "The proposed DMBGN model showed 10% to 16% relative AUC improvement over Deep Neural Networks (DNN), and 2% to 4% AUC improvement over Deep Interest Network (DIN).", "future_impact": "The authors made the source code and a sample dataset publicly available, which should enable and facilitate future research in this area.", "venue": "SIGKDD", "year": 2021, "title": "DMBGN: Deep Multi-Behavior Graph Networks for Voucher Redemption Rate Prediction"} +{"pid": "edfd4f4b-9e12-4bb3-9167-d9569e6627d1", "context": "Although Euclidean distance between two nodes is commonly used in outlier detection, it only captures the physical distance between them, not their local neighborhood densities. Traditional outlier detection methods often fail to capture outlying clusters and are sensitive to noise.", "key_idea": "The authors present a method for outlier detection using 'commute distance' from a random walk on a graph, which captures both the distance between nodes and their local neighborhood densities. The commute distance is shown as the Euclidean distance in the space spanned by eigenvectors of the graph Laplacian matrix.", "method": "The authors propose a graph component sampling and an eigenspace approximation, combined with pruning technique to reduce the computation time of commute distance. The efficiency of the method is validated through analysis and experiments.", "outcome": "The proposed method effectively detects both global and local outliers and outlying clusters, demonstrating higher resistance to noise than local outlier detection methods. Moreover, the computational time for commute distance is reduced to O(nlogn) while preserving the outlier ranking.", "future_impact": "The presented method provides a robust approach for outlier detection, which is expected to be a valuable tool for efficient, accurate detection of both global and local outliers, as well as for identifying outlying clusters, in potentially noisy datasets.", "venue": "SIGKDD", "year": 2010, "title": "Robust outlier detection using commute time and eigenspace embedding"} +{"pid": "d1537e91-49e8-457f-95ac-ddb548ffc96e", "context": "Previous studies lacked a robust mechanism to identify emerging threats in time-dependent, unlabelled data sets collected from the Dark Web.", "key_idea": "This paper presents a new approach for unsupervised anomaly detection in data collected from the Dark Web using a dynamical systems approach and using finite-time Lyapunov exponents.", "method": "The paper uses finite-time Lyapunov exponents to characterize the time evolution of both the directed network structure and the distribution of text attributes in the forum messages.", "outcome": "The paper provides initial anomaly detection results using the proposed method but does not provide specific results or benchmarks.", "future_impact": "The conclusion of the paper discusses promising avenues for future research, indicating the potential impact of the presented method in enhancing current practices of anomaly detection.", "venue": "SIGKDD", "year": 2010, "title": "Anomaly detection in extremist web forums using a dynamical systems approach"} +{"pid": "629d70385aee126c0f302471", "context": "Although numerous outlier detection (OD) algorithms exist in the literature, a majority of them are designed for single-machine operation. This is a gap, as large, cloud-resident datasets necessitate the need for distributed OD techniques. However, this area is not only understudied, but there is also a lack of public-domain implementations for practical use.", "key_idea": "This paper introduces Sparx, a data-parallel OD algorithm suitable for shared-nothing infrastructures, specifically implemented in Apache Spark.", "method": "The authors carry out extensive experiments on three real-world datasets, containing several billions of points and millions of features, in order to test the scalability and effectiveness of Sparx and to compare it with existing open-source solutions.", "outcome": "The experiments demonstrated that existing open-source solutions failed to scale up to large numbers of points or high dimensionality. In contrast, Sparx demonstrated scalable and effective performance.", "future_impact": "By open-sourcing Sparx under the Apache license, the authors aim to facilitate practical use of OD on modern-scale datasets.", "venue": "SIGKDD", "year": 2022, "title": "Sparx: Distributed Outlier Detection at Scale"} +{"pid": "d83b3a88-5ca5-4d3b-a138-29e6319de81e", "context": "Controlling the quality of worker output in crowdsourcing systems is a critical aspect, typically consuming a substantial fraction of the time and money invested in such systems.", "key_idea": "The authors propose techniques to generate confidence intervals for worker error rate estimates which allows for a better evaluation of worker quality.", "method": "The authors apply their devised techniques on a range of real-world datasets, and further they use these techniques to evict poorly performing workers and provide confidence intervals on the accuracy of the answers.", "outcome": "The authors demonstrated that their techniques generate correct confidence intervals on a variety of real-world datasets.", "future_impact": "The proposed techniques can be used in the future to evict poorly performing workers from crowdsourcing tasks and to provide confidence intervals on the accuracy of answers, enhancing the overall evaluation process of worker quality in crowdsourcing systems.", "venue": "SIGKDD", "year": 2013, "title": "Evaluating the crowd with confidence"} +{"pid": "60d996c70abde95dc965f577", "context": "Graph Neural Networks (GNNs) have achieved great results in several applications but have significant shortcomings in capturing heterogeneous structures and attributes of an underlying graph. Many Heterogeneous GNN (HGNN) variants have been proposed and have achieved state-of-the-art results, but there is a lack of theoretical understanding of their properties.", "key_idea": "The authors introduce a Heterogeneous Graph Kernel-based Graph Neural Networks (HGK-GNN), implementing graph kernel to HGNNs. They incorporate Mahalanobis distance to build a Heterogeneous Graph Kernel (HGK), and use it in deep neural architectures, thus creating a heterogeneous GNN with a heterogeneous aggregation scheme.", "method": "The authors incorporate mathematical linkage between HGK-GNN and metapath-based HGNNs, which are highly effective variants of HGNNs, to analyze HGK-GNN. Both graph and node classification experiments are conducted to evaluate HGK-GNN on six real-world datasets.", "outcome": "HGK-GNN outperforms a wide range of baselines on six real-world datasets, showcasing the effectiveness of the solution.", "future_impact": "The introduction of HGK into the field of HGNNs marks a first step toward the theoretical understanding and analysis of HGNNs, which has potential to guide future research.", "venue": "SIGKDD", "year": 2021, "title": "HGK-GNN: Heterogeneous Graph Kernel based Graph Neural Networks"} +{"pid": "bae86adc-5bda-4497-a160-e43dcba13325", "context": "The representation of knowledge is usually not straightforward to comprehend, necessitating an easy-to-understand technique.", "key_idea": "The authors propose determinations as a knowledge representation that is easy to understand, and they describe an algorithm called CONDET that uses feature selection to construct determinations from training data.", "method": "The authors develop a method named CONDET which makes use of a condensation process to create simpler structures from training data. They conduct a series of experiments to evaluate the effectiveness of this condensation process.", "outcome": "The experimental results indicate that condensation reduces complexity without causing any loss in accuracy.", "future_impact": "The authors discuss the relation of CONDET to other works and suggest direction for future studies, implying there may be room for refinement or different applications of the algorithm.", "venue": "SIGKDD", "year": 1996, "title": "Induction of condensed determinations"} +{"pid": "5f0ec68f91e011ead966520a", "context": "Current methods for measuring algorithmic bias in computer vision use observational datasets and therefore fuse algorithmic bias with dataset bias, which can hinder the fair assessment and improvement of algorithms.", "key_idea": "The researchers propose a new experimental method for measuring algorithmic bias in face analysis algorithms which directly manipulates attributes like gender and skin tone to reveal causal links between attribute variation and performance.", "method": "The method revolves around generating synthetic image grids that vary specific attributes while keeping others unchanged, with human perception providing control for synthesis errors when quantifying algorithmic bias. The method is compared to an observational bias analysis in gender classification algorithms.", "outcome": "The two methods show conflicting results. The observational method reports gender and skin color biases, while the experimental method reveals biases related to gender, hair length, age, and facial hair. The synthetic transects also enable easier bias analysis on minority and intersectional groups.", "future_impact": "The experimental method developed could guide improvement in biases of algorithm performance in face analysis systems, especially for minority and intersectional groups.", "venue": "ECCV", "year": 2020, "title": "Towards Causal Benchmarking of Bias in Face Analysis Algorithms"} +{"pid": "b6a108a8-a7e6-4ec2-bd13-f28f2f6498a6", "context": "Existing shape-from-intensity methods estimate depth and surface normals separately and have difficulty dealing with discontinuous surfaces, as they rely on surface normal integration.", "key_idea": "The authors propose a new method to recover scene geometry from the second-bounce of light transport. The method does so by deriving form factors from the second-bounce component in a Lambertian case which carry information about the geometric relationship between all pairs of points in the scene.", "method": "The authors model the scene as polygonal and develop an optimization method to recover the scene geometry from the form factors up to a scaling ambiguity. The proposed theory and the method were demonstrated through a variety of simulations and real-world experiments.", "outcome": "The proposed method has been validated through various simulations and real-world experiments, demonstrating the correctness of the proposed theory of shape recovery from light transport.", "future_impact": "The proposed method paves the way for more advanced scene reconstruction technologies, particularly useful for areas where depth and surface normal characterizations are of prime importance.", "venue": "ECCV", "year": 2010, "title": "Shape from second-bounce of light transport"} +{"pid": "e4abfde6-a2cb-4fc9-b57f-0358020678ee", "context": "Traditional object detection approaches require intensive supervision for shape model construction and use parametric distributions to describe the spatial layout between parts, resulting in inflexibility and complexity in learning.", "key_idea": "The authors propose an efficient approach to construct shape models with partially-supervised learning that transfers part structures across different object classes with similar shapes. The spatial layout between parts is described by a non-parametric density for flexibility and ease of learning. Object detection is transformed into state estimation inference executed using a novel Particle Filters (PF) framework different from previous PF methods.", "method": "The proposed PF algorithm linearizes the fully connected graph underpinning the model by exploring the conditional dependencies of nodes representing contour parts.", "outcome": "Experimental results reveal that the proposed approach achieves excellent detection results and accurately locates contours of target objects in cluttered images.", "future_impact": "While not explicitly stated, the method may be adaptable to various shape-based object detection tasks and could improve the efficiency of detecting objects in cluttered images.", "venue": "ECCV", "year": 2010, "title": "Weakly supervised shape based object detection with particle filter"} +{"pid": "5f0d8e1e91e011047aff998f", "context": "Existing methods for few-shot semantic segmentation either focus on one-way few-shot segmentation or suffer from incomplete coverage of object regions.", "key_idea": "The authors propose a novel few-shot semantic segmentation framework that decomposes the holistic class representation into a set of part-aware prototypes for capturing diverse and fine-grained object features, and leverages unlabeled data to enrich these prototypes and better model intra-class variations of semantic objects.", "method": "A novel graph neural network model is developed to generate and enhance the proposed part-aware prototypes based on labeled and unlabeled images.", "outcome": "Experimental evaluations on two benchmarks show that the proposed method outperforms prior techniques.", "future_impact": "The proposed few-shot semantic segmentation framework could lead to improved segmentation of new object classes with few annotated examples, which has significant implications for real-world applications.", "venue": "ECCV", "year": 2020, "title": "Part-Aware Prototype Network for Few-Shot Semantic Segmentation"} +{"pid": "8a8dacfa-3ce5-4a3c-b6ac-b127723abded", "context": "Long-term human motion can be represented as a series of motion modes\u2014motion sequences that capture short-term temporal dynamics\u2014with transitions between them.", "key_idea": "The authors propose a Motion Transformation Variational Auto-Encoders (MT-VAE) model that jointly learns a feature embedding for motion modes and a feature transformation that represents the transition of one motion mode to the next.", "method": "The authors present MT-VAE for learning motion sequence generation and they apply this approach to both facial and full body motion.", "outcome": "The proposed MT-VAE model is able to generate multiple diverse and plausible motion sequences in the future from the same input.", "future_impact": "The authors demonstrate successful applications such as analogy-based motion transfer and video synthesis, presumably opening the door for further development and deployment in related areas.", "venue": "ECCV", "year": 2018, "title": "MT-VAE: Learning Motion Transformations to Generate Multimodal Human Dynamics"} +{"pid": "bf38840d-3307-4c9d-8168-288061876fe4", "context": "Structure from Motion (SfM) with rolling shutter cameras has various issues including camera configurations, like parallel readout directions, which allow for a large class of ambiguities in multi-view reconstruction.", "key_idea": "The study tackles problems in Structure from Motion with rolling shutter cameras, pointing out that some common camera configurations become critical and result in ambiguities in multi-view reconstructions.", "method": "The authors provide mathematical analysis for one, two and some multi-view cases and verify these through synthetic experiments. They also evaluate practical photographic techniques with rolling shutter cameras and provide a quantitative analysis of their performance in real experiments.", "outcome": "The study found that bundle adjustment with rolling shutter cameras, even ones close to critical configurations, can produce drastically deformed reconstructions. The authors' results show their practical recipes can aid in properly photographing with rolling shutter cameras to avoid scene deformations in Structure from Motion.", "future_impact": "The findings of this study can have practical implications in guiding photographers on the best practices for photographing with rolling shutter cameras when dealing with Structure from Motion, leading to correct 3D model reconstructions.", "venue": "ECCV", "year": 2016, "title": "Degeneracies in Rolling Shutter SfM"} +{"pid": "fc9b4536-4733-4c1f-82a6-24fdaf0f965a", "context": "Generative Adversarial Networks (GANs) have demonstrated remarkable results in image-to-image translation, but these models often only transfer low-level information (color, texture, etc.) while failing to manipulate higher-level semantic elements. Existing models can synthesize compelling images given class labels or captions, but they don't support adjusting shapes or structures, which limits their use cases and interpretability.", "key_idea": "The authors introduce a new generative model for semantic manipulation, termed as 'Contrast-GAN.' It aims to change the semantic elements of an object category while preserving its characteristics such as viewpoints and shapes (e.g., cow to sheep, motor to bicycle, cat to dog).", "method": "The proposed contrast-GAN operates under a novel adversarial contrasting objective, optimizes over the distance comparisons between samples, and is designed to enable disentangling of image background with object semantic changes. The authors evaluate this approach through extensive qualitative and quantitative experiments on multiple semantic manipulation tasks using ImageNet and the MSCOCO dataset.", "outcome": "Experimental results show that contrast-GAN significantly outperforms other conditional GAN models for semantic manipulation tasks on the ImageNet and MSCOCO datasets.", "future_impact": "As contrast-GAN can manipulate high-level semantics in images whilst retaining characteristics such as shapes and structures, it could broaden the application scenarios and interpretability of generative models, opening up new possibilities for advanced image manipulation techniques.", "venue": "ECCV", "year": 2018, "title": "Generative Semantic Manipulation with Mask-Contrasting GAN"} +{"pid": "635024b890e50fcafd303bd3", "context": "Most lesion segmentation methods from CT slices focus on one specific lesion type, and the existing universal lesion segmentation methods are weakly-supervised because of the lack of pixel-level annotation data.", "key_idea": "The authors establish a large-scale universal lesion dataset called SegLesion and propose a baseline method for the task. They introduce a Knowledge Embedding Module (KEM) and a Knowledge Embedding Network (KEN) to increase the distinguishability of CT slices for universal lesion segmentation.", "method": "They learn the knowledge encoding of CT slices using KEM and embed the encoded knowledge into the deep features of a CT slice. Furthermore, they build a large benchmark for SegLesion to extensively compare KEN with previous segmentation methods.", "outcome": "KEN, incorporating KEM achieves state-of-the-art performance for universal lesion segmentation.", "future_impact": "The proposed KEN serves as a strong baseline for future research in universal lesion segmentation. The established large-scale universal lesion segmentation dataset, SegLesion, will also benefit the community.", "venue": "ECCV", "year": 2022, "title": "Delving into Universal Lesion Segmentation: Method, Dataset, and Benchmark."} +{"pid": "fc103813-c19b-4654-bfca-8db2e06f10df", "context": "Current mobility aids for visually impaired people may be cumbersome and ineffective in providing safe and autonomous navigation in unknown environments.", "key_idea": "The authors propose a compact, lightweight, and energy-efficient mobility aid system that uses dense depth maps from a compact stereo vision system installed in the user's glasses. The system provides real-time vibration and audio feedback about obstacles via bone-conductive speakers.", "method": "The system was deployed and tested in both indoor and outdoor environments with users simulating visual impairment, including a blind person.", "outcome": "The experimental evaluation indicated that the system provides an ergonomic and comfortable user interface with a fast learning curve for effective deployment.", "future_impact": "This system could potentially improve the quality of life of individuals with visual impairments by enabling safe and autonomous navigation, and it has the potential for further improvements and wider applications.", "venue": "ECCV", "year": 2014, "title": "3D Glasses as Mobility Aid for Visually Impaired People"} +{"pid": "de33fe2c-b598-437c-9d82-4ff31cc476ce", "context": "Scene recognition methods traditionally struggle with encoding spatial information and finding a suitable similarity measure that respects the manifold structure of the semantic space. There is a need to improve these aspects while maintaining low computational complexity.", "key_idea": "The authors propose a new architecture, Spatial Pyramid Matching on the Semantic Manifold (SPMSM), which refers to an image representation on a semantic probability simplex coupled with a rough encoding of spatial information. It integrates geodesic distance between two manifold points as a natural measure of image similarity.", "method": "Several experiments are reported in which SPMSM is compared with state-of-the-art recognition methods on two large datasets, MIT Indoor and SUN, and several smaller datasets.", "outcome": "The presented SPMSM approach shows superior recognition rates for both the MIT Indoor and SUN datasets, and equivalent or superior performance on several smaller datasets. Importantly, the resulting SVM has a smaller dimensionality and requires fewer support vectors than previous classifiers, indicating smaller complexity.", "future_impact": "The reduced complexity and faculty for improved generalization of the SPMSM architecture may enable more effective scene recognition applications beyond the tested datasets.", "venue": "ECCV", "year": 2012, "title": "Scene recognition on the semantic manifold"} +{"pid": "b5d232ad-9aaa-4a9b-8032-c55466060606", "context": "Average precision (AP), the area under the recall-precision (RP) curve, is the standard performance measure for object detection. However, it has shortcomings such as the inability to distinguish very different RP curves and the lack of directly measuring bounding box localization accuracy.", "key_idea": "The authors propose a new metric, \"Localization Recall Precision (LRP) Error\", specifically designed for object detection. LRP Error is composed of three components: localization, false negative rate, and false positive rate.", "method": "Based on LRP, the authors introduce \"Optimal LRP\" (oLRP), which determines the \"best\" confidence score threshold for a class. The authors also present the performances of object detectors using class-specific optimized thresholds. They use PASCAL VOC AND MSCOCO datasets for conducting these experiments.", "outcome": "The experiments show that oLRP provides richer and more discriminative information than AP. The best confidence score thresholds vary significantly among classes and detectors, and using class-specific optimized thresholds increases the accuracy against using a general threshold for all classes.", "future_impact": "The findings from the experiments could cause a shift from AP to LRP as the performance measure for object detection due to its ability to provide more valuable information, leading to more accurate and superior models.", "venue": "ECCV", "year": 2018, "title": "Localization Recall Precision (LRP): A New Performance Metric for Object Detection"} +{"pid": "635024af90e50fcafd302917", "context": "In self-supervised learning methods (SSL), maximizing the mutual information between two augmented views has achieved significant success, with cropping being a popular augmentation technique. However, the contribution of unused regions after cropping has been barely explored in existing methods, though they contribute to the description of the category within the same image instance.", "key_idea": "The authors propose a novel pretext task called Region Contrastive Learning (RegionCL) that emphasizes the importance of considering both cropped and remained regions in an image instance for self-supervised learning.", "method": "RegionCL randomly crops a region from each input image and swaps it with different images to compose new images that involve both cropped and remain regions. These regions are then employed to efficiently construct contrastive pairs for positive or negative samples, taking into consideration augmentation from the same or different original images.", "outcome": "Experiments on ImageNet, MS COCO, and Cityscapes demonstrate that RegionCL significantly improves other SSL methods like MoCov2, DenseCL, and SimSiam, and achieves state-of-the-art performance on classification, detection, and segmentation tasks.", "future_impact": "The novel methodology of RegionCL facilitating the use of both cropped and remained regions in an image validates an effective way to enhance visual representation learning which might spur future advancements in self-supervised learning methods.", "venue": "ECCV", "year": 2022, "title": "RegionCL: Exploring Contrastive Region Pairs for Self-supervised Representation Learning."} +{"pid": "b60c5968-66ac-46b3-bb3c-39d72c4d2704", "context": "Efficient access to information contained in video databases requires a structured representation of the video content to be built beforehand, which is a challenge given the complexity and diversity of video contents.", "key_idea": "The authors propose a method for building a structured spatio-temporal representation of video content, which includes video partition into shots, characterization of camera motion, object extraction and tracking. The content of each shot is visualized on a synoptic frame composed of a mosaic image of the background scene, with the trajectories of mobile objects superimposed.", "method": "The proposed method relies on robust motion estimation, statistical tests and contextual statistical labeling. It is designed to keep computational cost low while being able to cope with general video content. Experimental testing was conducted on real-world sequences.", "outcome": "The paper demonstrates a structured approach for video indexing and browsing, and provides efficient access to information in video databases.", "future_impact": "The structured output of the method opens up possibilities for various extensions, such as higher-level interpretation of video content, which can significantly benefit tasks related to video content analysis.", "venue": "ECCV", "year": 1998, "title": "Determining a Structured Spatio-Temporal Representation of Video Content for Efficient Visualization and Indexing"} +{"pid": "ea3b26ad-6190-4e67-aeee-475594e636cd", "context": "Previous approaches to recognizing objects in cluttered scenes differ depending on the dimensionality of the objects and the sensors used to capture the object data. The efficient online recognition of these objects remains a challenge.", "key_idea": "The authors discuss the Geometric Hashing paradigm for model-based recognition. It allows for a unified approach to rigid object recognition under different viewing transformation assumptions for both 2-D and 3-D objects, obtained from different sensors. The paradigm relies on intensive offline model preprocessing and transformation invariant features.", "method": "The authors indexed model information into a hash-table using a preprocessing stage and then used this for online recognition. The method is parallelizable and was tested on both 2-D and 3-D objects in cluttered scenes from an arbitrary viewpoint.", "outcome": "The initial experimentation of this technique successfully recognized both 2-D and 3-D objects in cluttered scenes from an arbitrary viewpoint.", "future_impact": "Extensions of the basic Geometric Hashing paradigm are mentioned that promise to reduce its worst case recognition complexity, thus a potential for further research and advancements in object recognition.", "venue": "ECCV", "year": 1990, "title": "Model-based object recognition by geometric hashing"} +{"pid": "aa7cae53-b4dc-42b7-91cf-2f3698ed3540", "context": "Realistic rendering of human behavior for applications such as video animations, virtual reality, and gaming engines usually relies on manually created 3D models based on sequences of coarse body shape representations. These 3D models are typically designed manually or reconstructed from costly sensor data.", "key_idea": "The authors propose a holistic learning framework for rendering human behavior where all components are learned from easily available data. This framework includes control over the generated behavior based on motion capture data or direct copying of behavior from videos.", "method": "The authors use motion capture data and RGB camera data to train their learning framework. They improve data efficiency by training on multiple characters simultaneously.", "outcome": "The authors have shown their approach to successfully render human behavior, and the methodology improves data efficiency when trained with multiple characters.", "future_impact": "This novel approach may streamline the process of creating animations and games, and it shows a new path towards easily available, personalized avatar creation.", "venue": "ECCV", "year": 2018, "title": "Towards Learning a Realistic Rendering of Human Behavior"} +{"pid": "c043cc58-1389-42a9-8a3c-09c2a1ec4e2d", "context": "The lack of eye contact in desktop video teleconferencing reduces the effectiveness of video content. While hardware solutions exist, they tend to be expensive and bulky. Researchers have been attempting to provide a practical software-based solution.", "key_idea": "The paper presents a novel approach to correcting eye gaze, based on stereo analysis and a personalized face model which together allow for a 3D description of the face and hence the synthesis of virtual views that maintain eye contact.", "method": "The authors use stereo techniques that include both feature matching and template matching. The accuracy and robustness of the stereo analysis are improved by the personalized face model which substantially reduces the search range.", "outcome": "The system can generate an eye-gaze corrected video stream at about 5 frames per second on a typical PC.", "future_impact": "N/A.", "venue": "ECCV", "year": 2002, "title": "Eye Gaze Correction with Stereovision for Video-Teleconferencing"} +{"pid": "45348ba0-2cf1-4f81-92b0-7af5963808fa", "context": "The existing algorithms for structure from motion from tracked features over a number of images typically have complexity that depends on the number of images and do not handle changing features or missing data well.", "key_idea": "The paper presents a new recursive algorithm for structure from motion that is independent of the number of images, allows for dynamic inclusion and removal of features, can handle missing data, and does not require heuristic segmentation strategies for outlier elimination.", "method": "The algorithm is tested with different kinds of tracked features such as edges and corners, demonstrating its versatility.", "outcome": "The algorithm is found to be efficient with its time complexity linear to the number of tracked features and independent of the number of images. It successfully handles dynamic feature changes and missing data.", "future_impact": "The proposed algorithm's ability to employ different types of tracked features opens up possibilities for more flexible and robust structure-from-motion applications.", "venue": "ECCV", "year": 1994, "title": "Recursive affine structure and motion from image sequences"} +{"pid": "2332da07-189b-4138-9584-24dc37a88a6d", "context": "The problem of graph matching, which plays a crucial role in many computer vision applications, is generally NP-hard and most existing work is focused on suboptimal solutions between two graphs. Traditionally, pairwise affinity matching score and global matching consistency that affect the overall accuracy are either enforced early in iterative optimization or tackled separately in two steps.", "key_idea": "This paper proposes a more general approach to the problem of matching N attributed graphs to each other, a method that jointly tackles the affinity score and the global matching consistency - the two key elements affecting the overall accuracy of graph matchings.", "method": "A novel multi-graph matching algorithm is designed to iteratively approximate the global-optimal affinity score, gradually incorporating the consistency as a regularizer. This method improves the performance of the initial solutions obtained by existing pairwise graph matching solvers.", "outcome": "The proposed algorithm shows notable efficacy on both synthetic and public image datasets, and it has a theoretically proven convergence.", "future_impact": "The proposed approach of jointly addressing the affinity score and matching consistency could pave the way for more robust and effective solutions to the multi-graph matching problem in the field of computer vision.", "venue": "ECCV", "year": 2014, "title": "Graduated Consistency-Regularized Optimization for Multi-graph Matching"} +{"pid": "78f8479f-c4f5-4993-9f64-2d69532f15e3", "context": "Eye typing has become an alternative form of text entry for physically challenged people with severe motor disabilities, but additional eye-tracking devices are usually needed to track eye movements, which can be inconvenient.", "key_idea": "The authors propose an appearance-based method to estimate a person's gaze point using a standard webcam, eliminating the need for additional eye-tracking devices.", "method": "The authors investigate practical issues of the proposed method but the specific methodology is not mentioned in the abstract.", "outcome": "Experimental results demonstrate the feasibility of eye typing using the proposed webcam-based gaze estimation method.", "future_impact": "The provided information in the abstract does not explicitly discuss potential future impacts.", "venue": "ECCV", "year": 2016, "title": "Feasibility Analysis of Eye Typing with a Standard Webcam"} +{"pid": "62d8c4555aee126c0f7626be", "context": "The roles of shape, texture, and color in object classification with reference to the human visual system have not been explicitly separated and studied in computer vision systems.", "key_idea": "The authors propose a humanoid vision engine (HVE) that calculates shape, texture, and color features from images separately and explicitly and then combines them for final object classification.", "method": "The authors set up experiments to use HVE to rank the contributions of shape, texture, and color features to object recognition and simulate human task-specific and class-specific feature utilization. They also use HVE to simulate open-world zero-shot learning abilities and human imagination capabilities.", "outcome": "Experiments confirm that HVE, like humans, relies on specific features for classifying certain objects. The HVE can identify the most crucial features for the entire task (e.g., color for the CUB dataset) and for each class (e.g., shape for recognizing boats in the iLab-20M dataset).", "future_impact": "The open-sourcing of the HVE engine and corresponding datasets could facilitate further research in this area.", "venue": "ECCV", "year": 2022, "title": "Contributions of Shape, Texture, and Color in Visual Recognition."} +{"pid": "61ca80355244ab9dcba699f6", "context": "Responsive feedbacks of a listener during face-to-face conversation have seldomly been studied in literature, despite their importance to applications such as digital human, virtual agents, and social robots.", "key_idea": "The authors propose a novel dataset 'ViCo' for listening head generation during face-to-face conversation, which highlights the importance of non-verbal feedbacks such as head motions and facial expressions in response to the speaker's audio and visual signals.", "method": "The dataset contains 92 identities (67 speakers and 76 listeners), featuring 483 clips in a paired 'speaking-listening' pattern where listeners show three listening styles: positive, neutral, negative. The baseline model for listening head generation is conditioned on these listening attitudes.", "outcome": "The authors introduce a novel dataset 'ViCo' and release a listening head generation baseline, conditioning on different listening attitudes.", "future_impact": "The ViCo dataset can support a wide range of applications such as human-to-human interaction, video-to-video translation, cross-modal understanding and generation, enabling future research in these areas.", "venue": "ECCV", "year": 2022, "title": "Responsive Listening Head Generation: A Benchmark Dataset and Baseline."} +{"pid": "0d9253c0-17ad-4fde-a7c2-344a84413420", "context": "Structure from Motion (SFM) suffers from the problem of noise sensitivity. Inherent motion ambiguities and errors in motion estimates often make accurate metric depth estimate difficult to obtain.", "key_idea": "Despite inaccuracies in metric depth estimates, this paper investigates the possibility of extracting some valid and useful depth information, specifically, the resolution of ordinal depth.", "method": "The authors derive a sufficient condition based on a general depth distortion model to validate the extraction of ordinal depth. The geometry and statistics of image regions satisfying this condition are studied.", "outcome": "It is found that even with inaccurate metric depth estimates, ordinal depth can still be discerned locally if physical metric depth difference is beyond a certain discrimination threshold. However, various factors such as visual angle, speed of motion carrying depth information, and points' distance from the camera affect the resolution level of discernible ordinal depth.", "future_impact": "The findings suggest that accurate knowledge of qualitative 3D structure is ensured in a small local image neighborhood, which might account for biological foveated vision and enhance understanding of the nature of the perceived visual space.", "venue": "ECCV", "year": 2008, "title": "Behind the Depth Uncertainty: Resolving Ordinal Depth in SFM"} +{"pid": "ad621c3b-c395-4e2f-9a09-fbddc02edfb0", "context": "The deformation of applicable surfaces like sheets of paper satisfies differential geometric constraints. Understanding this deformation has typically required complex calculations and multiple views of the surface.", "key_idea": "The authors demonstrate that the constraints of isometry and vanishing Gaussian curvature provide a closed set of equations that allow recovery of the full geometric structure from a single image of the surface and knowledge of its undeformed shape.", "method": "The authors reduce the partial differential equations to the Hopf equation. They develop a new exact integration of these equations that relates the 3D structure of the applicable surface to an image and test it by comparing with particular exact solutions.", "outcome": "The authors successfully obtained the 3-D structure of applicable surfaces from a single image using newly developed equations, and presented results for both the forward and the inverse 3D structure recovery problem.", "future_impact": "The new ability to recover 3-D structure from a single image of applicable surfaces could revolutionize multiple fields such as manufacturing, animation, and design.", "venue": "ECCV", "year": 2004, "title": "Structure of Applicable Surfaces from Single Views"} +{"pid": "60742028e4510cd7c8726b26", "context": "The rapid change in fashion trends and the need for garment models in applications like virtual try-on have escalated the importance of having a general cloth model that can represent a diverse set of garments. However, previous studies make assumptions about the garment such as the topology or suited body shape.", "key_idea": "The authors propose a unified method using a generative network that can be applied to different garment topologies with different sewing patterns and fabric materials. They also introduce a novel image representation of garment models and a reliable mapping algorithm for data representation of the cloth.", "method": "The authors provide a large garment appearance dataset for garment reconstruction, and demonstrate the generative model's performance using this dataset.", "outcome": "The generated garment model can be easily retargeted to another body, enabling garment customization. It is shown that the generative model has high reconstruction accuracy and can provide rich variations of virtual garments.", "future_impact": "The research provides a unified method for garment generation suitable for diverse garment topologies, which could be useful in garment reconstruction, garment capturing, virtual try-ons, and other applications. The novel image representation and mapping algorithm can regularize data representation of clothes, potentially enhancing garment customization.", "venue": "ECCV", "year": 2020, "title": "GAN-Based Garment Generation Using Sewing Pattern Images"} +{"pid": "635024b790e50fcafd3038f7", "context": "Automatic Check-Out (ACO) models aim to accurately predict the presence and count of each category of products in check-out images, but there is a significant domain gap between training data (single-product exemplars) and test data (check-out images).", "key_idea": "The authors propose Prototype-based classifier learning from Single-Product exemplars (PSP), a method designed to reveal the advantages of representing category semantics and manage fine-grained classifications. This includes a multi-label recognition loss to model the co-occurrence of products in check-out images.", "method": "The authors conduct experiments using the large-scale RPC dataset to evaluate the proposed PSP method's impact on Automatic Check-Out (ACO) accuracy, with improvements bolstered by discriminative re-ranking of predicted product proposal scores.", "outcome": "The result of the ACO using the proposed PSP method achieves 86.69%, seeing significant improvement over the state of the art by 6.18%.", "future_impact": "The provided method and source code have the potential to advance Automatic Check-Out (ACO) systems and improve product recognition capabilities, making the checkout process more efficient and accurate.", "venue": "ECCV", "year": 2022, "title": "Automatic Check-Out via Prototype-Based Classifier Learning from Single-Product Exemplars."} +{"pid": "62008da15aee126c0fbd19fd", "context": "General Purpose Vision (GPV) systems, which are designed to solve a wide array of visual tasks without architectural changes, primarily learn skills and concepts from fully supervised datasets. However, scaling GPVs to tens of thousands of concepts by acquiring data to learn each concept for every skill quickly becomes prohibitive.", "key_idea": "This study proposes an inexpensive alternative for GPV systems to learn skills from supervised datasets and concepts from web image search, leveraging the characteristic of GPVs to transfer visual knowledge across skills. The authors also propose an architecture, GPV-2, that supports various tasks from vision tasks to vision+language tasks.", "method": "A dataset of over 1 million images spanning more than 10,000 visual concepts is used to demonstrate webly-supervised concept expansion for two existing GPVs (GPV-1 and VL-T5) on three benchmarks: 5 Coco-based datasets, a series of 5 datasets based on the OpenImages and VisualGenome repositories, and the web-derived dataset.", "outcome": "The proposed GPV-2 architecture demonstrates significant benefits from web data, outperforming GPV-1 and VL-T5 across these benchmarks.", "future_impact": "The proposed approach and new architecture could pave the path for more efficient and scalable learning for General Purpose Vision systems, enabling them to handle a wide array of tasks without extensive and prohibitive data gathering for each concept and skill.", "venue": "ECCV", "year": 2022, "title": "Webly Supervised Concept Expansion for General Purpose Vision Models."} +{"pid": "623a900f5aee126c0f36d0cc", "context": "Pre-training on large supervised datasets such as ImageNet has been essential for effective domain transfer in deep learning. However, the effects of pre-training specifics such as network architectures, size, pre-training loss, and datasets on these tasks have not been well-studied.", "key_idea": "The paper aims to provide an in-depth analysis and broad study of various factors related to pre-training for domain generalization and adaptation, including network architecture, size, pre-training loss, and the dataset used for pre-training.", "method": "The research involves testing different network backbones and pre-training configurations on the Office-Home and DomainNet domain adaptation tasks. Comparisons are made against existing state-of-the-art baselines.", "outcome": "The study concludes that merely using a state-of-the-art backbone can outperform previous domain adaptation baselines, improving results on Office-Home and DomainNet by 10.7% and 5.5%, respectively.", "future_impact": "The insights drawn from this study could guide and influence future domain transfer research.", "venue": "ECCV", "year": 2022, "title": "A Broad Study of Pre-training for Domain Generalization and Adaptation."} +{"pid": "bfc32a75-6083-4245-98b0-4ba51c115374", "context": "Advances in photo editing and manipulation tools have made it easier to create fake imagery. The detection of such manipulations is a challenging problem due to the lack of manipulated training data.", "key_idea": "The paper proposes a learning algorithm for detecting visual image manipulations using only a large dataset of real photographs. The model uses photo EXIF metadata for training a model to determine whether an image content is self-consistent.", "method": "The method uses photo EXIF metadata as supervisory signal for training a model. This self-consistency model is then applied to the task of detecting and localizing image splices.", "outcome": "The proposed method performs at state-of-the-art level on several image forensics benchmarks, despite never being trained on any manipulated images.", "future_impact": "The paper signifies the step towards achieving a truly general purpose visual forensics tool.", "venue": "ECCV", "year": 2018, "title": "Fighting Fake News: Image Splice Detection via Learned Self-Consistency"} +{"pid": "7afad8c7-caf7-4ee4-b162-b5231b44f7a8", "context": "Patch descriptors are used for various tasks such as finding corresponding points across images and describing object category parts. Existing works often use histograms of gradients.", "key_idea": "The study proposes an image patch descriptor based on edge position, orientation, and local linear length, unlike previous works which use relative gradient magnitudes. The algorithm locally normalizes the patch gradients and uses orientation dependent binning.", "method": "To validate the approach, the method is tested on previously published object instance patch data sets, as well as a new patch data set that models intra-category appearance variations.", "outcome": "The proposed algorithm produces state-of-the-art results on the test data sets.", "future_impact": "Fast PCA dimensionality reduction and a min-hash approach for fast patch retrieval are proposed as possible extensions, which could potentially enhance the scalability and effectiveness of the method.", "venue": "ECCV", "year": 2010, "title": "Binary coherent edge descriptors"} +{"pid": "68b54dfa-3705-4f0f-a75e-b3669267239b", "context": "The challenge of estimating the three-dimensional shape and radiance of a surface in space from images obtained with different focal settings has traditionally been addressed using shift-invariant imaging models or methods that only work with equifocal surfaces.", "key_idea": "In this paper, the authors propose using an infinite-dimensional optimization to model the problem, which they solve through a carefully crafted partial differential equation (PDE). This approach allows them to deterministically govern regularization, galvanizes the process by utilizing level set methods for the PDE solving, and circumvents the constraint of assuming a shift-invariant imaging model or being limited to equifocal surfaces.", "method": "In their study, the authors address the problem of 3D shape and radiance estimation of a surface in space from images taken at different focal settings by numerically solving a PDE that models this as an infinite-dimensional optimization problem.", "outcome": "The authors do not provide specific outcomes or results in the abstract.", "future_impact": "The authors do not speculate on the future impact of their work in the abstract.", "venue": "ECCV", "year": 2002, "title": "A Variational Approach to Shape from Defocus"} +{"pid": "05c9b5e9-502a-4b5a-8a49-0f37118f2529", "context": "The problem of recognizing human actions under view changes is challenging. Most solutions focus on either structure recovery or multi-view correspondence estimation, both of which can be complex and computationally demanding.", "key_idea": "The authors propose a novel action descriptor that captures the structure of temporal similarities and dissimilarities within an action sequence. This descriptor leverages the observed stability of self-similarities of action sequences across different views.", "method": "The proposed method combines weak geometric properties with machine learning techniques for efficient cross-view action recognition, and is validated through experiments on three public datasets.", "outcome": "The proposed method exhibits similar or superior performance compared to related methods and performs well even in extreme conditions such as recognizing actions from top views while using side views for training only.", "future_impact": "This simple action recognition approach, which does not require complex operations such as structure recovery or multi-view correspondence estimation, has the potential to improve the efficiency and accuracy of action recognition tasks across various viewpoints.", "venue": "ECCV", "year": 2008, "title": "Cross-View Action Recognition from Temporal Self-similarities"} +{"pid": "60001c8891e011c8f78fd36d", "context": "The development of computational methods for understanding bodily-expressed emotions is an ongoing effort in diverse fields like computer vision, robotics, psychology/psychiatry, graphics, data mining, machine learning, and movement analysis.", "key_idea": "The paper presents a panel discussion involving active researchers in closely-related fields to discuss the future of this research area.", "method": "The paper documents the opinions expressed by the individual panelists involved in the discussion.", "outcome": "The document encapsulates the perspectives and opinions shared by the panelists on the subject of computational methods for emotion understanding.", "future_impact": "Though the abstract does not explicitly mention the future impact, it alludes towards a collaborative multidisciplinary approach to the ongoing development in decoding bodily expressed emotions.", "venue": "ECCV", "year": 2020, "title": "Panel - Bodily Expressed Emotion Understanding Research - A Multidisciplinary Perspective."} +{"pid": "5f44d54f91e011872f85ec18", "context": "Existing deep networks designed for point cloud data analysis often struggle with out-of-distribution samples, noise, and distortions that exist in test data. These networks are prone to overfitting and unpredictable behavior when test data is partially changed.", "key_idea": "The authors introduce a new deep network for 3D model analysis, which utilises the theory of 'orderly disorder'. The method extracts deep patterns within a 3D object, seeking the most stable patterns and discarding the unstable ones.", "method": "The analysis is based on an innovative cloning decomposition technique whereby features are extracted and then linked to each other to form stable complex patterns. Experiments were carried out on several benchmark datasets to assess network performance, particularly in the presence of noise.", "outcome": "The new model alleviates the vanishing-gradient problem, strengthens dynamic link propagation, and substantially reduces the number of parameters. The network's performance drops less than 10% in the presence of noise, whereas the performance of state-of-the-art networks significantly diminished.", "future_impact": "The proposed light network may improve the segmentation and classification tasks of 3D models, particularly in noisy environments, potentially leading to more robust and stable solutions in the field of 3D data analysis.", "venue": "ECCV", "year": 2020, "title": "Orderly Disorder in Point Cloud Domain"} +{"pid": "601a76a591e011689f55223a", "context": "The current literature involves the pruning of neural networks however they do not adaptively prune the network based on the convolutional weights.", "key_idea": "The authors propose a novel weight-dependent gates (W-Gates) framework for network pruning that prunes filters in a network based on the convolutional weights, effectively learning whether to prune or keep filters automatically.", "method": "The authors trained a Latency Predict Net (LPNet) to estimate the hardware latency of candidate pruned networks, and optimized the W-Gates and the pruning ratio of each layer under latency constraints. The method was tested on ImageNet using the models Resnet34 and Resnet50.", "outcome": "The weight-dependent gates framework demonstrated superior performance compared to state-of-the-art pruning methods showing an increase of up to 1.33/1.28 higher Top-1 accuracy with lower hardware latency on ImageNet when tested with Resnet34 and Resnet50.", "future_impact": "The proposed method brings about a new perspective on neural network pruning, introducing the use of weight-dependent gates. This can inspire new designs adaptive pruning methods and lead to more efficient neural networks.", "venue": "ECCV", "year": 2020, "title": "Weight-Dependent Gates for Differentiable Neural Network Pruning."} +{"pid": "a000e8b0-bd54-4075-942b-ef0f970cc9a2", "context": "Existing techniques in surface matching and computing distance maps on manifolds can be unstable or inefficient, and may struggle with large deformations and topological changes.", "key_idea": "The authors propose a formulation for geodesic distance propagation of surfaces that involves setting up a partial differential equation for the propagation of surfaces on a 3-manifold, and a surface matching method that generalizes a curve matching method.", "method": "The authors apply an Eulerian formulation with level-sets to derive numerically stable algorithms for computing distance maps, which are then used to conduct surface matching.", "outcome": "The proposed method can provide matching paths as the orbits of the vector field defined by the sum of two distance maps' gradient values, hence it can handle surface matching cases involving large deformations and topological changes.", "future_impact": "The proposed theory and method provide a more general solution for surface matching problems, potentially benefiting tasks that deal with large deformations and topological changes.", "venue": "ECCV", "year": 2000, "title": "Surface Matching with Large Deformations and Arbitrary Topology: A Geodesic Distance Evolution Scheme on a 3-Manifold"} +{"pid": "ea6012dd-4681-44ba-8f0a-dade30559bc9", "context": "Separating foreground from background in video images is useful in many applications, but it typically requires the processing of large amounts of information, leading to high computational cost and/or manual interaction.", "key_idea": "The authors propose a generative model that can separate foreground from background in video images in an efficient and visually pleasing manner.", "method": "The proposed method selects salient appearance poses of the foreground across the entire sequence in an unsupervised way and uses them to better extract the foreground from the single frames.", "outcome": "The results from the model validate the effectiveness of the approach in separating foreground and background elements.", "future_impact": "The proposed model, designed for efficiency and visually pleasing results, can alleviate computational costs and improve results in applications like video compression, human-computer interaction, and object tracking.", "venue": "ECCV", "year": 2008, "title": "Capturing video structure with mixture of probabilistic index maps"} +{"pid": "5f1ea6fc91e0119a0f6b6551", "context": "The challenge of teaching robots or virtual assistants to gesture naturally and in the style of a specific speaker, in line with corresponding speech (co-speech gestures), remains an unexplored area. Additional complexity in this field involves learning to generate gestures for multiple speakers and retaining the uniqueness of each speaker's style.", "key_idea": "This paper introduces Mix-StAGE, a model that teaches a single system to generate gestures for multiple speakers while learning unique style embeddings for each speaker\u2019s gestures, in an end-to-end manner. Uniquely, it learns a mixture of generative models, allowing for conditioning on each speaker's unique gesture style, and can alter gesturing styles for the same input speech by merely switching the style embeddings.", "method": "The authors introduce the Pose-Audio-Transcript-Style (PATS) dataset, which is designed to study gesture generation and style transfer, and use this dataset to train the Mix-StAGE model.", "outcome": "The proposed Mix-StAGE model significantly outperforms the previous state-of-the-art approach in gesture generation and allows for gesture style transfer across multiple speakers.", "future_impact": "The Mix-StAGE model offers a path towards performing gesture style transfer across multiple speakers and could improve the authenticity of gesturing in robots or virtual assistants.", "venue": "ECCV", "year": 2020, "title": "Style Transfer for Co-speech Gesture Animation: A Multi-speaker Conditional-Mixture Approach"} +{"pid": "5f53687891e0110c40a7be14", "context": "Deep neural networks are highly effective at various discriminative classification tasks, but they are poorly calibrated and often assign high probability to misclassified predictions. Combining deep neural networks and Bayesian inference has been suggested as a solution to quantify uncertainty over model predictions for ambiguous data points.", "key_idea": "The authors propose to validate and test the efficacy of models based on likelihood in the task of out-of-distribution detection (OoD). They also hypothesise that sensitivity to unseen inputs in neural networks maybe associated to multiple factors, which are amplified by the curse of dimensionality.", "method": "The authors test Bayesian deep learning models across different datasets and metrics, to ascertain their ability to outperform conventional neural networks. They also assess models in scenarios where there's minimal overlap between in and out of distribution classes. A study is also performed to find the effect of adversarial noise resistance methods on in and out-of-distribution performance.", "outcome": "Bayesian deep learning models only marginally outperform conventional neural networks on certain occasions and exhibit a reduction in AUC scores in detecting OoD data when the overlap between in and out of distribution classes is minimal.", "future_impact": "The findings indicate an inherent role of bias due to choices of initialisation, architecture, and activation functions, suggesting the need for further investigation. Also, this work contributes to understanding the robustness of Bayesian deep learners against adversarial noise, sparking further research in this area.", "venue": "ECCV", "year": 2020, "title": "Ramifications of Approximate Posterior Inference for Bayesian Deep Learning in Adversarial and Out-of-Distribution Settings"} +{"pid": "9dd9546f-b765-481d-a915-254e6d05308c", "context": "State-of-the-art face detectors based on single-stage networks are potentially vulnerable to attacks. Existing adversarial perturbation methods are not efficient in performing such attacks, particularly when multiple faces are present in the input image, leading to the Instance Perturbation Interference (IPI) problem.", "key_idea": "This paper focuses on the IPI problem and addresses it by analyzing the relationship between the deep neural network receptive field and the adversarial perturbation, leading to the development of a solution called Localized Instance Perturbation (LIP).", "method": "The effectiveness of the proposed LIP is tested by confining the adversarial perturbation inside the Effective Receptive Field (ERF) of a target and then comparing its performance against existing adversarial perturbation generation methods.", "outcome": "Experimental results indicate that the proposed LIP method massively outperforms existing adversarial perturbation generation methods \u2013 often by a factor of 2 to 10.", "future_impact": "The proposed method contributes to enhancing the security of smart surveillance systems utilizing face detectors and privacy protection, as it prevents faces from being harvested and stored in servers.", "venue": "ECCV", "year": 2018, "title": "Using LIP to Gloss Over Faces in Single-Stage Face Detection Networks"} +{"pid": "5ff8803a91e011c832668d07", "context": "Sign Language (SL) is vital for communication for individuals with hearing impairment, and it relies heavily on facial expressions for its grammatical structure. Automatic Sign Language Recognition (ASLR) technology is used to translate sign language gestures into written or spoken sentences of a target language, but the accuracy of ASLR can sometimes be impacted due to the lack of recognition of facial expressions.", "key_idea": "The paper proposes an approach for the facial recognition in sign language, where the Brazilian Sign Language (Libras) is used as a case study. The facial expressions in Libras are coded using the Facial Action Coding System (FACS).", "method": "Two convolutional neural networks - a standard CNN and a hybrid CNN+LSTM are evaluated for AU (action unit) recognition in the Facial Action Coding System. The models are evaluated on a challenging real-world video dataset of facial expressions in Libras.", "outcome": "The proposed system achieved an average f1-score of 0.87, indicating its potential in recognizing Libras' facial expressions.", "future_impact": "The results suggest that this proposed approach could improve the accuracy of Automatic Sign Language Recognition technology by enhancing its ability to recognize critical facial expressions, thereby improving communication between deaf and hearing individuals.", "venue": "ECCV", "year": 2020, "title": "Recognition of Affective and Grammatical Facial Expressions - A Study for Brazilian Sign Language."} +{"pid": "60001c8891e011c8f78fd360", "context": "The increasing use of 3D data in deep learning for tasks such as robotics, drone control, and autonomous driving has emphasized the need for developing robust neural networks. In particular, 3D point clouds, which represent shape information, are a commonly used type of 3D data.", "key_idea": "The authors focus on the perspective of the attacker to understand how neural networks can be exploited and explore two categories of attacks on 3D point clouds: distributional attacks that consist of imperceptible perturbations to the points' distribution, and shape attacks that deform the shape represented by a point cloud.", "method": "The study explores three possible shape attacks for attacking the classification of 3D point clouds and evaluates their effectiveness, including against known defense methods such as point-removal defenses.", "outcome": "The authors find that some shape attacks are able to effectively exploit vulnerabilities in the neural network models, including those with preprocessing steps such as point-removal defenses.", "future_impact": "Understanding of these vulnerabilities can help in the development of more robust neural network models for 3D point cloud data, especially in critical applications, such as robotics, drone control, and autonomous driving.", "venue": "ECCV", "year": 2020, "title": "Adversarial Shape Perturbations on 3D Point Clouds."} +{"pid": "5f0c2ba391e0115455a34b6d", "context": "The advancement of deep learning techniques has enhanced the spatial resolution of hyperspectral image super-resolution (HSI-SR), but developing unsupervised deep networks for this task remains a challenge.", "key_idea": "The authors propose a novel coupled unmixing network, CUCaNet, which enhances the spatial resolution of hyperspectral image super-resolution (HSI-SR) using a multispectral image (MSI). It uses a two-stream convolutional autoencoder framework, decomposing multispectral and hyperspectral data into a spectrally meaningful basis and corresponding coefficients.", "method": "CUCaNet is tested on three widely-used HS-MS datasets, with its performance compared to that of state-of-the-art models in HSI-SR.", "outcome": "Experiments demonstrate the superiority of the CUCaNet in comparison to other state-of-the-art HSI-SR models.", "future_impact": "The proposed model and the availability of the code and datasets online could promote further development and experimentation in the area of HSI-SR.", "venue": "ECCV", "year": 2020, "title": "Cross-Attention in Coupled Unmixing Nets for Unsupervised Hyperspectral Super-Resolution"} +{"pid": "5ff68bbad4150a363cd02e7b", "context": "Photon-efficient imaging allows 3D image capture with as few photons as one per pixel. Measurements of low photon counts, however, can get mixed with heavy background noise, which is difficult for existing computational reconstruction algorithms to handle.", "key_idea": "This paper proposes a non-local neural network for depth reconstruction, designed to exploit long-range correlations in both spatial and temporal dimensions of measurements.", "method": "The proposed network is tested under photon counts as low as 1 photon/pixel and with a signal-to-background ratio as low as 0.01. Moreover, the generalizability of the network is assessed by training it on simulated data and applying it to different real-world imaging systems.", "outcome": "The proposed non-local neural network demonstrated good reconstruction fidelity even under low photon counts and signal-to-background ratio, surpassing the existing state-of-the-art. It was also found to be well generalizable to different real-world imaging systems.", "future_impact": "This work could extend the application scope of photon-efficient imaging in scenarios where there is a strict limit on optical flux.", "venue": "ECCV", "year": 2020, "title": "Photon-Efficient 3D Imaging with A Non-local Neural Network"} +{"pid": "8720ac0f-db63-4069-8361-6b74ace59b50", "context": "Recent methods for visual place recognition based on Convolutional Neural Networks (CNN) are able to handle the place recognition task in RGB-D images. However, these methods require many aligned and annotated intensity and depth images to train joint detectors.", "key_idea": "The authors propose a new approach augmenting the place recognition process using separate intensity and depth networks trained on synthetic data, which requires only a handful of aligned RGB-D frames for competitive place recognition performance.", "method": "The authors propose integrating intensity and depth into a joint robust matching framework for place recognition and evaluate the utility of prediction from each modality.", "outcome": "The new approach shows that a competitive place recognition performance can be achieved with only a handful of aligned RGB-D frames.", "future_impact": "This is the first CNN approach that integrates intensity and depth for place recognition, suggesting a potential shift towards using synthetic data and separate modalities for training place recognition models.", "venue": "ECCV", "year": 2016, "title": "Enhancing Place Recognition Using Joint Intensity - Depth Analysis and Synthetic Data"} +{"pid": "80e10e66-309c-48b7-8b84-64474da85c1e", "context": "Inertial navigation from the data of an inertial measurement unit (IMU) in a smartphone relies largely on heuristic-based approaches and faces issues with bias errors especially with low-frequency linear accelerations.", "key_idea": "The authors propose RIDI, a novel data-driven approach to inertial navigation that learns to estimate trajectories of natural human motions derived solely from an IMU by regressing a velocity vector from the history of linear accelerations and angular velocities and then correcting low-frequency bias.", "method": "Training data with ground truth motion trajectories across multiple human subjects in multiple phone placements (e.g., in a bag or a hand) were collected and used to benchmark the algorithm.", "outcome": "The RIDI algorithm demonstrated improved qualitative and quantitative performance in comparison to existing heuristic-based approaches, showing results comparable to full Visual Inertial Navigation.", "future_impact": "By introducing supervised training for inertial navigation, this paper opens a new line of research in the domain of data-driven inertial navigation. The code and data will be publicly shared for further research.", "venue": "ECCV", "year": 2018, "title": "RIDI: Robust IMU Double Integration"} +{"pid": "60800bb891e011772654f86a", "context": "Monocular 3D object detection is a challenging task in the self-driving and computer vision community. Most previous works use manually annotated 3D box labels, but the annotation process is expensive.", "key_idea": "The authors challenge the belief that precisely and carefully annotated labels are necessary in monocular 3D detection by proving that using roughly disturbed labels, the detector can achieve a very close accuracy compared to the one using ground-truth labels. They propose a simple and effective framework, LiDAR point cloud guided monocular 3D object detection (LPCG), that can either reduce annotation costs or boost detection accuracy without extra annotation costs.", "method": "The LiDAR point cloud guided monocular 3D object detection (LPCG) framework generates pseudo labels from unlabeled LiDAR point clouds. The model was tested on the KITTI and Waymo benchmarks.", "outcome": "LPCG took first place on both monocular 3D and BEV (bird's-eye-view) detection on the KITTI benchmark with a significant margin. On the Waymo benchmark, LPCG using 10% labeled data achieved comparable accuracy to the baseline detector using 100% labeled data.", "future_impact": "LPCG can be applied into any monocular 3D detector to fully use massive unlabeled data in a self-driving system, potentially providing a significant reduction in the cost and time required for manual data annotation.", "venue": "ECCV", "year": 2022, "title": "Lidar Point Cloud Guided Monocular 3D Object Detection."} +{"pid": "62c64f255aee126c0f6cc9b4", "context": "Current transformer frameworks for camouflaged instance segmentation (CIS) are typically two-stage systems, which may require extensive training data and may not converge efficiently.", "key_idea": "The paper introduces OSFormer, the first one-stage transformer framework for CIS, based on two key designs: a location-sensing transformer (LST) for obtaining location labels and instance-aware parameters, and a coarse-to-fine fusion (CFF) mechanism to blend information from the LST encoder and CNN backbone.", "method": "The effectiveness of OSFormer is validated through comparisons with traditional two-stage frameworks, with its efficiency assessed in terms of convergence and data requirements.", "outcome": "OSFormer achieves an average precision (AP) of 41% and demonstrates good convergence efficiency, requiring only 3,040 samples under 60 epochs.", "future_impact": "The design and structure of OSFormer can potentially shape the development of future camouflaged instance segmentation models and approaches, making them more efficient and data-economical.", "venue": "ECCV", "year": 2022, "title": "OSFormer: One-Stage Camouflaged Instance Segmentation with Transformers."} +{"pid": "4c411842-2098-405d-9e11-6ce2727eb9d1", "context": "The increasing volume of data being shared on the web, particularly on social media, poses a threat to user privacy.", "key_idea": "This work explores privacy implications of data sharing on social media by studying how well people can be recognized in varying scenarios where factors such as how many heads of a person are tagged, and if those heads are obfuscated or not, are considered. The study proposes a robust person recognition system that can handle large variations in pose and clothing, and requires few training samples.", "method": "The researchers define various scenarios for a systematic study considering factors like the number of person's heads tagged and the state of their obfuscation. Using these scenarios, they experiment with a robust person recognition system they have proposed.", "outcome": "The experimental results indicate that few images are enough to compromise user privacy, even when obfuscation techniques are applied. Detailed experimental results are provided.", "future_impact": "The authors' findings highlight critical privacy concerns in social media data, contributing to understanding and discussions on privacy in the face of increasing data sharing in the virtual domain.", "venue": "ECCV", "year": 2016, "title": "Faceless Person Recognition: Privacy Implications in Social Media"} +{"pid": "f93e2182-c002-4f39-bec8-a39a325c847b", "context": "Current methods for 3D pose estimation from RGB images struggle with structural errors and inaccuracies.", "key_idea": "The authors propose a deep learning architecture named propagating Long Short-Term Memory networks (p-LSTMs), which incorporate joint interdependency to learn high spatial correlations of human posture from RGB images and reduce structural errors.", "method": "In p-LSTMs, each LSTM is connected sequentially to reconstruct 3D depth from the centroid to edge joints. The first LSTM creates seed joints of 3D pose, which are then reconstructed into whole-body joints through the connected LSTMs.", "outcome": "With p-LSTMs, they achieve higher accuracy (about 11.2% more) than existing state-of-the-art methods on the largest publicly available database. It also significantly reduces structural errors at body edges.", "future_impact": "The use of joint interdependency in 3D pose estimation could lead to improved models in detecting human movement.", "venue": "ECCV", "year": 2018, "title": "Propagating LSTM: 3D Pose Estimation Based on Joint Interdependency"} +{"pid": "8d68d09e-004b-496c-a333-a6a74e656618", "context": "Existing image-matching mathematical models for mammogram registration may be vulnerable to initialization inaccuracies and do not specifically focus on regions of interest or adequately combine intensity and segmentation constraints.", "key_idea": "Authors introduced a new image-matching mathematical model for mammogram registration that combines region matching and segmentation via energy minimization formulation with free boundary conditions, and also contains a new registration constraint.", "method": "The novel model was tested on a simulated mammogram pair and its performances were compared with models without free boundaries.", "outcome": "The experimental results suggest that the new model with free boundary is more robust to initialization inaccuracies as compared to model without free boundaries.", "future_impact": "The new model could improve the robustness of mammogram registration, particularly in terms of its resilience to initialization inaccuracies, which may contribute to advancements in mammography techniques.", "venue": "ECCV", "year": 2002, "title": "A New Image Registration Technique with Free Boundary Constraints: Application to Mammography"} +{"pid": "5f02f0d191e011ee5e0258b0", "context": "Deep neural models often perform best as ensembles of multiple base-level networks. However, ensemble learning for domain adaptive person re-identification (re-ID) has not been explored yet.", "key_idea": "The authors propose a multiple expert brainstorming network (MEB-Net) for domain adaptive person re-ID. It uses a mutual learning strategy where multiple networks with different architectures are pre-trained, and adaptation is accomplished through mutual learning among expert models.", "method": "The proposed method introduces a regularization scheme about the authority of experts to accommodate the heterogeneity of experts learned with different architectures and to enhance the discrimination capability of the adapted re-ID model. Experiments were performed on large-scale datasets, including Market-1501 and DukeMTMC-reID.", "outcome": "Experiments on large-scale datasets show that the MEB-Net outperforms the state-of-the-art models.", "future_impact": "The proposed MEB-Net for domain adaptive person re-ID opens up a promising direction for model ensemble problems in unsupervised conditions.", "venue": "ECCV", "year": 2020, "title": "Multiple Expert Brainstorming for Domain Adaptive Person Re-Identification"} +{"pid": "5ff68a47d4150a363ccc80ce", "context": "The problem of using referential language to identify common objects in real-world 3D scenes has been studied before, but the scarcity and unsuitability of existing 3D-oriented linguistic resources has presented a major challenge, especially when the referred object belongs to a fine-grained object class and is present in a scene with multiple object instances of the same class.", "key_idea": "The authors have developed two large-scale complementary visio-linguistic datasets (Sr3D and Nr3D) and they propose novel neural listeners that can comprehend object-centric natural language and identify the referred object directly in a 3D scene by combining linguistic and geometric information.", "method": "The authors develop their proposed neural listeners using utterances from their created datasets. The listeners are designed to combine linguistic and geometric (3D point clouds) information, and architectures promoting object-to-object communication via graph neural networks are also explored.", "outcome": "Human listeners can recognize the referred object with high accuracy (>86%, 92%) using utterances from the developed datasets. The architectures promoting object-to-object communication via graph neural networks outperformed less context-aware alternatives.", "future_impact": "The paper identifies that improving fine-grained object classification could further enhance the performance of language-assisted 3D object identification.", "venue": "ECCV", "year": 2020, "title": "ReferIt3D: Neural Listeners for Fine-Grained 3D Object Identification in Real-World Scenes"} +{"pid": "62df81d75aee126c0f8734de", "context": "While large-scale datasets have boosted advancements in face generation/editing and other emerging research fields, there is a lack of video datasets with diverse facial attribute annotations, which is important for face-related video research.", "key_idea": "The authors suggest a large-scale, high-quality, and diverse video dataset with comprehensive facial attribute annotations, named the High-Quality Celebrity Video Dataset (CelebV-HQ).", "method": "The authors manually label 35,666 video clips with 83 facial attributes, including appearance, action, and emotion, and conduct a comprehensive analysis in terms of age, ethnicity, brightness stability, motion smoothness, head pose diversity, and data quality.", "outcome": "CelebV-HQ demonstrates both diversity and temporal coherence. Additionally, its versatility and potential have been confirmed on two representative tasks, namely, unconditional video generation and video facial attribute editing.", "future_impact": "The authors anticipate that CelebV-HQ will open up new possibilities and challenges for related research directions.", "venue": "ECCV", "year": 2022, "title": "CelebV-HQ: A Large-Scale Video Facial Attributes Dataset."} +{"pid": "5f1acdc591e011c09af1b6ce", "context": "In the TextVQA task, where the goal is to reason about text in images to answer a question, existing approaches are limited due to their use of fully-connected transformer-based architectures that try to implicitly learn the spatial structure of a scene.", "key_idea": "The authors propose a novel spatially aware self-attention layer for multimodal transformers, wherein each visual entity only looks at neighboring entities defined by a spatial graph and each head in the multi-head self-attention layer focuses on a different subset of relations.", "method": "The authors test their model on the TextVQA and ST-VQA datasets, comparing its performance to existing state-of-the-art methods.", "outcome": "The proposed model improves the absolute accuracy of current state-of-the-art methods on TextVQA by 2.2% overall, and 4.62% on questions that involve spatial reasoning and can be answered correctly using OCR tokens. On ST-VQA, it improves the absolute accuracy by 4.2%.", "future_impact": "The proposed spatially aware self-attention mechanism has the potential to improve visual grounding in multimodal transformers, potentially influencing the development of more advanced transformers in the future.", "venue": "ECCV", "year": 2020, "title": "Spatially Aware Multimodal Transformers for TextVQA"} +{"pid": "80a20988-bb16-41a5-b76a-66f630df7171", "context": "Large-scale instance-level image retrieval involves retrieving specific instances of objects or scenes, and retrieving multiple objects in a test image is a difficult task, especially if the objects are visually similar.", "key_idea": "An efficient approach for per-exemplar multi-label image classification is introduced, which aims to recognize and localize products in retail store images using discriminative random forests, deformable dense pixel matching and genetic algorithm optimization.", "method": "The approach is tested by performing cross-dataset recognition, with training images taken in ideal conditions with one single training image per product label, and the evaluation set taken using a mobile phone in real-life conditions.", "outcome": "The approach achieves promising results in terms of both accuracy and runtime efficiency on 680 annotated images of our dataset, and 885 test images of GroZi-120 dataset.", "future_impact": "The authors provide a large novel dataset of product images, as well as labeling tools for product image search, to motivate further research on multi-label retail products image classification. The dataset of 8350 different product images and the 680 test images from retail stores with complete annotations will also be made available to the wider community.", "venue": "ECCV", "year": 2014, "title": "Recognizing Products: A Per-exemplar Multi-label Image Classification Approach"} +{"pid": "635024b790e50fcafd303939", "context": "State-of-the-art methods for text-to-image synthesis are evaluated with existing metrics which present several issues. These include miscalibration of the Inception Score for single-object cases, misuse of the said score for multi-object cases, overfitting in R-precision and Semantic Object Accuracy metrics, and overlooking of vital factors such as object fidelity, positional alignment, and counting alignment in multi-object cases.", "key_idea": "The authors introduce a framework to evaluate text-to-image synthesis techniques using a combined bag of existing and new metrics. They also propose an improved version of Inception Score (IS*), which uses temperature scaling to calibrate the confidence of the classifier used by IS, and offer a solution to mitigate overfitting issues of R-precision and Semantic Object Accuracy metrics.", "method": "The authors implement and test new metrics such as counting alignment, positional alignment, object-centric IS, and object-centric FID for evaluating multi-object cases. They also create AttnGAN++, a benchmark model for text-to-image synthesis, by stabilizing the training of AttnGAN using spectral normalization.", "outcome": "Benchmarking with the proposed bag of metrics results in a highly consistent ranking among existing methods that is well-aligned with human evaluation. The AttnGAN++ model serves as a new strong baseline for future evaluation.", "future_impact": "The authors release their toolbox, TISE, for promoting fair and consistent evaluation of text-to-image models, which could potentially improve the quality and accuracy of future text-to-image synthesis methods.", "venue": "ECCV", "year": 2022, "title": "TISE: Bag of Metrics for Text-to-Image Synthesis Evaluation."} +{"pid": "62281ae95aee126c0f7aae9e", "context": "Intelligent assistants such as AR glasses/robots aim to assist users in affordance-centric real-world scenarios, but there is still no clear task definition and suitable benchmarks.", "key_idea": "The authors propose a new task called Affordance-centric Question-driven Task Completion for AI assistant which should learn from instructional videos to provide step-by-step guidance in the user's view. A new dataset called 'AssistQ' and a model 'Question-to-Actions (Q2A)' are also introduced to address this new task.", "method": "The Q2A model is developed to address the Affordance-centric Question-driven Task Completion task and validated on the AssistQ dataset which is composed of 531 question-answer samples from 100 newly filmed instructional videos.", "outcome": "The results show that the Question-to-Actions (Q2A) model significantly outperforms several VQA-related baselines in addressing the Affordance-centric Question-driven Task Completion task.", "future_impact": "The introduction of this task and the AssistQ dataset is expected to contribute to the development of future Egocentric AI Assistant.", "venue": "ECCV", "year": 2022, "title": "AssistQ: Affordance-Centric Question-Driven Task Completion for Egocentric Assistant."} +{"pid": "d826efdf-c64c-4d8c-836f-9a4753ffb4d0", "context": "Most existing single image super resolution (SR) techniques primarily focus on enforcing image priors or synthesizing details, often neglecting the final color assignment, which can lead to color aberrations in the upsampled image.", "key_idea": "This paper presents a new procedure for handling color in SR, that is based on image colorization and back-projection to perform color assignment guided by the super-resolution luminance channel.", "method": "The authors outline a procedure based on image colorization and back-projection to perform color assignment guided by the SR luminance channel.", "outcome": "The proposed procedure produces better results both quantitatively and qualitatively than existing approaches, according to the authors.", "future_impact": "This approach is generic and can be incorporated into any existing single image super resolution techniques, potentially improving their outcomes.", "venue": "ECCV", "year": 2010, "title": "Colorization for single image super resolution"} +{"pid": "7df966c1-2bca-49b7-9367-149a8d307d1e", "context": "Structure from Motion (SfM) is an approach to recover 3D structure from 2D image sequences, but it is inherently plagued by ambiguities and uncertainties.", "key_idea": "The main idea of the paper is to quantify the ambiguities in SfM by examining the eigenvectors associated with null or small eigenvalues of the Hessian matrix, and predicting how these ambiguities impact the accuracy of the reconstructed shape.", "method": "The authors propose a method based on eigenvectors and Hessian matrix for orthographic and perspective cameras in SfM to investigate the nature of structure reconstruction ambiguities.", "outcome": "The findings reveal that the bas-relief ambiguity is significant for orthographic cameras even with many images, unless a large degree of rotation is present. Moreover, it is found that three or more frames and substantial rotation are mandatory for metrically accurate reconstruction in perspective cameras.", "future_impact": "This analysis of ambiguities in SfM could guide researchers in enhancing image reconstruction techniques by considering these inherent uncertainties.", "venue": "ECCV", "year": 1996, "title": "Shape Ambiguities in Structure from Motion"} +{"pid": "e1b538f1-90da-4fdf-b8f6-58bfa7d6c695", "context": "In human-robot interaction, emotion perception and interpretation plays a key role. But, it's challenging to model affective bodily expressions especially in naturalistic settings due to the variety of expressive patterns and difficulty in acquiring reliable data.", "key_idea": "This paper investigates the dimensional emotion prediction problem in a child-robot interaction scenario and proposes a method that considers emotion elicitation, data acquisition, 3D skeletal representation, feature design and machine learning algorithms.", "method": "The study involves the elicitation of emotions, gathering of data, the use of 3D skeletal representation, designing features, and implementing machine learning algorithms in a child-robot interaction setting.", "outcome": "The proposed method shows good predictive performance on the variation trends of emotional dimensions, especially the arousal dimension.", "future_impact": "The findings of this study could enhance the quality and naturalness in human-robot interaction by enabling robots to perceive and interpret human emotions more effectively.", "venue": "ECCV", "year": 2014, "title": "Real-Time Emotion Recognition from Natural Bodily Expressions in Child-Robot Interaction"} +{"pid": "6391890590e50fcafd2b42c2", "context": "Bipartite graphs are widely used to model relationships between different entities. Identifying dense subgraphs in such models can be significant but few models consider similarity between vertices from the same side.", "key_idea": "The authors propose a new concept known as 'similar-bicliques' in which all vertices from a designated side in a bipartite graph are similar to each other. The aim is to enumerate all similar-bicliques.", "method": "An efficient backtracking algorithm is proposed to directly enumerate maximal similar-bicliques, powered by vertex reduction and optimization techniques. A novel index structure and efficient index construction algorithms are developed. Algorithms are also provided for dynamic graph updates, and index construction algorithms are parallelized to exploit multiple CPU cores.", "outcome": "Experiments on 17 bipartite graphs demonstrate the effectiveness and efficiency of the proposed model and algorithms.", "future_impact": "The proposed model and algorithms could significantly enhance the finding of dense subgraphs in bipartite graphs, thereby improving the efficiency in modeling relationships between different entities.", "venue": "VLDB", "year": 2024, "title": "Identifying similar-bicliques in bipartite graphs"} +{"pid": "5b40b728-85c4-42ee-ae81-ab8fce4130f2", "context": "Data in many application domains need a system for storing and querying faceted hierarchies.", "key_idea": "The authors have developed a general faceted domain model and a query language specifically designed for hierarchically classified data.", "method": "The authors demonstrate the applicability and utility of their system by using it on two real archaeological datasets containing thousands of artifacts.", "outcome": "The system can provide global access to large datasets in a queryable format, demonstrating its capability to serve as a valuable tool for data analysis and research.", "future_impact": "This system, as a sharable and evolvable resource, can potentially be applied and useful to many other application domains beyond archaeology.", "venue": "VLDB", "year": 2005, "title": "A faceted query engine applied to archaeology"} +{"pid": "26c0481a-152f-4546-b0b2-d8d0a4a8a558", "context": "The existing literature differentiates between self-contained and embedded database languages. However, there is no clear comparison and examination of their differences.", "key_idea": "This paper revolves around the comparison of self-contained and embedded database languages, and observes that the major differences originate from the fact that many details are prespecified in self-contained languages.", "method": "By analysing the key features and specifications of these language types, the authors make arguments and provide several examples illustrating the application of these improved embedded query languages.", "outcome": "The study finds that the prespecifications in self-contained languages can be transferred to embedded languages, achieving better compactness without compromising flexibility, given the possibility of overriding the prespecifications.", "future_impact": "This paper opens up avenues for improving the compactness and flexibility of embedded database languages by integrating features from self-contained languages, potentially influencing the advancement of existing database languages.", "venue": "VLDB", "year": 1987, "title": "A Comparison of Self-Contained and Embedded Database Languages"} +{"pid": "49226566-29d0-4ad1-8465-b2ed363299bf", "context": "XML and XQuery semantics are sensitive to the output order, but current pattern-tree based algebraic approaches for XML evaluation lack a universally accepted technique to guarantee both a correct output order and efficient alternative plans. Additionally, duplicate elimination raises concerns similar to order maintenance as it is costly and a single collection type does not provide the flexibility to optimize this effectively.", "key_idea": "The authors propose a method by using hybrid collections of trees that can be sets, sequences, or intermediate types. Each collection is coupled with an Ordering Specification for sorting mechanisms and a Duplicate Specification for managing duplicate elements in the collections, allowing a query plan to maintain no order, partial order, or full order.", "method": "The authors extend an existing bulk tree algebra, TLC, to use Ordering and Duplicate specifications to produce correctly ordered results. They suggest some optimizations and experimentally demonstrate the increase in performance due to these changes.", "outcome": "The adapted TLC using the authors' Ordering and Duplicate specifications demonstrated an increase in performance in the experimental demonstrations.", "future_impact": "The proposed method can potentially streamline order maintenance and duplicate elimination in XML and XQuery semantics, suggesting efficient alternative query plan development and execution.", "venue": "VLDB", "year": 2005, "title": "Pattern tree algebras: sets or sequences?"} +{"pid": "5a493605-3b96-4701-ae55-2433c29a5387", "context": "Previous algorithms for garbage collecting object-oriented databases in a client-server environment require locks on data and callbacks to clients, which can impact performance and useability. Most of these algorithms lack fault tolerance and extensive logging.", "key_idea": "The authors propose an efficient server-based algorithm for garbage collection that is incremental, can run concurrently with client transactions, does not hold any locks on data, does not require callbacks to clients, and is fault tolerant with minimal logging.", "method": "The algorithm has been designed to be integrated into existing Object-Oriented Database (OODB) and it was implemented in the EXODUS storage manager for evaluation.", "outcome": "The abstract does not provide specific results or outcomes of the implementation.", "future_impact": "Potential integration into existing OODB systems, which could lead to better client-server performance optimizations such as client caching and flexible management of client buffers in the field of Object-Oriented Databases.", "venue": "VLDB", "year": 1995, "title": "Efficient Incremental Garbage Collection for Client-Server Object Database Systems"} +{"pid": "dc774a3d-137c-4409-938a-7bb71c587a48", "context": "The lack of a method to design scalable and representative moving object data (MOD) and sets of queries for benchmarking spatio-temporal DBMS motivated the authors to contribute in this area.", "key_idea": "The authors offer a method to create scalable and representative moving object data in addition to two sets of queries, using the existing Secondo DBMS, for benchmarking spatio-temporal DBMS. The benchmark's foundation is a simulation scenario involving the positions of vehicles observed within the street network of Berlin over an indefinite period of time.", "method": "The authors used the existing Secondo DBMS to create the benchmark data and showed how to achieve more natural movement generation patterns, disturb vehicles' positions to create noisy data and how to extend the benchmark to accommodate new classes of queries using nearest neighbour queries as a preliminary extension.", "outcome": "The authors provided well-defined data sets and query libraries for experimental evaluations and presented initial benchmarking results for the Secondo DBMS.", "future_impact": "The BerlinMOD benchmark is expected to simplify experimental repeatability, emphasize the development of complete systems, reveal weaknesses in existing systems thereby facilitating further research, and make it possible to compare various representations of the same moving objects.", "venue": "VLDB", "year": 2009, "title": "BerlinMOD: a benchmark for moving object databases"} +{"pid": "01ff0345-6576-417b-99ce-6b426ee122fa", "context": "The advent of high throughput systems to determine the three-dimensional (3-D) structure of proteins creates a growing need for systems to automatically store, maintain, and explore the vast structural databases.", "key_idea": "The authors have developed the Capri/MR system which automatically indexes and searches a protein database by three-dimensional shape, structural and/or physicochemical properties. This system creates a compact, rotation and translation invariant index (or signature) for diverse protein structure representations.", "method": "The authors demonstrate the speed and accuracy of the Capri/MR system via exhaustive search against the entire database, utilizing the compact signatures to rapidly find protein structures similar in 3-D shape and/or two-dimensional (2-D) properties.", "outcome": "When tested against the Protein Data Bank, Capri/MR was able to accurately group protein structures into the correct families with very high precision and recall. The system also demonstrates the ability to dynamically process new protein structures as they become available.", "future_impact": "The Capri/MR system could be significant in structural proteomics, protein evolution and mutation studies, drug design (in particular for studying docking problems and the computer-aided design of non-toxic drugs).", "venue": "VLDB", "year": 2008, "title": "Capri/MR: exploring protein databases from a structural and physicochemical point of view"} +{"pid": "496ce002-d509-4f92-bbe4-feb59d382003", "context": "Reverse k nearest neighbors (RkNN) queries in facility-user scenarios return every user for which the facility is one of k closest facilities. Existing techniques to solve RkNN queries generally use a prune-and-verify approach with two main modes: region-based and half-space pruning, with the latter considered superior.", "key_idea": "The authors observe that the commonly neglected regions-based pruning has certain positives which were overlooked, and propose a new regions-based pruning algorithm called 'Slice'. Additionally, they extend this idea to solve spatial reverse top-k (SRTk) queries using a novel regions-based pruning algorithm in the Slice framework.", "method": "The authors extend half-space-based pruning to answer SRTk queries and then implement a new algorithm called 'Slice', a regions-based pruning mechanism. The varied strategies are tested through an extensive experimental study on synthetic and real datasets.", "outcome": "The results show that the proposed 'Slice' pruning strategy is significantly more efficient than all existing RkNN and SRTk algorithms.", "future_impact": "The method proposed could be utilized in future research to improve the efficiency of spatial related queries like RkNN and SRTk.", "venue": "VLDB", "year": 2017, "title": "Reverse k nearest neighbors queries and spatial reverse top-k queries"} +{"pid": "c49ce789-0f33-48b4-b9ae-1a1ed4f821d1", "context": "Existing algorithms for computing core and truss decompositions in graph analysis are incremental in nature and require global information at each step, which hinders scalable parallelization and approximations. Moreover, the densest regions are not revealed until the end of the computation.", "key_idea": "This paper offers a framework of local algorithms to obtain the core, truss, and nucleus decompositions based on the iterative computation of the $h$-indices of neighbor vertex degrees, a concept which was initially introduced by Lu et al.", "method": "The proposed framework and algorithms are implemented in shared-memory and then tested on real-world networks to verify their efficiency, scalability, and effectiveness.", "outcome": "The shared-memory implementation confirmed the proposed local algorithms' efficiency, scalability, and effectiveness on real-world networks.", "future_impact": "The proposed local, parallel algorithms have high scalability and offer potential for approximations, allowing time and quality trade-offs in future network analysis applications.", "venue": "VLDB", "year": 2018, "title": "Local algorithms for hierarchical dense subgraph discovery"} +{"pid": "6c0208af-7935-433b-9265-f85f38abb85d", "context": "The requirements for a main memory data storage model are both compactness and efficient processing for all database operations.", "key_idea": "The authors propose DBGraph, a storage model that represents the entire database in a graph-based structure, fully exploiting the direct-access capability of main memory systems.", "method": "Operations like Selection, Join and Transitive closure over base or temporary relations are tested by DBGraph traversal without tuple comparison and move. DBGraph's decomposability is also examined by loading only the useful subset of the database from disk without format conversion. Database queries are processed by either set-oriented or pipelined mode depending on the graph's traversal.", "outcome": "Analysis shows DBGraph has good storage occupancy and excellent performance for both update and retrieval operations.", "future_impact": "Complex database queries can be processed effectively by either set-oriented or pipelined mode depending on the way the graph is traversed, which can enhance the efficiency of database management systems in the future.", "venue": "VLDB", "year": 1990, "title": "Efficient main memory data management using the DBgraph storage model"} +{"pid": "b64cb122-ad51-4550-a338-bd56efbd8deb", "context": "The problem of horizontally partitioning a dynamic relation across a large number of disks/nodes by the use of range partitioning is common and important in large-scale parallel databases and peer-to-peer (P2P) systems. As tuples are inserted and deleted, the partitions may need to be adjusted, and data moved, in order to achieve storage balance across the participant disks/nodes.", "key_idea": "The authors propose efficient, asymptotically optimal algorithms that ensure storage balance at all times, even against an adversarial insertion and deletion of tuples. They combine these algorithms with distributed routing structures to architect a P2P system.", "method": "The authors architect a P2P system that supports efficient range queries while simultaneously guaranteeing storage balance, using their proposed algorithms combined with distributed routing structures.", "outcome": "The authors developed a P2P system that supports efficient range queries and guarantees storage balance using their proposed algorithms and distributed routing structures.", "future_impact": "The proposed system and algorithms could potentially be used for partitioning in large-scale parallel databases and P2P systems, improving the efficiency and optimality of these systems.", "venue": "VLDB", "year": 2004, "title": "Online balancing of range-partitioned data with applications to peer-to-peer systems"} +{"pid": "98fc7300-117a-4d64-bf59-a99a19ddb5fd", "context": "Managing and integrating different types of clinical data in ovarian cancer research remains a challenging task.", "key_idea": "The authors propose to build an i2b2-based integrated data repository to centralize data management for ovarian cancer clinical research. The process involves collection of multiple types of heterogeneous clinical data and extensive data normalization through reusing standard codes and mapping local codes to standard vocabularies.", "method": "The authors use extraction, transformation, and loading scripts to load the data into an i2b2 instance. They then perform analytic practices to evaluate the system by addressing common clinical research needs such as cohort query and identification, clinical data-based hypothesis-testing, and exploratory data mining.", "outcome": "The authors identified multiple expectations of the system through analytic practices, including utility for cohort query and identification, clinical data-based hypothesis-testing, and exploratory data-mining. Some outstanding issues were also identified and discussed.", "future_impact": "Lessons learned from this study regarding the evaluation and enhancement of a cancer-specific data repository can benefit future generic cancer research.", "venue": "VLDB", "year": 2016, "title": "Building an i2b2-Based Integrated Data Repository for Cancer Research: A Case Study of Ovarian Cancer Registry"} +{"pid": "6391890690e50fcafd2b4388", "context": "In National University Hospital (NUH) in Singapore, the available modalities evolve or change intermittently while conducting healthcare analytics to support clinical decision-making on a daily basis. Existing models fail to adapt to such dynamic changes.", "key_idea": "The paper presents DyHealth, a modular, multimodal and interpretable framework that enables dynamic healthcare analytics in clinical practice. This framework can adapt to dynamic modality changes via exponential increasing/decreasing mechanisms supporting modality 'hot-plug' and also provides a fine-grained interpretation of results with a novel modality-based attention mechanism.", "method": "DyHealth was validated by conducting pilot evaluation on the patients' EMR data from NUH and on two public EMR datasets. Different modalities are processed within their respective data modules adhering to DyHealth interface definition, and then the information extracted from them is integrated in Multimodal Fusion Module of DyHealth.", "outcome": "The pilot evaluation of DyHealth on the patients' EMR data from NUH showcased superior performance. Experimental results on two public EMR datasets confirm the effectiveness, flexibility, and extensibility of DyHealth for multimodal and interpretable healthcare analytics.", "future_impact": "DyHealth shows promise to roll out for hospital-wide deployment due to its dynamic adaptability, which can significantly enhance the effectiveness of healthcare analytics and clinical decision-making.", "venue": "VLDB", "year": 2022, "title": "DyHealth: Making Neural Networks Dynamic for Effective Healthcare Analytics."} +{"pid": "cdf5e699-f8d4-4de7-9360-bd9f0e769bc8", "context": "Information integration, which provides a competitive advantage to businesses, is fundamental to on demand computing. It is a strategic area of investment by software companies today aiming to provide a unified view of data regardless of format, location, and access interfaces.", "key_idea": "The paper describes the motivation for integrating information for on demand computing and explains its requirements.", "method": "The study's scope is explained through usage scenarios which illustrate the value of integrating information for on demand computing.", "outcome": "The paper points to the existing challenges and complexity of achieving a comprehensive information integration solution, with emphasis on the need for more research, engineering, and development work to realize the full vision of information integration.", "future_impact": "It is anticipated that software companies will continue to heavily invest in pursuing the information integration vision, heralding an extensive amount of future research and development in the field.", "venue": "VLDB", "year": 2003, "title": "Integrating information for on demand computing"} +{"pid": "e3ebbd62-762f-47d7-a7e0-f95b27a22c03", "context": "The existing studies on the performance of concurrency control algorithms in literature have been either simulation-based or analytical. While Bayer's Time Interval concurrency control method has been proposed, no actual implementation had been reported until now.", "key_idea": "The authors implement and evaluate Bayer\u2019s Time Interval concurrency control method for the first time, comparing it with a conventional timestamp method.", "method": "The implementation and comparison were performed on the Eden experimental local area network, and the early and late serialization schemes within the Time Interval method were further explored and compared.", "outcome": "The Time Interval method showed higher throughput, caused one-third as many distributed transaction aborts, and required very little additional overhead compared to timestamps. Early and late serialization with Time Intervals showed comparable performance over a range of read/write ratios and multiprocessing levels.", "future_impact": "The demonstrated superiority and efficiency of the Time Interval method over the conventional timestamp method may inform future implementations of concurrency control methods to consider adopting this approach.", "venue": "VLDB", "year": 1987, "title": "Measured Performance of Time Interval Concurrency Control Techniques"} +{"pid": "a8fa8ebc-f3b2-42f2-adda-ea1ded36a638", "context": "Handling memory resident persistent objects in virtual memory in the context of the E programming language has presented a challenge, with existing software architectures not providing optimal performance.", "key_idea": "The authors propose an efficient method for accessing memory resident persistent objects which adopts pointer swizzling to enhance performance. The method is characterized by on-demand copying from the buffer pool and updates propagation back to the object manager via a single write operation at the end of each transaction.", "method": "A performance study was conducted comparing this method with several alternative software architectures including ObjectStore V1.2, a commercially available OODBMS. These comparisons looked at the tradeoffs between software vs. memory-mapped support for pointer swizzling and quantified the effects of pointer swizzling on overall performance.", "outcome": "The experiment results showed that in many situations the software approach to pointer swizzling can outperform the memory-mapped approach.", "future_impact": "The method has the potential to change how memory resident persistent objects in virtual memory are traditionally handled, promoting a software approach over current memory-mapped solutions for pointer swizzling.", "venue": "VLDB", "year": 1992, "title": "A Performance Study of Alternative Object Faulting and Pointer Swizzling Strategies"} +{"pid": "aa023033-5194-4799-b588-70407eb33668", "context": "Processing and querying continuous data streams (comprising data-item deletions as well as insertions) with limited memory resources is of interest but challenging. Estimating the cardinality of set expressions defined over several update streams is a fundamental problem. However, earlier studies have only focused on the special case of insert-only streams and specific operators.", "key_idea": "The authors propose the first space-efficient algorithmic solution for estimating the cardinality of set expressions over general update streams. This is based on a novel hash-based synopsis data structure, the '2-level hash sketch'.", "method": "The authors demonstrate how the 2-level hash sketch synopses can provide low-error, high-confidence estimates for the cardinality of set expressions (including operators such as set union, intersection, and difference) over continuous update streams. The effectiveness of their approach is verified through an empirical study.", "outcome": "The proposed solution allows estimates for the cardinality of set expressions over continuous update streams using only space that is significantly sublinear in the sizes of the streaming input (multi-)sets. The proposed solutions never require rescanning or resampling of past stream items, regardless of the number of deletions in the stream. The empirical study verifies the effectiveness of this approach.", "future_impact": "The techniques presented offer an optimized, time-efficient stream synopsis that provides strong accuracy-space guarantees while requiring only guaranteed logarithmic maintenance time per update. This makes the methods applicable for truly rapid-rate data streams and potentially impactful in domains dealing with continuous data streams.", "venue": "VLDB", "year": 2004, "title": "Tracking set-expression cardinalities over continuous update streams"} +{"pid": "9b09b7f6-a32c-484e-a2f2-79d6101f2e8f", "context": "Labeling schemes, used in network routing, programming, classification, and knowledge representation systems, have also interested the XML community. However, there's a trade-off between storage capacities and runtime performance.", "key_idea": "The authors provide a taxonomy of numbering schemes for XML based on the types of supported queries, encoding technique, and the scheme's robustness properties per updates. The focus is on prefix-based schemes.", "method": "The authors describe some of the numbering techniques proposed for XML and give a qualitative comparison of the existing numbering schemes, discussing their advantages and drawbacks.", "outcome": "Comparison of the numerous numbering schemes shows the trade-offs between storage requirements and performance, and provides insights into their relative advantages and disadvantages.", "future_impact": "The paper suggests new research directions that likely could benefit from numbering scheme techniques.", "venue": "VLDB", "year": 2008, "title": "Prefix based numbering schemes for XML: techniques, applications and performances"} +{"pid": "1bc48841-57ba-4d4e-80f5-2738350c9c58", "context": "Wikipedia's multi-lingual articles provide a resource for cross-language information retrieval and machine translation. However, leveraging such data for querying across different languages remains a challenge due to the lack of mappings between attributes from infoboxes across different languages.", "key_idea": "The authors propose a method for identifying mappings between attributes from infoboxes across different languages, which is automated, scalable, and does not require syntactic similarity between attribute names.", "method": "The method was evaluated through extensive experiments using a corpus consisting of Portuguese, Vietnamese, and English Wikipedia pages. It was compared with state-of-the-art techniques.", "outcome": "The results reveal that the proposed method achieves high precision and recall, and outperforms state-of-the-art techniques. The derived multilingual mappings lead to significant improvements in answer quality and coverage for structured queries over Wikipedia content.", "future_impact": "The proposed technique, by facilitating structured querying across languages, can open new opportunities for querying structured Wikipedia content and enable answers that straddle different languages.", "venue": "VLDB", "year": 2011, "title": "Multilingual schema matching for Wikipedia infoboxes"} +{"pid": "ac30795c-ff49-49fc-88fd-3812c5c655ad", "context": "Decision-support applications often need to execute natural queries, typically group queries. The current state of database systems does not offer optimized plans specifically for group queries where data could be processed in a group-by-group fashion.", "key_idea": "The authors define a particular class of queries called group queries that can be executed in a group-by-group fashion. For these queries, they propose a syntactic criterion to identify such queries and demonstrate an evaluation technique to optimize their execution.", "method": "The authors provide proof for their syntactic criterion to identify group queries and show an equivalent formulation that satisfies their criterion. They demonstrate two methods to identify group query components within complex queries. They perform comparative performance analysis for group queries expressed in standard SQL, using both a commercial database system and an optimized plan on top of the same commercial system.", "outcome": "The authors have proven that every group query has an equivalent formulation that satisfies their syntactic criterion. Their performance results show that their optimized plan offers potential performance improvements compare to a commercial database system.", "future_impact": "The authors propose their techniques and evaluations as a basis for further investigations, implying the potential advancements in the field of database systems, specifically in improving performance of executing group queries.", "venue": "VLDB", "year": 1997, "title": "Groupwise Processing of Relational Queries"} +{"pid": "d2ed2a7c-0de4-4256-ab9c-b14c0f75456b", "context": "With the dominance of the relational model along with the requirements of modern applications and veracity of data, handling NULLs in relational databases is a significant challenge. In SQL database systems, primary key columns are NOT NULL by default, however, NULL columns may occur in unique constraints providing a different function than primary keys.", "key_idea": "The authors propose the concept of possible and certain keys in the context of relational databases dealing with NULLs. Possible keys, which are consistent with an SQL's unique constraint, provide a semantic interpretation for their syntactic definition in SQL standard. Certain keys expand primary keys to include NULL columns and represent a necessary and sufficient condition for unique identification of tuples.", "method": "The authors characterize, provide an axiom for, and describe discovery approaches for possible and certain keys. The paper also presents an indexing scheme for upholding certain keys. The paper evaluates these concepts through experiments on real-world databases.", "outcome": "The experiments confirm that certain keys with NULLs do occur in real-world databases, and the related computational problems can be efficiently solved.", "future_impact": "Certain keys are semantically solid and capable of maintaining data quality in terms of Coddu0027s entity integrity rule while dealing with realistic needs of modern applications, such as handling larger volumes of incomplete data from diverse formats.", "venue": "VLDB", "year": 2015, "title": "Possible and certain SQL keys"} +{"pid": "6e36dc9d-6096-4039-bea8-fdfbac9b73c9", "context": "Top-k query processing is an important part of ranked retrieval and operates on index lists for a query's elementary conditions and aggregate scores for result candidates. Existing threshold algorithms terminate index scans early based on final score bounds and perform sequential and random disk accesses for sorted index scans resulting in scheduling concerns. Previous literature has addressed these scheduling issues, but only for each type of access separately.", "key_idea": "This paper proposes integrating both types of scheduling issues into a single system, and introduces new principled scheduling methods based on a Knapsack-related optimization for sequential accesses and a cost model for random accesses.", "method": "The proposed methods are tested on three different datasets: TREC Terabyte, HTTP server logs, and IMDB, and compared to the best methods previously known.", "outcome": "The performance experiments showed that the proposed methods achieved significant performance gains compared to the best previously known methods.", "future_impact": "The improved top-k query processing methods can effectively boost ranked retrieval in various applications, including text and data integration and distributed aggregation of network logs and sensor data.", "venue": "VLDB", "year": 2006, "title": "IO-Top-k: index-access optimized top-k query processing"} +{"pid": "0684764f-b334-41d5-8b5a-786a93bcaead", "context": "Relational database systems often require the processing of predicates in crucial aspects such as satisfiability testing (as in processing predicate locks), and preprocessing predicates in queries to reduce the number of database operations when the query is answered.", "key_idea": "The authors examine predicates consisting of conjunctions of comparisons, particularly focusing on conjunctions of =, <, and >= comparisons where a variable can be compared with a constant or with another variable. The study explores the complexity introduced when unequal comparisons between variables are allowed.", "method": "The authors provide algorithms for satisfiability, equivalence, and minimizing the number of comparisons in a predicate. They also analyze scenarios when unequal comparisons between variables are allowed, and discuss the complexity in satisfiability, equivalence, and minimization.", "outcome": "The authors present efficient algorithms for satisfiability, equivalence, and minimizing the number of comparisons in a predicate. However, they show that the inclusion of unequal comparisons between variables makes satisfiability, equivalence, and minimization NP-hard.", "future_impact": "Understanding the complexity introduced by unequal comparisons between variables in conjunctions of comparisons can inform the development of more efficient algorithms for relational database systems.", "venue": "VLDB", "year": 1980, "title": "Processing conjunctive predicates and queries"} +{"pid": "1b0cfd97-97d5-433a-acc5-50e7db82c968", "context": "Adaptive query processing has become a focus in recent work, especially in the emerging fields of data integration and data streams.", "key_idea": "The authors focus on providing an overview of the work in adaptive query processing and aim to identify common themes and explain where it's most appropriately used.", "method": "The study presents an overview of adaptive query processing by examining, comparing, and discussing various implementations within the field, including the exploration of different query plans.", "outcome": "The paper serves as a comprehensive discussion on why adaptive query processing is needed, how it is implemented, and where it is most effectively applied.", "future_impact": "The authors identify several open research problems in adaptive query processing, potentially guiding future studies in this area.", "venue": "VLDB", "year": 2007, "title": "Adaptive query processing: why, how, when, what next?"} +{"pid": "045c8c1a-b850-40b5-80e8-8ebb0fa42aa5", "context": "The research takes place in the context of data cleaning, for which there is a need for benchmarking algorithms. However, a challenge arises in creating errors in clean databases for this purpose.", "key_idea": "The paper proposes a new approach to introduce errors into clean databases for benchmarking data-cleaning algorithms, aiming to provide a high level of user control over the process and to scale to large databases.", "method": "The problem is tackled by developing a correct and efficient greedy algorithm that relies on non-trivial optimizations, including a newly discovered symmetry property of data quality constraints.", "outcome": "The paper reveals that the error-generation problem is NP-complete, which makes it surprisingly challenging. The proposed greedy algorithm, although it sacrifices completeness, is shown to succeed under reasonable assumptions.", "future_impact": "The proposed scalable solution and the highlighted trade-off between control over the process and scalability can serve as significant technical contributions in the field of data cleaning and error generation for algorithm benchmarking.", "venue": "VLDB", "year": 2015, "title": "Messing up with BART: error generation for evaluating data-cleaning algorithms"} +{"pid": "6391890890e50fcafd2b4636", "context": "Performance comparisons across different systems rely on well-known benchmarks such as TPC-C and YCSB. However, interpreting and comparing performance numbers remains a challenge due to tunability of benchmark parameters, system features, and hardware settings, which affects system behaviors.", "key_idea": "This study seeks to understand performance sensitivity to settings by reproducing 11 previous works evaluated under TPC-C and YCSB benchmarks and investigating performance under a wider range of settings.", "method": "The study reproduces 11 works evaluated under the TPC-C and YCSB benchmarks and measures their performance under an extensive range of settings. The changes in performance numbers are investigated to understand the factors behind such differences.", "outcome": "The study finds that varying experiment settings leads to different performance results, and it emphasizes the need for a discussion on how to address the issue of system behavior variance impacted by benchmark parameter settings, system features, and hardware settings.", "future_impact": "The study suggests concrete steps to improve the state of art despite not providing a complete solution, which may encourage future research in this direction for refining the process of comparing system performance.", "venue": "VLDB", "year": 2022, "title": "A Study of Database Performance Sensitivity to Experiment Settings."} +{"pid": "0e97fbe3-8602-4da5-85c4-44bc919eed3d", "context": "The prevalence of social media and GPS-enabled devices has led to the generation of a massive amount of geo-textual data in a stream fashion. This poses challenges to existing systems in processing and delivering the most relevant geo-textual messages in real-time.", "key_idea": "The paper presents a novel real-time top-k monitoring system, called Skype (Top-k Spatial-keyword Publish/Subscribe), that continuously maintains the top-k most relevant geo-textual messages for a large number of spatial-keyword subscriptions over a sliding window of streaming data.", "method": "The authors propose a new indexing structure for maintaining top-k results for large-scale subscriptions and develop a cost-based k-skyband technique to reduce the number of re-evaluations triggered by message expiration. They verify these techniques through extensive experiments.", "outcome": "The proposed Skype system and the associated techniques demonstrate great efficiency and effectiveness in maintaining top-k geo-textual results and reducing re-evaluation costs, as shown in experimental results.", "future_impact": "The novel system and techniques proposed in this study could be useful for a variety of applications such as location-based recommendation and information dissemination, potentially benefiting those utilizing large-scale geo-textual data.", "venue": "VLDB", "year": 2016, "title": "Skype: top-k spatial-keyword publish/subscribe over sliding window"} +{"pid": "23d8567f-9e63-46c2-b0ed-dc91cd617287", "context": "The selection and restriction operation of database management systems involves linearly searching large files, a process which may be slow and efficiency-reducing.", "key_idea": "The authors propose a special processor for performing the selection and restriction operation of data base management systems by linearly searching large files at high speed.", "method": "The paper describes the reasoning, architecture, and structure of the proposed search processor.", "outcome": "The authors provide a status report and future plans for a research project, presumably about the proposed processor.", "future_impact": "The paper does not explicitly mention the future impact of this research. However, it implies that if successful, the proposed processor could significantly improve the selection and restriction operations in database management systems.", "venue": "VLDB", "year": 1978, "title": "A search processor for data base management systems"} +{"pid": "5c6e41d3-442d-41d0-9389-7a6ed3a139f8", "context": "Histogram queries in differential privacy often lack precision because of the obfuscation introduced to maintain privacy.", "key_idea": "The authors propose to enhance accuracy by choosing a set of queries to evaluate and exploiting consistency constraints that should hold over the noisy output. In a post-processing phase, they compute the consistent input most likely to have produced the noisy output.", "method": "The authors test the strategy by applying it to estimating the degree sequence of a graph and computing a histogram that can support arbitrary range queries.", "outcome": "The final output is not only differentially-private and consistent, but it is also much more accurate, both theoretically and experimentally.", "future_impact": "This strategy could significantly improve the precision of histogram queries made under differential privacy constraints.", "venue": "VLDB", "year": 2010, "title": "Boosting the accuracy of differentially private histograms through consistency"} +{"pid": "9b81c74d-5c7c-4279-af9f-c7696e94f493", "context": "The search for occurrences of a twig pattern query (TPQ) in an XML document is a core task in all XML database query languages. The generalized twig pattern (GTP) extends the TPQ model to include additional semantics. Preorder filtering holistic algorithms such as TwigStack and postorder filtering holistic algorithms like $$\text{ Twig}^2$$ Stack are existing approaches to TPQ processing.", "key_idea": "The authors propose GTPStack, a holistic algorithm that leverages a combination of preorder and postorder filtering to process a GTP with a linear worst-case I/O complexity relative to the GTP result size.", "method": "The authors introduce a new perspective on algorithm optimality, demonstrating its relation to not just query classes but also XML document characteristics, through an experimental study.", "outcome": "The article proves GTPStack's optimality for any GTP given a specific XML document and under certain circumstances, it outperforms other holistic approaches.", "future_impact": "The extension of general knowledge about the type of queries for which holistic algorithms are optimal, as well as a novel perspective on algorithm optimality, may facilitate advancements in the field.", "venue": "VLDB", "year": 2013, "title": "Optimal and efficient generalized twig pattern processing: a combination of preorder and postorder filterings"} +{"pid": "fe32984e-ee27-42a6-8278-22ee70cedf8c", "context": "Recent work in data streams has focused on systems where new data is continuously processed and streamed in real time, but there is a need for processing data that arrives prior to query submission or during a period of disconnection in emerging applications.", "key_idea": "PSoup is developed to process ad hoc and continuous queries by treating data and queries symmetrically. It allows new queries to be applied to old data and new data to be applied to old queries, and supports intermittent connectivity by separating query results computation from their delivery.", "method": "Authors build PSoup using adaptive query-processing techniques from the Telegraph project at UC Berkeley and conduct experiments to validate the effectiveness of the approach.", "outcome": "Experiments demonstrate the effectiveness of PSoup for processing both old and new data with old and new queries.", "future_impact": "PSoup has the potential to influence the approach towards stream data querying by addressing the challenge of intermittent connectivity and the need for processing data that arrives prior to query submission.", "venue": "VLDB", "year": 2003, "title": "PSoup: a system for streaming queries over streaming data"} +{"pid": "ef2a67f5-a361-43da-b168-a4967ab55856", "context": "Graph data management tools are rapidly evolving, but current graph generators provide limited support for workload generation or are limited to fixed use-cases.", "key_idea": "The authors propose gMark, the first domain- and query language-independent framework for synthetic graph and query workload generation, allowing for fine-grained control via user-defined schemas and supporting expressive graph query languages.", "method": "A demonstration is carried out, showcasing the highly tunable generation of graphs and queries through various user-defined schemas and targeted selectivities, and the variety of supported practical graph query languages. A performance comparison of four state-of-the-art graph database engines is also executed.", "outcome": "The demonstration of gMark showcases the flexible and tunable generation of graphs and queries. The performance comparison helps understand the strengths of the graph database engines and where they could be further extended.", "future_impact": "The insights obtained from the demonstration and performance comparison could influence the design of future extensions of graph database engines.", "venue": "VLDB", "year": 2016, "title": "Generating flexible workloads for graph databases"} +{"pid": "ac3ef110-ae9c-4172-858e-d5438fabf535", "context": "Natural language interfaces to databases are not in common use today as they pose challenges due to their difficulty in use and their high cost in building and maintenance.", "key_idea": "The authors propose a new kind of natural language interface that combines a menu-based interaction technique with a semantic grammar-driven natural language system. It leverages a core grammar with parameters supplied by the data dictionary and then automatically generates semantic grammars covering a selected part of the user's data. This combination is intended to ease use and reduce cost.", "method": "The authors designed a core grammar with parameters supplied by the data dictionary and automatically generated semantic grammars covering a selected subpart of the user's data.", "outcome": "The authors present a functional overview of the new kind of natural language interface that would solve 'ease-of-use' and 'creation and maintenance' issues broadly experienced in current interfaces.", "future_impact": "Automatically generated natural language interfaces offer attractive features like grouping semantically related tables together, modeling a user\u2019s access rights, and supporting joins paths in a database. This advances the user's view of a database and may shape future database interaction designs.", "venue": "VLDB", "year": 1983, "title": "Building Usable Menu-Based Natural Language Interfaces To Databases"} +{"pid": "96f1cf5f-b182-4ae0-9db6-94feb79ef22e", "context": "The emerging standards for the electronic representation of clinical data, such as the Clinical Document Architecture (CDA), are very complex making it challenging to extract and utilize the valuable information contained in these documents.", "key_idea": "The authors propose HIWAS, a research prototype tool, that structured summaries of a collection of XML documents and enables users to find information within these complex XML documents.", "method": "The authors demonstrate the functionality of HIWAS by showing how a user can create a target model that contains only the needed information in a simplified representation that is queryable and compatible with existing relational business intelligence technology.", "outcome": "The HIWAS tool simplifies complex XML documents to be easily digestible with conventional tools.", "future_impact": "The authors anticipate that HIWAS can lower a key barrier to meaningful use of aggregated clinical data, providing a stepping stone for more complex analyses in the future.", "venue": "VLDB", "year": 2011, "title": "HIWAS: enabling technology for analysis of clinical data in XML documents"} +{"pid": "1f82bf3f-92a6-4337-8f8a-09364fe42a4a", "context": "Prior to this study, structured data stored in files were not typically managed using standard database technology.", "key_idea": "The authors introduce the concept of a structuring schema, which consists of a grammar annotated with database programs, that allows a file to be viewed as a database structure and queried and updated accordingly.", "method": "The authors explore the use of almost standard database optimization techniques to answer queries without needing to construct the entire database. They also dig deep into the process of propagating updates to the file which are specified on the database view of this file.", "outcome": "The authors found that propagating updates to the file is not feasible in general, though they present methods that can efficiently propagate updates under certain locality conditions on the structuring schemas.", "future_impact": "This novel way of viewing and manipulating data stored in files as if they are database structures could lead to more efficient and effective data management.", "venue": "VLDB", "year": 1998, "title": "A logical view of structured files"} +{"pid": "8af6a422-2d11-4f73-bddc-38dc4810ad1a", "context": "Current generation of search technology does not fully utilize structures and semantics of documents generated by application programs, leading often to inaccurate search results.", "key_idea": "The paper presents a new method to extend inverted file index structures with predicates which allows the encoding of circumstances under which certain keywords in a document become visible to the user, thus helping in better exploitation of document structures.", "method": "The authors provide a framework to express the semantics of structures in documents and algorithms to construct enhanced, predicate-based indexes. They also show how keyword and phrase queries can be processed efficiently on such enhanced indexes.", "outcome": "The proposed approach showed superior retrieval performance in terms of both recall and precision with tolerable space overhead and query running time.", "future_impact": "This paper provides a new framework for desktop search technologies and could guide future research in encoding the semantics of structures into search indexes.", "venue": "VLDB", "year": 2010, "title": "Predicate-based indexing for desktop search"} +{"pid": "ee726e71-5ef8-450a-a7f8-28aa9528f178", "context": "Query-by-Example is a query language designed for non-programmers to interact with a relational database. Previously, this was introduced with the assumption that the database was predefined and available to the user.", "key_idea": "The authors propose using Query-by-Example to not only query the database but also define it, including data descriptions and declarations of various integrity constraints.", "method": "The authors demonstrate the functionality of their proposal in two parts: firstly, by illustrating how Query-by-Example can be used to define a database; and secondly, by showcasing how one can use preliminary interrogations to retrieve a subset of the database of interest, which is especially useful in large databases.", "outcome": "The authors were able to show that Query-by-Example can be used effectively for defining databases and retrieving subsets of databases.", "future_impact": "The proposed approach could provide users with a simplified, unified interface for various functions, making tasks like scanning large tables for relevant information less daunting.", "venue": "VLDB", "year": 1975, "title": "Query-by-example: the invocation and definition of tables and forms"} +{"pid": "e946c3ee-550f-457e-a50b-954fc522414a", "context": "Conventional approaches to execution of database queries on general-purpose multiprocessors maximize system throughput using inter-query parallelism with a fixed number of processors and standard uniprocessor optimization techniques.", "key_idea": "The paper proposes a method to increase performance by using intra-query parallelism and minimizing overall resource requirements, specifically through coordination of the order in which data pages are read into memory and page joins are assigned to available processors.", "method": "The paper presents a scheduling strategy based on join indices, along with a heuristic for estimating the required number of processors to complete join execution in minimal time. The approaches' resource requirements are validated through lower and upper bound proofs and simulated performance.", "outcome": "Simulation results indicate that the proposed techniques effectively utilize processors and buffer requirements.", "future_impact": "The methodology and techniques proposed in this paper, especially the heuristic for estimating the number of processors required, could lead to significant performance improvements and resource optimizations in the execution of database queries on multiprocessor systems.", "venue": "VLDB", "year": 1989, "title": "Effective resource utilization for multiprocessor join execution"} +{"pid": "e8404eee-e940-4700-a054-4b3a4d9d7804", "context": "Nearest neighbor search is a crucial computational task in database systems and data analysis which often represents a major computational bottleneck. Numerous research has been conducted on data structures and algorithms for accelerating this task. The prevalent approach is to run brute force search on a highly parallel system using graphics hardware, given the recent advancements in this area.", "key_idea": "This paper explores a unique approach by proposing a new data structure, a Random Ball Cover, specifically designed to make efficient use of parallel systems such as graphics cards.", "method": "The authors construct a novel data structure and compare its performance to a GPU-based brute force nearest neighbor search approach.", "outcome": "Despite potential challenges and doubts, the authors demonstrate that their proposed data structure, the Random Ball Cover, yields significant speed improvements when compared to the GPU-based brute force approach.", "future_impact": "This work presents a novel approach to handling the complexities of graphics hardware, which could lead to the development of more efficient algorithms and data structures for other computational tasks.", "venue": "VLDB", "year": 2010, "title": "A Nearest Neighbor Data Structure for Graphics Hardware"} +{"pid": "6b61fddd-6392-4ed9-ad78-4ed2714ea064", "context": "Role-Based Access Control (RBAC) models have been available since the early 1990s, but there is no well-understood methodology for using RBAC with an arbitrary application program.", "key_idea": "The authors seek to bridge this gap by discussing tradeoffs between the ANSI RBAC model and the Role Graph Model, and by enumerating different versions of each.", "method": "The authors discuss various alternatives for bridging between an RBAC model and an arbitrary program, and provide an example of applying one of these alternatives.", "outcome": "The paper elaborates different versions of the ANSI RBAC model and the Role Graph Model, and showcases how to apply one of the bridging alternatives.", "future_impact": "The methodology provided in this paper may pave the way for better understanding and integration of RBAC models with arbitrary application programs.", "venue": "VLDB", "year": 2008, "title": "A Methodology for Bridging between RBAC and an Arbitrary Application Program"} +{"pid": "b84191cf-df09-46f0-8995-ca32f8bbb369", "context": "The current price gap between disk and flash memory drives has led to a consideration of flash memory drives supplementing rather than replacing disk drives in applications dealing with large scale data.", "key_idea": "The authors propose a new low-overhead caching strategy called FaCE that uses flash memory as an extension to the DRAM buffer, aiming at improving transaction throughput and shortening recovery time from system failure.", "method": "The authors devise two novel algorithms for flash cache management, Multi-Version FIFO replacement and Group Second Chance, and implement FaCE in the PostgreSQL open source database server. To demonstrate its effectiveness, they apply it to TPC-C benchmarks.", "outcome": "The FaCE strategy demonstrates that a small flash memory drive used as a caching device delivers higher throughput than a large flash memory drive storing entire database tables. This is achieved through flash write optimization and disk access reduction.", "future_impact": "FaCE takes advantage of flash memory's non-volatility to expand the scope of a persistent database to include the data pages stored in the flash cache and fully support database recovery.", "venue": "VLDB", "year": 2012, "title": "Flash-based extended cache for higher throughput and faster recovery"} +{"pid": "9eeb4bcb-f715-4306-8df6-40a83a4812d8", "context": "Locality-sensitive hashing (LSH) has become a popular approach to solve the all-pairs similarity search problem, which involves finding all pairs of objects in a given collection that have a similarity greater than a user-specified threshold. However, LSH methods are typically only utilized in the initial phase of similarity search for efficient candidate generation.", "key_idea": "The authors introduce BayesLSH, a principled Bayesian algorithm that uses LSH in the subsequent phase of similarity search, specifically performing candidate pruning and similarity estimation.", "method": "Development of two models, BayesLSH which estimates similarities probabilistically and BayesLSH-Lite which calculates similarities exactly. They test these models on data accompanied with two state-of-the-art candidate generation algorithms, AllPairs and LSH.", "outcome": "Their algorithms showed the ability to quickly prune away most of the false-positive candidate pairs, resulting in significant speedups over baseline methods, typically between 2x and 20x. BayesLSH also provided probabilistic guarantees on the quality of the output, in terms of both accuracy and recall.", "future_impact": "The quality of BayesLSH outputs can be easily tuned and it doesn't require manual setting of hash numbers for similarity estimation, which could shape the future development of similarity search algorithms and could potentially negate the necessity for manual adjustments in similarity search tasks.", "venue": "VLDB", "year": 2012, "title": "Bayesian locality sensitive hashing for fast similarity search"} +{"pid": "612d300f-d790-4656-b9ba-6279df31fad9", "context": "Social Networking sites and Web portals have opened up their APIs to external developers, enabling them to design applications that can be integrated with the platform and shared with other users. However, these APIs target developers with programming expertise and are not accessible to those without a programming or database background.", "key_idea": "The researchers developed AppForge, a WYSIWYG application development platform that allows users to graphically specify the components of webpages which then generates the corresponding database schema and application logic automatically.", "method": "AppForge has been used to create prototype versions of a variety of applications, and the researchers also conducted a preliminary user study to identify and rectify usability issues.", "outcome": "The system successfully created prototypes of various applications, including an event planning system, a recruiting system, an item trading system, and an online course management system.", "future_impact": "The system allows a larger class of users to create new applications, expanding the pool of application developers beyond those with programming expertise and database knowledge.", "venue": "VLDB", "year": 2008, "title": "WYSIWYG development of data driven web applications"} +{"pid": "a1746166-fbfd-425f-9457-d10b6fe65221", "context": "The problem of data sharing between autonomous data sources in an environment where constraints cannot be placed on the shared contents of sources is considered. The challenge lies in retrieving data from different sources that are associated when queried against the schema of a single source.", "key_idea": "The proposed solution relies on the use of mapping tables that define how data from different sources are associated. This allows translation, through mapping tables, of local query into a set of queries that are executed against the associated sources.", "method": "The study proposes sound translations (which only retrieve correct answers) and complete translations (which retrieve all correct answers, and no incorrect answers), with algorithms to compute such translations. The developed solutions are implemented and tested experimentally.", "outcome": "The paper does not report specific outcomes or results from the implemented experiments.", "future_impact": "The paper does not explicitly mention the future impact of this method on related research.", "venue": "VLDB", "year": 2004, "title": "Data sharing through query translation in autonomous sources"} +{"pid": "50ed7b62-f7ae-4526-983b-f508a2f1ccc3", "context": "k-core decomposition is often used in large-scale network analysis and in many real-world applications, networks change over time. However, current methods do not support efficient incremental algorithms for streaming graph data.", "key_idea": "The paper proposes the first incremental k-core decomposition algorithms for streaming graph data, which locate a small subgraph that is guaranteed to contain the list of vertices whose maximum k-core values have to be updated.", "method": "The performance and efficiency of these algorithms are assessed in comparison to non-incremental alternatives on different types of real and synthetic graphs, at different scales.", "outcome": "The results show a significant reduction in run-time compared to non-incremental alternatives. Speedups reaching a million times were observed for a graph with 16 million vertices.", "future_impact": "The incremental k-core decomposition algorithms, due to their high efficiency and speed, could potentially improve the large-scale network analysis and related tasks, fostering advancements in real-world applications like protein function prediction and community detection.", "venue": "VLDB", "year": 2013, "title": "Streaming algorithms for k-core decomposition"} +{"pid": "eb337fa3-bab7-4d19-bbbd-295dbec86c2c", "context": "Businesses with branch offices can greatly benefit from distributed Enterprise Resource Planning (ERP) systems where local ERP systems use resources managed by other local systems. Such a system could theoretically be built with a distributed DBMS but such systems have been impractical due to low performance and local autonomy.", "key_idea": "This paper puts forth a solution to design a distributed ERP system by utilizing databases with relaxed ACID (Atomicity, Consistency, Isolation, Durability) properties, emphasizing the use of short duration locking to optimize performance, local autonomy, and availability.", "method": "The author has collaborated with a major ERP software company to analyze how a distributed version of their ERP system can be designed using relaxed ACID properties.", "outcome": "The paper describes how it's possible to build a distributed ERP system using databases with relaxed ACID properties. The techniques are general and can be used to integrate most package software running on separate computers.", "future_impact": "The techniques detailed in the paper are general and could potentially be used to integrate most package software running on separate computers, expanding the potential integration and efficiency of ERP systems.", "venue": "VLDB", "year": 2005, "title": "Architecture for distributed ERP systems"} +{"pid": "5aff8de4-7527-4b5c-b41d-ed63579bf7f5", "context": "The challenge addressed is portable and flexible privacy preserving access rights that permit access to a large collection of digital goods while maintaining privacy of customer activity and access rights.", "key_idea": "The authors propose two schemes to address the problem of privacy-preserving access rights. These schemes allow for a negligible amount of false positives due to constrained space resources.", "method": "The methodology involves the presentation and comparison of two proposed schemes. The first scheme does not place any constraints on the amount of space available on the limited-capacity storage device, and searches for the best representation that meets the requirements. The second scheme has modest requirements on the storage space available, but guarantees a low rate of false positives.", "outcome": "The first scheme works without any constraints on the available space and aims to find the best fitting solution. The second scheme, requiring a modest amount of space, guarantees a low rate of false positives.", "future_impact": "Possible future implications could include protection against the sharing of false positives information by dishonest users, providing users with transaction untraceability and unlinkability, and ensuring the schemes are forward compatible.", "venue": "VLDB", "year": 2006, "title": "Succinct representation of flexible and privacy-preserving access rights"} +{"pid": "628d2c065aee126c0f556d7b", "context": "Real-world networks are often evolving and can be modeled as temporal graphs to capture the time when relationships are established (temporal information) and the structure of the graph (topological information). However, there is a need for a more nuanced understanding of these two types of information.", "key_idea": "The authors propose a new metric named T-cohesiveness that evaluates the cohesiveness of a temporal subgraph from the time and topology dimensions jointly.", "method": "Two techniques using T-cohesiveness are proposed: 1) T-cohesiveness evolution tracking, which traces the evolution of the T-cohesiveness of a subgraph, and 2) combo searching which finds out all the subgraphs that contain the query vertex and have T-cohesiveness larger than a given threshold. A pruning strategy is proposed to improve the efficiency of combo searching.", "outcome": "Experimental results confirm the efficiency of the proposed time-topology analysis methods and the pruning strategy.", "future_impact": "The work could facilitate more effective time-topology analysis on temporal graphs, enabling the extraction of additional valuable data from networks evolving over time.", "venue": "VLDB", "year": 2021, "title": "Time-Topology Analysis"} +{"pid": "42e9e9b8-1034-4a3d-88b1-a228c508895d", "context": "Previous research on information extraction and constructing RDF knowledge bases creates structured, non-redundant sources of information by extracting and storing information from documents into an ontology. The ontology is then used to answer queries, making the original documents dispensable. However, users often have doubts about the validity of the facts and ontology subgraphs derived solely from the ontology without the context of the original document.", "key_idea": "The authors propose ROXXI, a system that operates on top of an existing knowledge base and revives the abandoned witness documents, tracing ontological facts back to the documents they were extracted from.", "method": "ROXXI offers interfaces for expert users (SPARQL) and non-experts (ontology browser), providing a ranked list of documents associated with a content snippet highlighting the queried facts in context. It will be demonstrated at the demonstration site to showcase its approach and benefits.", "outcome": "The authors demonstrate that ROXXI is able to trace back ontological facts to the documents they were extracted from, providing context for validation. It's not clear in the abstract about the quantitative outcome of the experiments.", "future_impact": "The demonstration of ROXXI aims to illustrate the benefits and potential of reviving abandoned documents from information extraction approaches and could shape future research towards the importance of keeping the witness documents alongside the structured ontology-based knowledge base.", "venue": "VLDB", "year": 2010, "title": "ROXXI: Reviving witness dOcuments to eXplore eXtracted Information"} +{"pid": "dcdbd802-f1f6-4874-913e-eec6ece3445c", "context": "Over the past decade, middleware has been developed for integrating and automating enterprise business processes, with a recent renewed focus on business process coordination due to growth in e-commerce and blurred enterprise boundaries.", "key_idea": "This paper provides a comprehensive historical perspective on technologies for intra- and inter-enterprise business processes, including process-based coordination and event/rule-based coordination.", "method": "The authors give a review of the current state of art and highlight open research issues in the field of business process coordination, including a discussion of corresponding products and standards activities.", "outcome": "The paper offers comprehensive summaries of advanced transaction models related to business processes and business process intelligence, exposing several open research issues.", "future_impact": "This review and discussion of open issues in the realm of business process coordination may stimulate further investigations and advancements in the field with potential impacts on e-commerce and enterprise operations.", "venue": "VLDB", "year": 2001, "title": "Business Process Coordination: State of the Art, Trends, and Open Issues"} +{"pid": "94c38eaa-02da-407e-9169-d588ea468ae9", "context": "Physical clustering of a database table on multiple dimensions is a powerful technique that offers significant performance benefits in many systems. However, the selection of clustering keys in single-dimensional and multidimensional relational databases can be combinatorially complex for significant dimensionality (3 or more).", "key_idea": "A model is proposed for automating the selection of clustering keys in databases using a cell/block storage structure for multidimensional clustering (MDC), based on what-if query cost modeling, data sampling, and a search algorithm.", "method": "Through data sampling and what-if query cost modeling, the model evaluates a large constellation of possible combinations. The effectiveness of the model is measured by comparing its design recommendations with those made by human experts.", "outcome": "The model effectively trades the benefits of potential combinations of clustering keys against data sparsity and performance, and it accurately selects the granularity at which dimensions should be used for clustering. The results from experiments demonstrate that the model's design recommendations are of comparable quality to those made by human experts.", "future_impact": "The model has been implemented in the IBM DB2 UDB for Linux, UNIX and Windows Version 8.2 release, suggesting potential widespread use in industry.", "venue": "VLDB", "year": 2004, "title": "Automated design of multidimensional clustering tables for relational databases"} +{"pid": "d615fe52-7fef-48ea-aaf4-cc5ac5e0eb75", "context": "String similarity join, a vital operation in data integration and cleansing, has seen the formulation of over ten algorithms in the past twenty years. However, these algorithms lack a comprehensive comparison under the same experimental framework, making it difficult for practitioners to decide the most suitable algorithm for varying scenarios.", "key_idea": "This study proposes a thorough survey of existing string similarity join algorithms, classifies them based on their primary techniques, and aims to compare them using extensive experiments on a wide range of real-world datasets.", "method": "The authors conducted extensive experiments on various real-world datasets, presenting different characteristics, to compare the classified string similarity join algorithms.", "outcome": "Through the experiments, the paper presents comprehensive findings, disclosing the strengths and weakness of existing similarity join algorithms.", "future_impact": "The insights derived from the research can guide practitioners in selecting the most appropriate algorithm for different scenarios, aiding effective data integration and cleansing.", "venue": "VLDB", "year": 2014, "title": "String similarity joins: an experimental evaluation"} +{"pid": "f62748da-7822-422a-aaba-5aa3c89cc185", "context": "Parallelism techniques in object-oriented database management systems (OODBMS) can enhance performance, but their effectiveness varies across different tasks such as graph traversals, which are key functions in such systems.", "key_idea": "The authors present ParSets, an approach designed to exploit parallelism in the SHORE OODBMS, focusing specifically on the graph traversal portion of the OO7 OODBMS benchmark.", "method": "The implementation of ParSets is tested in a parallelized SHORE setup running graph traversals of the OO7 OODBMS benchmark on a cluster of commodity workstations connected by a standard ethernet.", "outcome": "The implementation of ParSets showed excellent speedup and scaleup for some OO7 traversals, while only marginal speedup and scaleup occurred for others.", "future_impact": "The results and characteristics of these traversals can shed light on when the ParSet approach to parallelism can and cannot be applied to speed up an application in OODBMS context.", "venue": "VLDB", "year": 1996, "title": "Parallelizing OODBMS traversals: a performance evaluation"} +{"pid": "47307efb-58a0-47e9-b4c0-f91e4b425a7f", "context": "A significant amount of research has been conducted on the efficient and scalable processing of subgraph search queries on large networks, with the assumption that users have precise knowledge of the topological relationships between nodes in a query graph to formulate a connected query.", "key_idea": "This paper presents a novel graph querying paradigm called partial topology-based network search, proposing a query processing framework named PANDA that can process partial topology query (ptq), a type of disconnected query graph with multiple connected query components, on a single machine. PANDA allows users to formulate queries without requiring extensive knowledge of the complete topology of a query graph.", "method": "The authors propose an exact algorithm, SEN-PANDA, and an approximate algorithm, PO-PANDA, to generate top-k matches of a ptq. To further improve performance, a subgraph simulation-based optimization technique is also presented. The effectiveness of these methods is evaluated through experiments with real-life networks containing millions of nodes.", "outcome": "The experiments demonstrated that the proposed algorithms, SEN-PANDA and PO-PANDA, performed superior to multiple baseline techniques when used on real-life networks with millions of nodes.", "future_impact": "The authors have not explicitly mentioned the future impact of this work.", "venue": "VLDB", "year": 2017, "title": "PANDA: toward partial topology-based search on large networks in a single machine"} +{"pid": "6857f60b-5f99-40b7-9250-7a83a4d2a5a7", "context": "SimRank is an important measure for identifying similar vertices in a graph and is crucial for many data analysis applications. However, it is difficult to efficiently perform similarity search based on SimRank over large dynamic graphs due to high time and space costs.", "key_idea": "This paper proposes a novel two-stage random-walk sampling framework (TSF) specifically for SimRank-based similarity search. This new framework provides efficient solutions to the limitations of SimRank, specifically in the context of large dynamic graphs.", "method": "Random graphs are sampled and indexed using a novel method in the preprocessing stage, allowing future updates to accommodate graph modifications. The query stage allows for fast vertex similarity search by pruning unqualified vertices. The authors also implement the ability to store graphs on disk to improve scalability. TSF's performance is evaluated through extensive experimenting on dynamic billion-edge graphs.", "outcome": "The TSF framework was able to handle dynamic billion-edge graphs with high performance, providing an efficient solution for SimRank-based similarity search.", "future_impact": "The proposed TSF framework could serve as an effective solution for conducting efficient SimRank-based similarity searches over large, dynamic graphs, significantly impacting the realm of data analysis applications.", "venue": "VLDB", "year": 2015, "title": "An efficient similarity search framework for SimRank over large dynamic graphs"} +{"pid": "11b7c3bb-bafd-4289-b77a-3a5dc37d41f5", "context": "In a temporal Object-Oriented Database (OODB), an OID Index (OIDX) maps OID to the physical location of the object and indexes the object versions. This OIDX needs to be updated whenever an object is updated, unlike in a non-temporal OODB where it's only updated when an object is created. This repeated updating process has been identified as a potential bottleneck.", "key_idea": "The paper presents the Persistent Cache (PCache), a new approach intended to reduce the index update and lookup costs in temporal OODBs.", "method": "The authors establish a cost model for the PCache in order to demonstrate the potential cost reductions achievable with its use.", "outcome": "The cost model reveals that the use of PCache can reduce the average access cost to only a fraction of the cost compared to when it is not used.", "future_impact": "Although the primary context of the paper is OID indexing in a temporal OODB, the PCache can also be applied to general secondary indexing, and could be particularly beneficial for applications where updates are non-clustered.", "venue": "VLDB", "year": 1999, "title": "The Persistent Cache: Improving OID Indexing in Temporal Object-Oriented Database Systems"} +{"pid": "4408d3dc-a2c5-4b23-bfea-a3f56c2957fb", "context": "Traditional summarization techniques have been unable to provide a small, query-able summary of a dataset suitable for interactive data exploration.", "key_idea": "The authors propose a novel probabilistic approach that uses the Principle of Maximum Entropy to generate a probabilistic representation of data that enables approximate query answers.", "method": "The authors establish a theoretical framework and formulation, optimize the implementation for faster preprocessing time and better query accuracy, and evaluate it using a 5 GB dataset of flights within the US and a 210 GB astronomy particle simulation dataset.", "outcome": "The proposed technique successfully answers queries faster than sampling, with no more error on average, and it better differentiates between rare and nonexistent values. However, the current work only supports linear queries.", "future_impact": "Given the positive results, this probabilistic summarization technique could be further developed to support more complex query types beyond linear ones.", "venue": "VLDB", "year": 2017, "title": "Probabilistic database summarization for interactive data exploration"} +{"pid": "2a0d0da7-9a8b-4339-bded-400ecf2a9518", "context": "The Northern Ireland Housing Executive (NIHE) is developing a large relational database system for property management with a high expected transaction processing rate of approximately 37000 transactions per day from about 250 online users, and performance is critical to its success.", "key_idea": "The research studied the effect of the Ingres Search Accelerator on the transaction processing efficiency of the system.", "method": "The research assessed the performance enhancement brought about by SCAFS (XL\u2019s current version of the well-known Content Addressable File Store, the heart of the Ingres Search Accelerator) for different file organisations.", "outcome": "The study provided recommendations on how the performance of SCAFS can be improved by tuning certain parameters and provided guidelines as to when the Ingres Query Optimizer 'decides' to use SCAFS for different file organizations.", "future_impact": "The assessment of the Ingres Search Accelerator's impact on transaction processing efficiency, along with provided guidelines and parameter tuning recommendations, may influence future optimization strategies in large-scale database systems.", "venue": "VLDB", "year": 1994, "title": "An Empirical Performance Study of the Ingres Search Accelerator for a Large Property Management Database System"} +{"pid": "e8e3984b-3678-4c5a-ab1c-61f014d6883f", "context": "The need for better integration of data management and knowledge management has become more concrete and urgent with the increasing use of ontology in various applications. However, manipulating knowledge along with relational data in Database Management Systems (DBMSs) is challenging.", "key_idea": "A unified framework for managing data and domain knowledge is proposed in this paper. It offers a virtual view that unifies the data, the domain knowledge, and the knowledge that can be inferred from the data using the domain knowledge.", "method": "The approach leverages XML support in hybrid relational-XML DBMS like Microsoft SQL Server and IBM DB2 9 PureXML. A query rewriting mechanism is provided so that queries on the virtual view can be automatically transformed to components that execute on the hybrid relational-XML engine.", "outcome": "The method allows users to query the data and the knowledge in a seamlessly integrated manner, with logical to physical data modeling differences handled transparently for the user via an automated query rewriting mechanism.", "future_impact": "This unified approach could potentially ease the integration of data management and knowledge management within the database community and has the potential for further exploitation in systems leveraging hybrid relational-XML DBMS.", "venue": "VLDB", "year": 2007, "title": "Unifying data and domain knowledge using virtual views"} +{"pid": "cb302732-bfd1-4ef7-825c-b133c4171724", "context": "When testing applications that utilize production data containing sensitive information, the data must first be anonymized, a task made difficult by the constraints often present in the data that need to be applied to the anonymized version as well.", "key_idea": "The authors put forth a new strategy for the anonymization of constrained production data, drawing on a concept known as constraint satisfaction problems.", "method": "The approach suggested involves diving the constrained anonymization issue into independent problems that can be resolved as constraint satisfaction problems (CSPs) and solve iteratively by applying local modifications to the production data.", "outcome": "Simulations performed on synthetic production databases outlined how the proposed method was feasible and effective.", "future_impact": "The approach can support a wide range of mandatory integrity constraints as well as constraints that retain the resemblance of the anonymized data to the original production data, providing potential for further research and application in various fields.", "venue": "VLDB", "year": 2010, "title": "Constrained anonymization of production data: a constraint satisfaction problem approach"} +{"pid": "5ff883d991e011c832674446", "context": "Knowledge bases of semi-structured tables expressed as text lack an intuitive constraint satisfaction solver and interpreted language that can easily express complex compositional 'inference patterns', and support inference and explanation construction in downstream tasks.", "key_idea": "CoSaTa is introduced, an intuitive constraint satisfaction solver and interpreted language for semi-structured tables expressed as text. It aids in expressing complex compositional 'inference patterns' and includes advanced declarative features while operating over multiple representations of text.", "method": "The authors developed a hybrid imperative/declarative interpreted language for expressing simple models through minimally-specified simulations grounded in constraint patterns, thereby implementing CoSaTa.", "outcome": "CoSaTa was successfully created and includes features for expressing simple models, supporting inference and explanation construction, and operating over varying text representations. The solver and interpreter have been released as open source.", "future_impact": "CoSaTa can bridge the gap between question answering, question explanation, and model simulation, and its release as open source allows other researchers to utilize and build upon it in the future.", "venue": "EMNLP", "year": 2020, "title": "CoSaTa - A Constraint Satisfaction Solver and Interpreted Language for Semi-Structured Tables of Sentences."} +{"pid": "ac9cb887-6cd6-4753-b651-d6401972c644", "context": "The paper addresses the challenge in robust grammar engineering where there is a need for automatic acquisition of Multiword Expressions (MWEs). Previous attempts at MWE detection rely on statistical properties of component words, but the efficiency of different statistical measures is not thoroughly investigated.", "key_idea": "The key idea is to evaluate different statistical measures (mutual information (MI), chi-squared, and permutation entropy (PE)) for MWE detection, and to investigate the influence of corpus size and quality.", "method": "The authors explore these ideas by comparing the aforementioned statistical measures for MWE detection, and then examining the influence of the size and quality of different corpora, specifically the British National Corpus (BNC) and the web search engines Google and Yahoo.", "outcome": "The study concludes that at least two statistical measures, MI and PE, can differentiate MWEs from non-MWEs. In terms of language usage, the web generated corpora, in spite of their lack of control and balance, are similar to more carefully built corpora such as the BNC.", "future_impact": "The study suggests that a more compositional approach to grammar/lexicon automated extension, when automatically adding extracted MWEs to existing linguistic resources, could lead to qualitative improvements in grammar engineering.", "venue": "EMNLP", "year": 2007, "title": "Validation and Evaluation of Automatically Acquired Multiword Expressions for Grammar Engineering"} +{"pid": "618cfba991e011333c93276b", "context": "While consistency identification has been successful in open-domain dialogues for preventing inconsistent response generation, this aspect has received little attention in task-oriented dialogues.", "key_idea": "The authors propose the concept that the problem of consistency is more critical in task-oriented domains and introduce CI-ToD, a novel dataset for Consistency Identification in Task-oriented Dialog systems, supplementing it with fine-grained labels which judge system response contradictions.", "method": "State-of-the-art methods are applied to the CI-ToD dataset to evaluate their consistency identification ability. Exhaustive experiments and qualitative analysis are conducted to comprehend the key challenges.", "outcome": "Empirical results show that state-of-the-art methods only achieved a 51.3% success rate, significantly lower than the human performance of 93.2%, indicating much potential for improvements.", "future_impact": "The dataset, findings and qualitative analysis provide guidance for future research in improving consistency identification in task-oriented dialogues. Additionally, making the datasets and models publicly available may drive forward advancements in this field.", "venue": "EMNLP", "year": 2021, "title": "Don't be Contradicted with Anything! CI-ToD - Towards Benchmarking Consistency for Task-oriented Dialogue System."} +{"pid": "619799f191e011c82237313a", "context": "NLP systems often face difficulties with noisy, non-standard, and domain-specific corpora. The task of lexical normalization aims to standardize these kinds of corpora but there's a lack of suitable tools to acquire high-quality annotated data to support deep learning based approaches.", "key_idea": "The authors propose LexiClean, the first open-source web-based annotation tool for multi-task lexical normalization, which supports simultaneous in situ token-level modification and annotation that can be applied across the whole corpus.", "method": "The functionality and effectiveness of LexiClean were tested through a case study on two sets of noisy corpora derived from the specialized domain of industrial mining.", "outcome": "The case study shows that LexiClean allows for rapid and efficient development of high-quality parallel corpora.", "future_impact": "LexiClean has potential for future use in generating high-quality annotated data needed for deep learning based approaches in lexical normalization, especially beneficial for processing noisy, non-standard, and domain specific corpora.", "venue": "EMNLP", "year": 2021, "title": "LexiClean: An annotation tool for rapid multi-task lexical normalisation"} +{"pid": "65252a8e939a5f40826b7360", "context": "Text is everywhere in our visual world, providing crucial information in a variety of forms. Current machine understanding of visually-situated text is often dependent on optical character recognition (OCR) techniques and domain-specific pretraining and finetuning paradigms, which are resource-intensive.", "key_idea": "The authors propose UReader, an approach for universal OCR-free, visually-situated language understanding using a Multimodal Large Language Model (MLLM), which leverages the shallow text recognition ability of the MLLM and only requires finetuning of 1.2% parameters, reducing training cost.", "method": "UReader is jointly finetuned on a wide range of visually-situated Language Understanding tasks via a unified instruction format. Two auxiliary tasks, text reading and key points generation, are used to improve visual text and semantic understanding. A shape-adaptive cropping module is also designed before the encoder-decoder architecture of MLLM to manage high-resolution images.", "outcome": "Without downstream finetuning, UReader achieves state-of-the-art OCR-free performance in 8 out of 10 visually-situated language understanding tasks, across 5 domains: documents, tables, charts, natural images, and webpage screenshots.", "future_impact": "The proposed unified model, UReader, could serve as a universal solution for understanding visually-situated text across various formats, potentially revolutionizing how systems process and understand information. The authors plan to release the instruction-tuning datasets and codes which would benefit future researchers.", "venue": "EMNLP", "year": 2023, "title": "UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model."} +{"pid": "6579f104939a5f4082003a5b", "context": "Deep semantic retrieval has become increasingly important in online E-commerce applications, which predominantly rely on distinguishing positive and negative items via margin or softmax losses. However, such methods are sensitive to hyper-parameters (margin and temperature), which influence the similarity of negative pairs and the distribution of items in metric space. Existing methods of learning parameters through trainable/statistical means have proven to be unsuitable for retrieval scenarios due to the diversity and unknown nature of queries.", "key_idea": "The authors propose an adaptive metric learning method that allows for hyper-parameter learning based on batch similarity, without the need for fixed or extra-trainable hyper-parameters. The method also includes a symmetric metric learning component to prevent model collapse.", "method": "The authors test the adaptive metric learning method on a real-world dataset and compare its performance against previous methods. The method is also deployed on an online E-commerce platform.", "outcome": "Experimental results show that this adaptive learning method outperforms previous methods. Furthermore, the method has been successfully implemented on an online E-commerce search platform, contributing to significant economic benefits.", "future_impact": "The adaptive metric learning method is general in nature and may be applicable in numerous other fields, suggesting significant potential for future impacts and improvements in various applications.", "venue": "EMNLP", "year": 2023, "title": "Adaptive Hyper-parameter Learning for Deep Semantic Retrieval."} +{"pid": "6344dede90e50fcafd24d2f6", "context": "A recent introduction of benchmarks for embodied dialogue instruction following, a complex sequence of natural language tasks, raised the question of how best to train and evaluate models for this multi-turn, multi-agent, long-horizon task.", "key_idea": "The paper suggests that imitation learning and related low-level metrics are misleading and do not align with the goals of embodied dialogue research, potentially hindering progress. It offers findings related to action patterns, grounding of utterances, and a call for a focus on higher-level semantic goals.", "method": "The study involves empirical comparisons of metrics and the analysis of three models.", "outcome": "The authors observed that models trained with imitation learning take spurious actions during evaluation and that existing models fail to ground query utterances, which are essential for task completion.", "future_impact": "The paper's input could shift research focus towards achieving higher-level semantic goals and improve model evaluation methods and training in embodied dialogue instruction following.", "venue": "EMNLP", "year": 2022, "title": "Don\u2019t Copy the Teacher: Data and Model Challenges in Embodied Dialogue"} +{"pid": "645c5e48d68f896efa22cf27", "context": "Large language models (LLMs) are increasingly incorporating external references to generate and support their claims. However, attributing these claims to their proper sources and verifying their accuracy is an open problem. Current evaluations mainly rely on human judgement which is costly and time-consuming.", "key_idea": "This paper explores the automatic evaluation of attribution in LLMs, and introduces two methods for automatic evaluation: prompting LLMs and fine-tuning smaller LMs, making use of data from existing tasks such as question answering and fact-checking.", "method": "The authors curated a test set from a generative search engine, New Bing, across 12 domains. They also created simulated test examples from existing benchmark questions and used these sets to evaluate their new approaches for automatic attribution.", "outcome": "Results from the test sets showed promising signals for automatic evaluation of attribution, but also highlighted several remaining challenges in the field.", "future_impact": "The researchers hope their testbed, modeling methodology, and insights will help lay the foundation for future studies on automatically evaluating attribution in LLMs.", "venue": "EMNLP", "year": 2023, "title": "Automatic Evaluation of Attribution by Large Language Models."} +{"pid": "5ff8839291e011c832673a4e", "context": "The process of architecture search, which designs optimal model or cell structures for specific tasks, has shown improvements utilizing the Efficient Neural Architecture Search (ENAS) method. However, the latter is limited due to its relatively rigid search space and it can't reproduce LSTM and GRU architectures.", "key_idea": "The authors propose Flexible and Expressible Neural Architecture Search (FENAS), an architecture search algorithm with a more flexible and expressive search space, capable of reproducing LSTM and GRU architectures, and initializing with them for more efficient architecture discovery.", "method": "The authors explore the extended search space of FENAS via evolutionary search, testing the approach on several popular text classification tasks as well as a standard language model benchmark.", "outcome": "FENAS significantly outperformed ENAS on several text classification tasks while performing similarly on a standard language model benchmark.", "future_impact": "The extended flexibility and expressibility of the FENAS approach provides a promising avenue for improving the efficiency and effectiveness of architecture search in neural networks, and could spark further investigation in this direction.", "venue": "EMNLP", "year": 2020, "title": "FENAS - Flexible and Expressive Neural Architecture Search."} +{"pid": "63a1750d90e50fcafd1f3ab2", "context": "Disagreements are often studied either in terms of toxicity detection or analysis of argument structure; however, other dialogue acts such as asking questions and providing clarification that may help resolve disputes have been less explored.", "key_idea": "The authors propose a unified framework of dispute tactics that includes toxicity, argument structures, and other dialogue acts. This framework also includes a preferential ordering among rebuttal-type tactics, ranging from ad hominem attacks to refuting the central argument.", "method": "Using the proposed framework, the authors annotate 213 disagreements (3,865 utterances) from Wikipedia Talk pages. They develop models for multilabel prediction of dispute tactics in an utterance, with a transformer-based label powerset model, and add an auxiliary task to incorporate the ordering of rebuttal tactics.", "outcome": "The study provides empirical validation of the approach to disagreement recommended by Wikipedia. The authors' models show improved performance in predicting escalation when using the proposed dispute tactic annotations.", "future_impact": "The authors suggest that their dispute tactics annotations can provide useful signals to enhance performance on predicting escalation, potentially leading to better online discussion environments.", "venue": "EMNLP", "year": 2022, "title": "How to disagree well: Investigating the dispute tactics used on Wikipedia"} +{"pid": "4e9c267a-9b47-40fa-97e1-e27ae01b0ce2", "context": "In large collections of independently written documents, it's critical to detect passages of copied or similar text across documents, a task often associated with plagiarism detection. An effective and efficient methodology for this issue has not been adequately addressed so far.", "key_idea": "The key idea is using a statistical method for 'fingerprinting' text. Each text in a large collection of documents is associated with a unique fingerprint, with similarities between fingerprints indicating copied or similar text passages. This method makes use of the inherent distribution of word trigrams.", "method": "The paper suggests a method utilizing set theoretic principles to determine similarity based on the word trigram distribution characteristics. The method was developed using a broadcast news reports corpus and was tested on student work to detect plagiarism.", "outcome": "The system successfully identified plagiarism in students' work, proving its ability to detect not only identical but also similar sections of text.", "future_impact": "The method is a pioneering detection model for text plagiarism. It could be utilized in a wide variety of contexts where large scale plagiarism detection is needed, for instance, in academic or journalistic fields. With the ability to detect both similar and identical sections of text, it opens up further paths of exploration in plagiarism and similar content detection.", "venue": "EMNLP", "year": 2001, "title": "Detecting Short Passages of Similar Text in Large Document Collections"} +{"pid": "6268a6795aee126c0f143050", "context": "Transformer-based language models (LMs) are known for their opaque nature and unexplained behavior, leading to a wide interest in interpreting their predictions. However, current interpretation methods focus on probing models externally, executing behavioral tests, and analyzing salience input features, ignoring the internal prediction construction process.", "key_idea": "The authors introduce LM-Debugger, an interactive tool that provides a fine-grained interpretation of the internal prediction process of transformer-based language models, and a framework to intervene in LM behavior.", "method": "LM-Debugger uses a method that interprets internal token representations and updates them by the feed-forward layers in the vocabulary space. The authors demonstrated the utility of LM-Debugger for single-prediction debugging on GPT2, inspected the internal disambiguation process, and showed how the tool enables users to shift model behavior by identifying certain vectors in the network for interventions.", "outcome": "The paper demonstrates how the LM-Debugger facilitates the interpretation of internal processes and intervention in model behavior. The tool was found effective in disambiguation and prediction process interventions, allowing users to easily change the model\u2019s behavior as desired.", "future_impact": "LM-Debugger has been released as an open-source tool, where its future use will help improve understanding and control over transformer-based language model behavior.", "venue": "EMNLP", "year": 2022, "title": "LM-Debugger: An Interactive Tool for Inspection and Intervention in Transformer-Based Language Models"} +{"pid": "628ef0495aee126c0f82db2f", "context": "The DocRED dataset is a large-scale, widely used benchmark for document-level relation extraction (RE) with a recommended-revise annotation scheme, however, it suffers from having false negative samples due to its incomplete annotation.", "key_idea": "The authors re-annotated 4,053 documents in the DocRED dataset to address the issue of prevalent false negative samples, resulting in the revised dataset called Re-DocRED.", "method": "The authors conduct extensive experiments with state-of-the-art neural models on both the original DocRED and the revised Re-DocRED datasets to assess the impact of revisions on model performance.", "outcome": "The models trained and evaluated on the revised Re-DocRED dataset showed performance improvements of around 13 F1 points compared to models trained on the original DocRED dataset.", "future_impact": "The new dataset, Re-DocRed, and the analysis of potential areas for further improvement could guide future development in document-level relation extraction models.", "venue": "EMNLP", "year": 2022, "title": "Revisiting DocRED -- Addressing the False Negative Problem in Relation Extraction"} +{"pid": "9bf2c38c-05ff-439c-b534-99d235e4c893", "context": "Existing models for Part-Of-Speech (POS) tagging use a training set to predict tags for unseen text, but there are still challenges in accuracy and handling difficult tagging decisions.", "key_idea": "The authors present a statistical model classified as a Maximum Entropy model, able to use multiple contextual features concurrently to predict POS tags.", "method": "The paper is essentially about the implementation of specialized features in the model to address difficult tagging decision scenarios.", "outcome": "Their model assigns POS tags to unseen text with state-of-the-art accuracy.", "future_impact": "The authors propose a training direction that tackles corpus consistency problems discovered in the implementation phase, implying potential enhancements in the field of POS tagging.", "venue": "EMNLP", "year": 1996, "title": "A Maximum Entropy Model for Part-Of-Speech Tagging"} +{"pid": "b47aae96-f4d8-43db-a3bc-2495aecaed8e", "context": "Current methods for cross-lingual transfer learning for dependency parsing rely on aligned corpora or bilingual lexicons.", "key_idea": "The authors introduce a feature transfer learning method with minimal supervision, which adapts an existing parser for another language by transferring the parsing features from the source language into the target language, specifically from Japanese to Korean.", "method": "The authors employed a parsing algorithm called the Triplet/Quadruplet Model for Japanese, and then applied a delexicalized feature transfer for Korean. They conducted experiments on the Penn Korean Treebank.", "outcome": "Using only the transferred features from Japanese, the method achieved an accuracy of 81.6% for Korean dependency parsing. Further improvements were achieved when a small annotated Korean corpus was combined with the Japanese training corpus.", "future_impact": "The study demonstrates that efficient cross-lingual transfer learning can be achieved without expensive linguistic resources, suggesting a direction for future work focused on the adaptation of transfer learning techniques with limited resources.", "venue": "EMNLP", "year": 2014, "title": "Learning from a Neighbor: Adapting a Japanese Parser for Korean Through Feature Transfer Learning"} +{"pid": "1cb88245-e16e-4261-9690-3875afd49a80", "context": "There is an ongoing effort to construct a new dependency representation framework for the highly agglutinative morphology of Turkish as well as to allow the annotation of unedited web data.", "key_idea": "This paper presents a novel syntactic representation for morphosyntactic sub-word units (inflectional groups in Turkish) which allows inter-IG relations to be discerned without hiding lexical information and a better coordination structure annotation scheme with nearly 11% increase in recall scores.", "method": "This work is done as part of an ongoing project to develop a new dependency representation framework for Turkish, focusing on the development and evaluation of novel syntactic representations for morphosyntactic sub-word units and coordination structures.", "outcome": "The paper presents preliminary findings, indicating that the proposed novel syntactic representations and annotation scheme provide an 11% increase in recall scores over the existing Turkish Treebank scheme.", "future_impact": "These insights can inform future efforts to improve the representation of morphosyntactic units and coordination structures in Turkish and potentially other agglutinative languages.", "venue": "EMNLP", "year": 2013, "title": "Representation of Morphosyntactic Units and Coordination Structures in the Turkish Dependency Treebank"} +{"pid": "3f44002f-a8e3-4dda-8a15-30805fe39e3a", "context": "Chinese word segmentation (CWS) and POS tagging have traditionally required task-specific feature engineering.", "key_idea": "The authors propose using deep layers of neural networks to discover relevant features to the tasks of CWS and POS tagging, thus eliminating the need for task-specific feature engineering. They also propose a perceptron-style algorithm for training the neural networks.", "method": "The authors applied deep learning on large-scale unlabeled data to improve internal representation of Chinese characters and enhance supervised word segmentation and POS tagging models", "outcome": "The proposed neural networks yielded close to state-of-the-art performance with minimal computational cost.", "future_impact": "The proposed perceptron-style algorithm could potentially speed up the training process and make the learning algorithm easier to implement.", "venue": "EMNLP", "year": 2013, "title": "Deep Learning for Chinese Word Segmentation and POS Tagging"} +{"pid": "297b73c1-7743-4fd8-863e-713ee456b851", "context": "Computing Distributional Thesauri (DTs) for very large corpora has been a challenge due to computational resource constraints.", "key_idea": "The authors introduce a new highly scalable approach to compute Distributional Thesauri by employing pruning techniques and a distributed framework.", "method": "The authors test their new approach for computing DTs on the entire vocabulary of Google Books syntactic n-grams and evaluate its performance against lexical resources using two measures.", "outcome": "The proposed approach produces higher quality Distributional Thesauri than previous approaches, and performs better in terms of speed and quality for large corpora.", "future_impact": "Given its efficiency and effectiveness, the proposed method for computing Distributional Thesauri has the potential to be the preferred choice for tackling large corpora.", "venue": "EMNLP", "year": 2013, "title": "Scaling to Large^3 Data: An Efficient and Effective Method to Compute Distributional Thesauri"} +{"pid": "26061c13-693d-47f5-8b10-5f6a2ee96159", "context": "The quality of machine translation (MT) systems is usually evaluated using automated metrics, but these may not always reflect actual user satisfaction or the efficiency of the system.", "key_idea": "This paper suggests that gaze metrics, derived from recording people's eye movements while they read machine translated texts, could serve as better measures of MT quality than current automated metrics.", "method": "The paper explores the use of gaze metrics to gauge MT quality effectiveness, leveraging the emergence of reliable eye-tracking technologies for home computers and mobile devices.", "outcome": "The study demonstrates that reading-based metrics, obtained through eye-tracking, are superior proxies for assessing machine translation quality than automated metrics currently in use.", "future_impact": "The findings suggest that eye-tracking metrics could provide a cheap and accessible method for getting feedback and guiding adaptation in machine translation systems.", "venue": "EMNLP", "year": 2015, "title": "Reading metrics for estimating task efficiency with MT output"} +{"pid": "5f9f6161-fd89-4ca8-8e20-4697a52c19cb", "context": "Machine translation evaluation relies on reference texts to gauge the quality of system outputs, but deciding which types of features are important is often done upfront and bears risk of not being adequate.", "key_idea": "The study puts forth a pairwise learning-to-rank approach that integrates several layers of tree-based linguistic information from both reference and system outputs and leverages preference re-ranking kernels to automatically learn important features.", "method": "The authors use the learning framework of preference re-ranking kernels to learn features automatically, combining both reference and system outputs. The performance of this method is then contrasted with traditional methods of machine translation evaluation.", "outcome": "The results show that using the proposed method of structural kernel learning yields better correlation with human translation evaluation than the traditional direct similarity method.", "future_impact": "The Structural Kernel Learning method could serve as a general framework for Machine Translation evaluation; syntactic and semantic information can be naturally incorporated into it.", "venue": "EMNLP", "year": 2014, "title": "Learning to Differentiate Better from Worse Translations"} +{"pid": "6346305e90e50fcafda07403", "context": "Mixture-of-Experts (MoE) networks have been proposed as an efficient way to scale up model capacity and implement conditional computing, but the use of MoE has mostly been confined to the feedforward layer in Transformer architecture.", "key_idea": "The paper proposes the Mixture of Attention Heads (MoA), a new architecture that combines multi-head attention with the MoE mechanism, allowing dynamic selection of attention heads per token and offering enhanced performance and potential for increased interpretability.", "method": "The authors tested the MoA on several tasks, including Machine Translation and Masked Language Modeling.", "outcome": "The experiments showed that the Mixture of Attention Heads achieved promising results against strong baselines involving large and very deep models.", "future_impact": "The sparsely gated MoA architecture offers the ability to scale up the number of attention heads and parameters while maintaining computational efficiency. Moreover, its ability to differentiate heads' utilities provides a new perspective for discussing the model's interpretability.", "venue": "EMNLP", "year": 2022, "title": "Mixture of Attention Heads: Selecting Attention Heads Per Token"} +{"pid": "63a2c50690e50fcafdb98fe9", "context": "Commonsense knowledge graphs (CKGs) in English are used in various natural language processing tasks. However, existing CKGs are limited to English language, impeding research in non-English languages. Directly deriving commonsense knowledge from pretrained language models has gained attention, but yet to be explored in non-English languages.", "key_idea": "The authors present an approach to generate a large-scale Chinese Commonsense Knowledge Graph (CKG) called CN-AutoMIC from multilingual pretrained language models (PLMs). They introduce a 'generate-by-category' strategy to lessen invalid generation, cascaded filters discard low-quality results, and a bootstrapping iteration process to reuse previously generated results.", "method": "The authors implement a series of strategies in CN-AutoMIC to generate, filter and increase the quality and diversity of the CKG. The analysis was conducted from various aspects to test the effectiveness of these strategies.", "outcome": "The proposed CN-AutoMIC surpasses the direct translation versions of similar English CKGs in terms of quality and diversity. The paper identifies specific issues and differences in relations that need to be addressed in future research for generating commonsense knowledge.", "future_impact": "The authors share their resources and models for future research, implying that their work could be further improved, extended and used as a basis for other research in the field of natural language processing, especially for non-English languages.", "venue": "EMNLP", "year": 2022, "title": "CN-AutoMIC: Distilling Chinese Commonsense Knowledge from Pretrained Language Models"} +{"pid": "20b908ef-a2fe-47b1-a0b4-60f00774744d", "context": "Prior work focuses on learning multiple related tasks but often fail to sufficiently exploit prior knowledge relating the output spaces of the tasks.", "key_idea": "The authors propose a new algorithmic framework for learning multiple related tasks which leverages prior knowledge that relates the output spaces of these tasks.", "method": "PAC learning results are presented to analyze the conditions under which such learning is possible and practical experiments including learning a shallow parser and named-entity recognition system that uses the new framework are also conducted.", "outcome": "The study yields consistent improvements over the existing baseline methods in the learning of a shallow parser and named-entity recognition system.", "future_impact": "The new learning framework with its ability to exploit prior knowledge relating to output spaces can significantly improve the learning accuracy in AI systems, facilitating precise task execution.", "venue": "EMNLP", "year": 2008, "title": "Cross-Task Knowledge-Constrained Self Training"} +{"pid": "618cfba991e011333c932592", "context": "While contextualized embeddings have driven increased interest in neural ranking approaches for information retrieval, two challenges persist: i) queries usually comprise a few keywords that increase ambiguity and complicate their contextualization, and ii) executing neural ranking on non-English documents is cumbersome due to the lack of labeled datasets.", "key_idea": "To navigate these issues, the paper introduces SIR (Sense-enhanced Information Retrieval), which leverages word sense information to enhance understanding. This includes a novel, multilingual query expansion mechanism grounded on Word Sense Disambiguation that proffers sense definitions as additional semantic data for the query.", "method": "The authors use senses as a bridge across languages, and compare the performance of their model against supervised and unsupervised alternatives across French, German, Italian, and Spanish languages on several CLEF benchmarks, all while training solely on English Robust04 data.", "outcome": "The SIR model performs significantly better than its supervised and unsupervised alternatives across several languages on various CLEF benchmarks, despite being trained solely on English data.", "future_impact": "The authors release SIR for public usage, suggesting potential broader applications and continuous developments in the field.", "venue": "EMNLP", "year": 2021, "title": "IR like a SIR - Sense-enhanced Information Retrieval for Multiple Languages."} +{"pid": "618cfba991e011333c932556", "context": "Pretrained Masked Language Models (MLMs) have been previously found to be ineffective as universal lexical and sentence encoders off-the-shelf, i.e., without further task-specific fine-tuning on tasks such as natural language inference, sentence similarity, or paraphrasing using annotated task data.", "key_idea": "The authors propose an easily implementable, quick, and effective contrastive learning technique known as Mirror-BERT, which aims to turn MLMs, like BERT and RoBERTa, into effective lexical and sentence encoders with no need for additional external knowledge or data, they simply use self-supervision.", "method": "Mirror-BERT relies on identical and slightly altered string pairs as positive (synonymous) fine-tuning examples and seeks to maximize their similarity during 'identity fine-tuning'. The authors apply Mirror-BERT to both lexical and sentence-level tasks across multiple domains and languages, particularly focusing on sentence similarity (STS) and question-answer entailment (QNLI) tasks.", "outcome": "Mirror-BERT dramatically improves the performance of off-the-shelf MLMs in both lexical and sentence-level tasks across varied domains and languages. The self-supervised Mirror-BERT model even matches the performance of Sentence-BERT models from prior work in STS and QNLI tasks that depend on annotated task data.", "future_impact": "The authors provide some evidence on why this straightforward Mirror-BERT fine-tuning approach can successfully produce effective universal lexical and sentence encoders, which could inform future research strategies and approaches for developing universal language encoders.", "venue": "EMNLP", "year": 2021, "title": "Fast, Effective, and Self-Supervised - Transforming Masked Language Models into Universal Lexical and Sentence Encoders."} +{"pid": "6576dbc9939a5f408216ef0f", "context": "BERT-based models are known for their strong performance in various applications, but they often exhibit poor generalizability in real-world settings, especially when the amount of training data is limited.", "key_idea": "This study examines the influence of the quality rather than the quantity of the training data on the generalization of a model, particularly focusing on two types of data: human-adversarial (h-adversarial), sample pairs with small differences but different ground-truth labels, and human-affable (h-affable), sample pairs with minor differences but the same ground-truth labels.", "method": "The authors conducted experiments to analyze the effects of including h-adversarial and h-affable instances in the training samples in a fixed percentage for tasks of text classification and relation extraction.", "outcome": "The study found that including 10-30% of h-adversarial instances in the training set can improve the precision and F1 score by up to 20 points in tasks of text classification and relation extraction. However, when the percentage of h-adversarial instances is increased beyond this range, the performance either plateaus or degrades. In contrast, h-affable instances may not enhance or may even degrade the generalization performance.", "future_impact": "The findings of this study can improve our understanding of the effects of training data quality on model generalization and potentially guide the design of more effective training data collection strategies for better generalizability.", "venue": "EMNLP", "year": 2023, "title": "Effects of Human Adversarial and Affable Samples on BERT Generalization"} +{"pid": "61397b715244ab9dcb323bd2", "context": "Predominantly, pretrained language models (PLMs) are combined with textual patterns to enhance performance in both the zero and few-shot learning settings. The patterns used either directly resemble the text seen during the PLM's pretraining (for zero-shot tasks) or they are adapted more flexibly to the PLM's unique characteristics (for supervised training tasks).", "key_idea": "The authors propose CONtinuous pAtterNs (CONAN), contrasting it with previous methods where a 'token' can be any continuous vector rather than having to make a discrete choice between vocabulary elements.", "method": "The authors evaluate CONAN on two established benchmarks for lexical inference in context (LIiC) a.k.a. predicate entailment, which is a challenging natural language understanding task with relatively small training sets.", "outcome": "CONAN consistently outperforms discrete patterns, setting a new state of the art on the LIiC task. It provides valuable insights into the kind of pattern that enhances a PLM's performance on LIiC.", "future_impact": "The approach raises important questions about our understanding of PLMs using text patterns and could lead to further investigation in this direction.", "venue": "EMNLP", "year": 2021, "title": "Continuous Entailment Patterns for Lexical Inference in Context."} +{"pid": "c41bfa6d-4c63-4aa9-942e-25ea33c2c5b7", "context": "Previous attempts to use natural language processing (NLP) for text categorization and information retrieval (IR) have generated mixed results, with no clear indication of NLP's utility in these tasks.", "key_idea": "The authors assert the significance of NLP in a specific task involving categorization of images captioned with sentences in which the subject and the predicate are critical elements.", "method": "The authors validate the importance of NLP for this task through two methods: by introducing a new system and comparing its performance with existing ones in literature, and by conducting experiments with human subjects.", "outcome": "The study presents evidence from experiments that indicate NLP can generally improve the speed and accuracy of the task.", "future_impact": "The study underscores the potential influence of NLP in improving text categorization and information retrieval tasks where linguistic elements are crucial, paving the way for further research in the area.", "venue": "EMNLP", "year": 2002, "title": "NLP Found Helpful (at least for one Text Categorization Task)"} +{"pid": "cc93b1b3-e294-46ed-9a49-a6d4455274cc", "context": "Neural models of language are often considered as 'black boxes', their intricate inner-workings and the representations they acquire are not fully understood.", "key_idea": "The first BlackboxNLP workshop focused on developing resources and techniques for analyzing and understanding the inner-workings and representations acquired by neural models of language.", "method": "Approaches included systematic manipulation of input to neural networks and investigating the impact on their performance, testing whether interpretable knowledge can be decoded from intermediate representations acquired by neural networks, proposing modifications to neural network architectures to make their knowledge state or generated output more explainable, and examining the performance of networks on simplified or formal languages.", "outcome": "The paper reviews a number of representative studies in each category, exploring techniques for interpreting and understanding neural models for natural language processing.", "future_impact": "This work might promote the understanding of how neural models process language, potentially leading to improvements in how these models are designed and interpreted.", "venue": "EMNLP", "year": 2019, "title": "Analyzing and interpreting neural networks for NLP: A report on the first BlackboxNLP workshop"} +{"pid": "956b3cf1-52db-4e4f-895c-cf7a5b02f429", "context": "Traditional analyses of web search queries have focused predominantly on semantic and pragmatic relationships between words, without considering the temporal patterns of these queries.", "key_idea": "The authors investigate temporal patterns of web search queries to reveal semantic and pragmatic relationships between words. In addition, they focus on two applications: query suggestion and query categorization.", "method": "They carry out several evaluations to analyze the properties of temporal profiles of queries, and apply measures of time-series similarity in the context of query suggestion and categorization.", "outcome": "The investigation shows that time-series similarity measures can enhance the effectiveness of query suggestion, achieving state-of-the-art performance. However, they found that the temporal profile alone is not a strong indicator of broad topical categories in query categorization.", "future_impact": "The insights derived about semantic relationships and query categorization from the investigation of temporal patterns in web search queries have potential applications in improving search engine functionality and user experience.", "venue": "EMNLP", "year": 2009, "title": "Gazpacho and summer rash: lexical relationships from temporal patterns of web search queries"} +{"pid": "f979c2b9-3e2f-4a05-a059-ceae01c9db17", "context": "Name filtering algorithms often require a balance between computational cost and the capability to filter based on phonetic similarity. Phonetic keys such as Soundex are commonly used for this role.", "key_idea": "The paper introduces a new algorithm called ETK (Ensemble of Transformation-based Keys) that maintains the advantages of phonetic keys like Soundex including low computational cost and filtering by phonetic similarity, and is also adaptable to alternative similarity models.", "method": "A preliminary empirical evaluation was conducted to assess the accuracy of ETK, in order to suggest its suitability for certain applications.", "outcome": "The ETK algorithm showed promising accuracy in a preliminary empirical evaluation, suggesting its potential use in phonetic filtering applications, including cross-lingual transliterations.", "future_impact": "The ETK algorithm could be adapted for various phonetic filtering applications, notably in recognizing alternative cross-lingual transliterations.", "venue": "EMNLP", "year": 2007, "title": "Inducing Search Keys for Name Filtering"} +{"pid": "5f969d2e91e01156ea5b369c", "context": "Natural Language Understanding (NLU) is a crucial component of conversational AI or digital assistant systems, tasked with semantically understanding user requests. The current state of these systems demands improvements in NLU.", "key_idea": "The authors propose a scalable and automatic approach to enhance NLU in large-scale conversational AI systems by utilizing implicit user feedback. They posit the idea that user interaction data and dialogue context contain rich information that can help infer user satisfaction and intentions.", "method": "The authors propose a domain-agnostic framework for curating new supervision data from live production traffic for NLU improvement. The effectiveness of the framework is tested through extensive experiments.", "outcome": "The proposed framework, when applied to a large-scale production system, showed improvements in NLU across 10 domains.", "future_impact": "This scalable, automatic approach for NLU improvement provides a way to improve the functionality and performance of large-scale conversational AI systems across various domains.", "venue": "EMNLP", "year": 2021, "title": "A Scalable Framework for Learning From Implicit User Feedback to Improve Natural Language Understanding in Large-Scale Conversational AI Systems."} +{"pid": "622abdd15aee126c0f56bbe0", "context": "Natural Language Generation (NLG) in non-English languages, especially Indic languages, is often hampered by the lack of suitable benchmark datasets.", "key_idea": "The paper presents the 'IndicNLG Benchmark', the first of its kind for Indic languages, which encompasses datasets of 5 diverse NLG tasks for 11 different Indic languages.", "method": "The authors use the datasets for benchmarking the performance of several monolingual and multilingual baselines leveraging pre-trained sequence-to-sequence models. The methods include data scraping, light cleaning, and pivoting via machine translation data.", "outcome": "The results demonstrate the strong performance of multilingual language-specific pre-trained models, and the utility of models trained on the IndicNLG datasets for various NLG tasks. The IndicNLG Benchmark is a diverse multilingual NLG dataset, with approximately 8M examples across 5 tasks and 11 languages.", "future_impact": "The paper's dataset creation methods can be applied to other resource-limited languages, potentially benefiting a wider range of linguistic and NLP studies. Also, making the datasets and models publicly available could encourage further improvements in NLG technology for Indic languages.", "venue": "EMNLP", "year": 2022, "title": "IndicNLG Benchmark: Multilingual Datasets for Diverse NLG Tasks in Indic Languages"} +{"pid": "636dbe6990e50fcafd79ab0d", "context": "In regards to the surviving literature of Plutarch of Chaeronea, there are several texts which are believed not to have been written by him and are instead attributed to Pseudo-Plutarch. These texts are difficult to attribute to an author or era, presenting challenges in historical and literary research.", "key_idea": "The authors propose the application of a BERT language model to the study of Ancient Greek. This model is intended to uncover statistical properties that can provide insights into the authorship and context of Pseudo-Plutarch texts.", "method": "The authors apply a BERT language model to Pseudo-Plutarch texts and other contemporaneous writings, examining the statistical properties uncovered.", "outcome": "The BERT model revealed that the work 'Placita Philosophorum', along with another Pseudo-Plutarch text, showed similarities with texts written by authors from an Alexandrian context in the 2nd/3rd century CE.", "future_impact": "The study implies that BERT and similar models can be effectively utilized to probe and resolve complex issues in the field of historical literature and philosophy, suggesting a future direction for such domain applications.", "venue": "EMNLP", "year": 2022, "title": "BERT in Plutarch's Shadows"} +{"pid": "607ffec791e011772654f736", "context": "Existing question-answering (QA) research has not adequately addressed the diversity of answer types that commonly occur in daily questioning.", "key_idea": "The authors propose GooAQ, a large-scale dataset with a variety of answer types comprising over 5 million questions and 3 million answers, collected semi-automatically from the Google search engine, thus resulting in a wide array of answer types.", "method": "The authors mined the answers from Google's responses to the collected questions and then used the T5 models on the GooAQ dataset to observe the model's performance on short-answer questions and on generating coherent and accurate responses for long-response questions.", "outcome": "From the experiments, it was found that the T5 models perform strongly on GooAQ's short-answer questions due to annotated data. However, for long-answer questions, the quality is less reliant on annotated data and is mainly supported by pre-training.", "future_impact": "The authors have released GooAQ, hoping to facilitate further research on improving QA systems with diverse response types.", "venue": "EMNLP", "year": 2021, "title": "GooAQ - Open Question Answering with Diverse Answer Types."} +{"pid": "afc7f515-6f14-4c85-83c2-cf136caf820b", "context": "An increasing number of new devices, specifically speech-operated ones, are being integrated into cars to help minimize distractions for drivers.", "key_idea": "The paper presents the development of an in-car dialog system for an MP3 player that has technologies like end-pointing with prosodic cues, error identification and recovery strategies, flexible multi-threaded, multi-device dialog management, and content optimization and organization strategies.", "method": "The dialog system incorporates numerous components including Nuance speech recognition engine, Nuance Vocalizer as the TTS engine, a Prosody module for better natural language understanding and dialog management, a Natural Language Understanding module, a Dialog Manager module, a Knowledge Manager and a Content Optimization module. The system was evaluated based on task completion rate in the MP3 player domain.", "outcome": "Early evaluation of the system shows a task completion rate of 80% on 11 tasks in the MP3 player domain, from playing requests to music database queries.", "future_impact": "The developed dialog system's knowledge representation in OWL, a Web Ontology language recommended by W3C, will facilitate its portability to new domains. The porting to a restaurant selection domain is currently in progress.", "venue": "EMNLP", "year": 2005, "title": "A Flexible Conversational Dialog System for MP3 Player"} +{"pid": "618cfbaa91e011333c9327b6", "context": "Statistical methods have traditionally been used to determine whether two documents were composed by the same author, also known as authorship verification. Recently, authorship representations learned using neural networks have outperformed alternatives in large-scale scenarios.", "key_idea": "The study conducts a large-scale examination of cross-domain transfer for authorship verification, exploring the question of whether representations learned in a particular domain can transfer to other domains.", "method": "The study is conducted using zero-shot transfers involving three disparate domains: Amazon reviews, fanfiction short stories, and Reddit comments. It examines the properties of these domains that influence generalization and proposes methods to improve transfer.", "outcome": "It was found that while a surprising degree of transfer is possible between certain domains, it is not as successful between others.", "future_impact": "The study's findings regarding domain properties that influence generalization for authorship verification, as well as the proposed methods for improving transfer, may lead to more effective authorship verification across multiple domains in the future.", "venue": "EMNLP", "year": 2021, "title": "Learning Universal Authorship Representations."} +{"pid": "61b80b745244ab9dcbf49817", "context": "Keyphrase generation in scholarly domains has largely focused on generating keyphrases using only the title and abstract of the articles.", "key_idea": "The authors explore if incorporating additional information from the full text of an article or from semantically similar articles can enhance a neural keyphrase generation model. They propose incorporating sentences from the full text, particularly in the form of the extractive summary of the article, to significantly improve keyphrase generation.", "method": "The authors experiment with three widely used models for keyphrase generation plus one of the latest transformer models suitable for long documents, Longformer Encoder-Decoder (LED). They also present a new large-scale scholarly dataset FullTextKP for keyphrase generation.", "outcome": "Adding sentences from the full text, especially in the form of the extractive summary of the article, significantly improves the generation of keyphrases that are either present or absent from the text.", "future_impact": "The FullTextKP dataset, which includes full text of the articles along with the title and abstract, and source code will be released for future work in this field.", "venue": "EMNLP", "year": 2021, "title": "Keyphrase Generation Beyond the Boundaries of Title and Abstract"} +{"pid": "7bbd8f66-b461-407c-b919-9b838814b012", "context": "In machine translation, syntactic pre-ordering is used to re-arrange source words into a target-language-like order. This process traditionally relies on supervised parsers trained on a tree-bank.", "key_idea": "The authors propose a new method to automatically induce parse trees from a parallel corpus instead of using a supervised parser, and use these induced parses to pre-order source sentences in machine translation.", "method": "The paper applies the method to a state-of-the-art phrase-based system with integrated reordering and compares the performance with a recent pre-ordering method based on a supervised parser.", "outcome": "The experiments demonstrate that the induced parser improves the state-of-the-art phrase-based system with integrated reordering and its performance approaches that of the pre-ordering method that uses a supervised parser.", "future_impact": "The study establishes a new application for unsupervised grammar induction, showing that syntactic structure relevant to machine translation pre-ordering can be learned automatically from parallel text.", "venue": "EMNLP", "year": 2011, "title": "Inducing Sentence Structure from Parallel Corpora for Reordering"} +{"pid": "0a3d9c1d-7f5f-4e66-9191-1884728a0620", "context": "Twitter is often utilized in quantitative studies for determining geographically-preferred topics, writing styles, and entities, relying on GPS coordinates attached to messages or user-supplied location profiles.", "key_idea": "The authors bring to focus the biases introduced by different data acquisition techniques in geotagged Twitter data. They aim to understand their impacts on linguistic analysis and text-based geolocation.", "method": "The authors compare GPS-tagging and self-reported locations and measure the biases they introduce. A latent variable model is used to induce age and gender", "outcome": "Findings show that the GPS-tagging and self-reported locations yield different corpora partially attributed to differences in dataset composition by age and gender. Furthermore, the accuracy of text-based geolocation varies with population demographics, giving the best results for men above the age of 40.", "future_impact": "The paper's findings provide insights into the biases of data acquisition techniques and their effects on linguistic analysis and text-based geolocation, providing potential considerations for future related studies.", "venue": "EMNLP", "year": 2015, "title": "Confounds and Consequences in Geotagged Twitter Data"} +{"pid": "5f7fe6d80205f07f689732de", "context": "Traditional entity linking models tend to be complex and non-scalable, and there's a need for a simple, scalability, and effective solution that can perform entity linking in a zero-shot manner which does not rely on explicit entity embeddings or manually engineered mention tables.", "key_idea": "The authors introduce a two-stage zero-shot entity linking model that operates by first retrieving entities in a dense space, defined by a bi-encoder which independently embeds the mention context and the entity descriptions. These candidates are then re-ranked using a cross-encoder that combines mention and entity text.", "method": "The authors evaluate the model's accuracy and speed trade-off via experiments on recent zero-shot benchmarks and more established non-zero-shot evaluations like TACKBP-2010. The experiments also investigate the speed of the bi-encoder in performing entity linking and explore the effectiveness of knowledge distillation in transferring cross-encoder's accuracy to the bi-encoder.", "outcome": "Experiments demonstrate that this approach achieves state-of-the-art performance on zero-shot benchmarks (6 point absolute gains) and non-zero-shot evaluations. The bi-encoder linking is found to be very fast in linking millions of entity candidates in a negligible amount of time, and the benefits from the more expensive cross-encoder can be effectively transferred to the bi-encoder via knowledge distillation.", "future_impact": "The simplicity, scalability and the impressive results of this entity linking model have the potential to improve various pertinent tasks in natural language processing, making entity linking and retrieval faster and more efficient in large-scale applications. The model and its source code being made publicly available might encourage further enhancements and related research.", "venue": "EMNLP", "year": 2020, "title": "Scalable Zero-shot Entity Linking with Dense Entity Retrieval"} +{"pid": "5f7d8e5191e011346ad27d78", "context": "Recent studies demonstrate that self-attention patterns in trained models contain a majority of non-linguistic regularities. However, there is currently a need for efficient self-supervised learning with bi-directional Transformers.", "key_idea": "The authors propose a computationally efficient auxiliary loss function to guide attention heads to conform to non-linguistic patterns in self-attention.", "method": "The proposed method is agnostic to the actual pre-training objective and its validity is evaluated on downstream tasks, specifically in low-resource settings.", "outcome": "The proposed method results in faster convergence of models and improved performance on downstream tasks, achieving state-of-the-art results in low-resource settings.", "future_impact": "The study offers the surprising finding that linguistic properties of attention heads are not necessarily correlated with language modeling performance, which may provide insights for future research in Transformer models and language modeling.", "venue": "EMNLP", "year": 2020, "title": "Guiding Attention for Self-Supervised Learning with Transformers"} +{"pid": "634e194090e50fcafd24e665", "context": "Existing automatic story evaluation methods focus on story lexical level coherence, which deviates from human preference.", "key_idea": "The authors proposed StoryER, a novel method for automatic story evaluation that is designed to mimic human preference when judging a story. StoryER fulfills three sub-tasks: ranking, rating and reasoning about the story.", "method": "StoryER uses the Longformer-Encoder-Decoder (LED) model finetuned on a new dataset comprising of 100k ranked story pairs and 46k ratings and comments on various aspects of the story.", "outcome": "The authors provide a competitive benchmark for each task, showing a high correlation with human preference. Joint learning of the preference scores, the aspect ratings, and comments results in gains in each single task.", "future_impact": "The authors offer this dataset and benchmarks for public use, which could serve to advance subsequent research in story evaluation tasks.", "venue": "EMNLP", "year": 2022, "title": "StoryER: Automatic Story Evaluation via Ranking, Rating and Reasoning"} +{"pid": "5e5f7c4791e011df604ec9a1", "context": "Prior to this study, there were no public large-scale monolingual language models pre-trained specifically for Vietnamese.", "key_idea": "The authors introduce PhoBERT which comes in two versions, 'base' and 'large' making it the first public large-scale monolingual language models pre-trained specifically for Vietnamese.", "method": "PhoBERT is applied to multiple Vietnamese-specific NLP tasks including Part-of-speech tagging, Named-entity recognition, and Natural language inference to test its performance.", "outcome": "PhoBERT improves the state-of-the-art results in multiple Vietnamese-specific NLP tasks.", "future_impact": "The authors release PhoBERT to facilitate future research and downstream applications for Vietnamese NLP.", "venue": "EMNLP", "year": 2020, "title": "PhoBERT: Pre-trained language models for Vietnamese"} +{"pid": "5ff8839291e011c8326739a3", "context": "Neural machine translation (NMT) has traditionally been approached as a classification task in an autoregressive setting, but the Zipfian nature of languages that causes imbalanced classes may impact performance.", "key_idea": "The paper explores the effect of vocabulary size on NMT performance, given the class imbalance caused by the Zipfian nature of languages.", "method": "The authors analyze the performance of NMT under different vocabulary sizes across multiple languages and data sizes.", "outcome": "The study reveals why certain vocabulary sizes are better suited for NMT and provides insights into how class imbalance impacts NMT performance.", "future_impact": "Finding the optimal vocabulary size for NMT can provide insights for future studies focusing on improving NMT models.", "venue": "EMNLP", "year": 2020, "title": "Finding the Optimal Vocabulary Size for Neural Machine Translation."} +{"pid": "6153e0335244ab9dcb39c30d", "context": "Improving the classification performance for challenging examples in machine learning models remains a challenge.", "key_idea": "The authors introduce a non-parametric post-processing step for classification called Classification with Alternating Normalization (CAN). It improves classification accuracy by re-adjusting the predicted class probability distribution using the predicted class distributions of high-confidence validation examples.", "method": "The properties of CAN are analyzed through simulated experiments, and its effectiveness is empirically demonstrated across a diverse set of classification tasks.", "outcome": "Empirical trials have shown CAN to be effective across a diverse set of classification tasks, although the abstract doesn't provide specific results.", "future_impact": "CAN can be easily applied to any probabilistic classifier, signaling potential wide applicability and impact in improving machine learning classification models.", "venue": "EMNLP", "year": 2021, "title": "When in Doubt - Improving Classification Performance with Alternating Normalization."} +{"pid": "618cfba991e011333c93263f", "context": "Building neural machine translation systems to perform well on a specific target domain is a common practice, but optimizing system performance for multiple, diverse target domains remains a challenge. The issue studied is in an adaptation setting where the goal is to preserve the existing system quality while incorporating data for domains that were not the focus of the original translation system.", "key_idea": "The key idea of this study is that it is possible to improve over the performance trade-off imposed by Elastic Weight Consolidation using a relatively simple data mixing strategy in neural machine translation systems.", "method": "The study includes a comparative analysis with Elastic Weight Consolidation and a proposed data mixing strategy in an adaptation setting.", "outcome": "The authors find that at comparable performance on the new domains, catastrophic forgetting is significantly mitigated on strong WMT baselines.", "future_impact": "No explicit future impact is mentioned, but since the study shows improvement in the performance of neural machine translation adaptations for diverse domains over the existing techniques, henceforth it will impact the optimization of such systems spanning multiple target domains.", "venue": "EMNLP", "year": 2021, "title": "Improving the Quality Trade-Off for Neural Machine Translation Multi-Domain Adaptation."} +{"pid": "8a306c6b-f995-40b8-a410-4bc618f9c402", "context": "Neural language models (NLMs) are generative and capable of modelling the distribution of grammatical sentences. However, their generative nature may not guarantee discrimination between 'good' and 'bad' sentences, leading to suboptimal performance.", "key_idea": "This study proposes transforming a generative Neural Language Model into a discriminative one, with the aim of increasing the margin between 'good' and 'bad' sentences.", "method": "The proposed method is trained end-to-end and applied to tasks involving the re-scoring of decoded text, specifically in ASR and SMT tasks.", "outcome": "Significant gains are observed in both Automatic Speech Recognition (ASR) and Statistical Machine Translation (SMT) tasks when using the proposed method.", "future_impact": "The proposed method can be widely applied for tasks that involve the re-scoring of decoded text.", "venue": "EMNLP", "year": 2018, "title": "Large Margin Neural Language Models"} +{"pid": "63a2c50d90e50fcafdb9a52f", "context": "Existing approaches for automatically generating video commentary have been generally concerned with specific fields where it is possible to leverage domain-specific information.", "key_idea": "The authors propose the task of generating video commentary in an open-domain fashion and detail the construction of a new large-scale dataset of transcribed commentary aligned with videos containing various human actions in a variety of domains.", "method": "The authors propose approaches based on well-known neural architectures to tackle the task and present an in-depth empirical study based on their data.", "outcome": "The results show clear trade-offs between textual and visual inputs for the models and highlight the importance of relying on external knowledge in this open-domain setting, offering robust baselines for this task.", "future_impact": "The newly proposed task and findings on the importance of relying on external knowledge in an open-domain setting may pave the way for improvements in future open-domain video commentary generation research.", "venue": "EMNLP", "year": 2022, "title": "Open-domain Video Commentary Generation"} +{"pid": "6350bc6690e50fcafdecea58", "context": "Active learning has been applied to natural language processing (NLP), but the specifics of applying AL to NLP problems have yet to be fully explored and understood.", "key_idea": "This paper presents a comprehensive survey of active learning strategies and their applications in NLP, covering several important aspects such as structured prediction tasks, annotation cost, model learning with deep neural models, and deciding when to start and stop active learning.", "method": "The authors conduct a detailed literature review and categorization of existing active learning strategies in the context of NLP tasks.", "outcome": "The paper provides a detailed categorization and discussion of various aspects of applying active learning to NLP problems and tasks.", "future_impact": "Identifies new areas for future research, including potential advancements in active learning for NLP, and their impacts on structured prediction tasks, annotation cost, model learning, and decisions on the initiation and termination of active learning.", "venue": "EMNLP", "year": 2022, "title": "A Survey of Active Learning for Natural Language Processing"} +{"pid": "634f6ae490e50fcafdcb657f", "context": "Large-scale language models, despite their excellent performance, are susceptible to various adversarial attacks. Many defense methods have been proposed, but they are limited due to redundant attack search spaces and inability to defend against various types of attacks.", "key_idea": "The paper introduces a new fine-tuning approach called 'RObust SEletive fine-tuning (ROSE)' that selectively updates parameters when adapting pre-trained models to downstream tasks, thereby filtering out invaluable and unrobust updates.", "method": "Two strategies are proposed under ROSE for selecting target robust parameters, referred to as first-order and second-order ROSE. This is tested against various downstream NLP tasks to evaluate improvements in adversarial robustness and integration with existing fine-tuning methods.", "outcome": "Experimental results show that ROSE significantly improves adversarial robustness across various downstream NLP tasks. Furthermore, an ensemble method surpasses both first-order and second-order ROSE variants. Empirical analysis confirms ROSE's ability to eliminate unrobust spurious updates, leading to flatter and wider optima than conventional methods.", "future_impact": "The authors anticipate that ROSE's ease of integration into existing fine-tuning methods will help enhance the adversarial robustness of these methods.", "venue": "EMNLP", "year": 2022, "title": "ROSE: Robust Selective Fine-tuning for Pre-trained Language Models"} +{"pid": "2243e32f-a2a5-4eb4-940d-a20b8d1e7733", "context": "Existing lexicons focus on categorizing words that express sentiments, without considering words that evoke or associate with a specific polarity of sentiment", "key_idea": "The authors introduce a connotation lexicon, which lists words with connotative polarity, i.e., words with positive or negative connotations. The key linguistic insight is selectional preference of connotative predicates.", "method": "The authors present graph-based algorithms using PageRank and HITS that collectively learn connotation lexicon together with connotative predicates.", "outcome": "The empirical study shows that the resulting connotation lexicon is valuable for sentiment analysis, complementing existing sentiment lexicons.", "future_impact": "The connotation lexicon alleviates the need for common sense and world knowledge in sentiment analysis and can enhance sentiment understanding in various applications.", "venue": "EMNLP", "year": 2011, "title": "Learning General Connotation of Words using Graph-based Algorithms"} +{"pid": "635753cc90e50fcafddddc2d", "context": "Cross-lingual pre-trained language models (XLMs) require accurate alignment between languages for improvement. Traditional code-switching, an approach for achieving word- or phrase-level linguistic alignment, relies on using dictionaries or parallel sentences with word alignment and involves randomly switching words in a sentence, but such methods can be suboptimal as they disregard semantics, and syntax might become invalid after random word switching.", "key_idea": "The authors introduce EntityCS, a method that focuses on Entity-level Code-Switching to capture fine-grained cross-lingual semantics without corrupting syntax, by switching entities to their counterparts in other languages.", "method": "The authors construct an entity-centric CS corpus using Wikidata and English Wikipedia by switching entities to their counterparts in other languages. They also propose entity-oriented masking strategies during intermediate model training on the EntityCS corpus for improving entity prediction.", "outcome": "When evaluated on four entity-centric downstream tasks, the trained models show consistent improvements over the baseline, with a notable increase of 10% in Fact Retrieval.", "future_impact": "The authors state that they are releasing the corpus and models to assist in research on code-switching and enriching XLMs with external knowledge, indicating their anticipation of further advances in these areas based on their work.", "venue": "EMNLP", "year": 2022, "title": "EntityCS: Improving Zero-Shot Cross-lingual Transfer with Entity-Centric Code Switching"} +{"pid": "5f7d8b6091e011346ad27d3b", "context": "Current NLP models are shown to lack robustness, as a model\u2019s prediction can be easily changed under small perturbations to the input.", "key_idea": "The authors present a Controlled Adversarial Text Generation (CAT-Gen) model that generates adversarial texts through controllable attributes that are invariant to task labels, to test and improve model robustness.", "method": "The CAT-Gen model is used to generate adversarial texts using controllable attributes, such as product categories in product review sentiment classification. The generated adversarial examples are used in adversarial training of different NLP models.", "outcome": "Experiments demonstrate that their method can generate more diverse and fluent adversarial texts compared to many existing adversarial text generation approaches and that the generated attacks improve the robustness of models against re-training and variations in model architecture.", "future_impact": "The text generation approach proposed in this study could be used widely for improving the robustness of NLP models against adversarial attacks.", "venue": "EMNLP", "year": 2020, "title": "CAT-Gen: Improving Robustness in NLP Models via Controlled Adversarial Text Generation"} +{"pid": "16d4b1b6-dd02-4f4c-98db-1c7169978e7a", "context": "MEANT is an automatic semantic frame-based Machine Translation (MT) evaluation metric with high correlation with human adequacy judgments. However, it uses simple word alignment heuristics which may not effectively score the phrasal similarity of semantic role fillers.", "key_idea": "The authors introduce a restructuring of MEANT using inversion transduction grammar (ITG) which can leverage ITG language biases to improve the correlation with human adequacy judgments.", "method": "The authors created a new metric called IMEANT, which uses bracketing ITGs to biparse the reference and machine translations, while maintaining the semantic frames in both.", "outcome": "The IMEANT metric shows improvements in evaluating MT output, providing a presumption that ITGs provide a better scoring system for phrasal similarity of semantic role fillers compared to the heuristics used in MEANT.", "future_impact": "The authors demonstrate that ITGs can provide extremely high coverage of cross-lingual semantic frame alternations, which can contribute to improving the accuracy of evaluating MT based on the degree of semantic frame preservation.", "venue": "EMNLP", "year": 2014, "title": "Better Semantic Frame Based MT Evaluation via Inversion Transduction Grammars"} +{"pid": "628707335aee126c0f78c4b2", "context": "The evaluation of summarization of long texts lacks appropriate frameworks. Current automatic and human evaluation methods fail to identify gaps in coherency, and there is a need for summaries to present a coherent narrative to be understandable by the reader.", "key_idea": "The authors introduce SNaC, a narrative coherence evaluation framework specifically designed for long summaries, based on fine-grained annotations.", "method": "The authors develop a taxonomy of coherence errors in generated narrative summaries and collect span-level annotations for 6.6k sentences across 150 book and movie screenplay summaries. They train a classifier to automatically localize coherence errors in generated summaries.", "outcome": "The authors offer the first characterization of coherence errors generated by state-of-the-art summarization models. They demonstrate that their collected annotations enable them to train a competent classifier for identifying coherence errors in summaries.", "future_impact": "The SNaC framework can support future work in long document summarization and coherence evaluation, including improved summarization modeling and post-hoc summary correction.", "venue": "EMNLP", "year": 2022, "title": "SNaC: Coherence Error Detection for Narrative Summarization"} +{"pid": "652dee7a939a5f4082b4494a", "context": "Recent text-to-image models such as Stable Diffusion struggle to effectively represent non-literal expressions. Current solutions mainly involve creating small-scale, human-annotated datasets, which are both demanding in terms of specialized expertise and highly inefficient.", "key_idea": "ViPE is introduced as a robust and lightweight series of language models, trained on a large set of lyrics with noisy visual descriptions that represent their implicit meaning. These synthetic visual descriptions are generated by GPT3.5, without relying on human annotations or images.", "method": "ViPE was trained on a large-scale set of lyrics with noisy visual descriptions generated by GPT3.5. The robustness of the model and its understanding of figurative expressions were tested.", "outcome": "ViPE effectively transforms text into a visualisable description and enables the generation of meaningful, high-quality images. Experiments show that ViPE is more robust than GPT3.5 in producing visual elaborations and its understanding of non-literal expressions is comparable to that of human experts.", "future_impact": "As an open-source solution, ViPE has potential for enhancing downstream applications like music video and caption generation by providing a more robust way of interpreting and visualising figurative expressions.", "venue": "EMNLP", "year": 2023, "title": "ViPE: Visualise Pretty-much Everything."} +{"pid": "9ed5eaa3-1fcd-422f-8c4d-afd474daf644", "context": "Previous work on extracting ideology from text has mainly been focused on domains where expression of political views is expected. However, it's unclear if these techniques are applicable in domains where displays of ideology are considered inappropriate.", "key_idea": "The authors propose a supervised ensemble n-gram model for ideology extraction with topic adjustments, which they apply to the domain of research papers written by academic economists.", "method": "The authors apply their proposed model to research papers in the field of academic economists. They then test their ability to predict the political leanings of these economists.", "outcome": "The model is able to successfully predict the political leanings of economists. These predictions are found to generalize well to new domains and correlate with public policy-relevant research findings.", "future_impact": "This study suggests that unsupervised models may under-perform in domains where ideological expression is discouraged, potentially guiding future research in this area.", "venue": "EMNLP", "year": 2014, "title": "Detecting Latent Ideology in Expert Text: Evidence From Academic Papers in Economics"} +{"pid": "618cfbaa91e011333c93280b", "context": "There has been recent research into quantum natural language processing (NLP), which involves designing algorithms that process natural language either in quantum computers or by leveraging quantum-inspired algorithms to improve NLP on classical computers.", "key_idea": "This study aims to provide a survey of the methods at the intersection of NLP and quantum physics from the past decade, categorizing them based on their use of quantum theory, the linguistic targets they model, and their downstream applications.", "method": "The authors conducted a literature review of methods that merge NLP and quantum physics from the past ten years, categorizing and discussing them based on key factors.", "outcome": "The authors successfully surveyed and categorized methods from the past ten years that combine quantum physics and NLP, discussing key factors to the success that has been achieved by existing work.", "future_impact": "The authors hint at future challenges and directions for the field, potentially promoting a better understanding of the promises and opportunities in merging quantum physics and NLP.", "venue": "EMNLP", "year": 2021, "title": "Natural Language Processing Meets Quantum Physics - A Survey and Categorization."} +{"pid": "635b486f90e50fcafd330d6b", "context": "While Multimodal Machine Translation (MMT) has gained interest, existing research struggles with the cost of training a separate model for each language pair, which is not feasible as the number of languages increases. Currently, the task of multilingual multimodal machine translation (Multilingual MMT), which provides a shared semantic space for multiple languages, has not been investigated.", "key_idea": "The authors propose the Multilingual MMT task and provide an effective baseline model LVP-M3 using visual prompts to support translations between different languages. This new task and model leverage the fact that the image modality, unlike languages, has no boundaries.", "method": "The authors establish two new Multilingual MMT benchmark datasets covering seven languages and conduct extensive experiments using their proposed LVP-M3 method, which includes token encoding, language-aware visual prompt generation, and language translation.", "outcome": "The experimental results demonstrate the effectiveness of the proposed LVP-M3 method for Multilingual MMT on the constructed benchmark datasets.", "future_impact": "As the first introduction of the Multilingual MMT task and the LVP-M3 model, this work opens up possibilities for future research on improving multilingual translation using visual inputs.", "venue": "EMNLP", "year": 2022, "title": "LVP-M3: Language-aware Visual Prompt for Multilingual Multimodal Machine Translation"} +{"pid": "edf684f1-c516-4af3-9ea3-0bdde25622e0", "context": "Classroom discussions in English Language Arts have a positive effect on students' reading, writing and reasoning skills. The existing literature mainly focuses on teacher talk and student-teacher interactions.", "key_idea": "This study introduces a new focus on three theoretically-motivated aspects of high-quality student talk: argumentation, specificity, and knowledge domain.", "method": "The authors introduce an annotation scheme and test its reliability and its ability to predict discussion quality.", "outcome": "The results show that the introduced annotation scheme can produce reliable annotations and is predictive of the quality of the discussion.", "future_impact": "The new annotation scheme presents new opportunities for both education and natural language processing research.", "venue": "NAACL", "year": 2018, "title": "Annotating Student Talk in Text-based Classroom Discussions"} +{"pid": "621454545aee126c0f20425f", "context": "The task of abductive natural language inference requires not only determining a causal relationship but also evaluating how reasonable an explanation is using common sense. All recent competitive systems for this task build on top of contextualized representations and make use of transformer architectures, and selecting the best model is a time-consuming and resource-intensive task.", "key_idea": "The authors propose a simple method for predicting the performance of an abductive natural language inference model without actually fine-tuning it, by comparing sentence embeddings with cosine similarity.", "method": "The authors test the proposed performance prediction method on the abductive natural language inference task by comparing the outcomes from the cosine similarity approach and a training a classifier on top of these embeddings.", "outcome": "The authors find that the accuracy of the cosine similarity approach correlates strongly with the accuracy of the classification method with a Pearson correlation coefficient of 0.65. The similarity computation is orders of magnitude faster to compute on a given dataset (less than a minute vs. hours).", "future_impact": "The proposed method can lead to significant time savings in the process of model selection for abductive natural language inference tasks.", "venue": "NAACL", "year": 2022, "title": "Embarrassingly Simple Performance Prediction for Abductive Natural Language Inference"} +{"pid": "2e83d82d-9e37-4c02-a787-5ee7af1b4cfc", "context": "Previous methods for bootstrapping gene name recognizers for FlyBase curation generally relied on fully supervised training on manually annotated biomedical text, but this approach may not be the most effective.", "key_idea": "The study explores a new method of bootstrapping a gene name recognizer using automatically annotated noisy text, rather than resorting to fully supervised training on manually annotated biomedical text.", "method": "The authors developed and use a new test set based on an annotation scheme marking the distinction between gene names and gene mentions, and use it to evaluate their recognizer. They then make extensions to the technique used to generate training data with the goal of improving performance on unseen genes.", "outcome": "The results showed that bootstrapping a gene name recognizer from automatically annotated noisy text was more effective. However, performance on unseen genes was identified as the method's main weakness.", "future_impact": "The approach used and insights obtained in this paper may guide further work in improving techniques for the training of biomedical named entity recognizers, specifically in handling unseen genes.", "venue": "NAACL", "year": 2006, "title": "Bootstrapping and Evaluating Named Entity Recognition in the Biomedical Domain"} +{"pid": "c889389e-45a0-45b0-b2cd-8f37f3606c0a", "context": "Normally, call-routing applications in the industry elicit speech from customers and then send them all to the same destination, such as an existing touch-tone menu, without considering what they have said. Traditional Wizard-of-Oz data collections are generally not fully integrated within existing services.", "key_idea": "The authors propose a new approach to data collection using a Wizard-of-Oz approach, where a simulated call-routing system replaces the existing touch-tone navigation system within a real service with actual customers and service agents.", "method": "The authors implemented the Wizard-of-Oz approach in a real service, replacing the existing touch-tone navigation system with a simulated call-routing system, and collected a corpus of 42,000 dialogues for a call-routing application.", "outcome": "They were able to collect a comprehensive dataset of real customer interactions, effectively situating the data collection process in a live, real-world service context.", "future_impact": "This methodology demonstrates a promising way to collect more relevant and realistic data in an actual service environment. This may shift how future data collection processes for developing AI applications are conducted.", "venue": "NAACL", "year": 2007, "title": "Experiences of an In-Service Wizard-of-Oz Data Collection for the Deployment of a Call-Routing Application"} +{"pid": "628afb4c5aee126c0f04e559", "context": "Contrastive learning (CL) has been successful in the computer vision, speech, and natural language processing fields with self-supervised learning. However, its application to supervised settings, particularly the natural language understanding classification task, is not fully explored.", "key_idea": "The authors propose a novel label anchored contrastive learning approach (LaCon) for language understanding, which utilizes the intrinsic ability of class labels to perform hard positive/negative mining.", "method": "The authors developed three contrastive objectives - multi-head instance-centered contrastive loss (ICL), label-centered contrastive loss (LCL), and a label embedding regularizer (LER). LaCon is tested with existing pre-trained language models and benchmarked on popular datasets of GLUE and CLUE as well as FewGLUE and FewCLUE for few-shot and data imbalance settings.", "outcome": "LaCon achieved improvements of up to 4.1% on the GLUE and CLUE benchmarks and up to 9.4% improvement on FewGLUE and FewCLUE in few-shot and data imbalance settings.", "future_impact": "As LaCon does not require any specialized network architecture or extra data augmentation, the authors anticipate that it can be easily implemented in existing powerful pre-trained language models, potentially enhancing their performance.", "venue": "NAACL", "year": 2022, "title": "Label Anchored Contrastive Learning for Language Understanding"} +{"pid": "628304435aee126c0f6ecf0f", "context": "Evaluating natural language generation (NLG) systems is difficult due to the many ways similar things can be expressed in text and the need to assess varying quality criteria depending on the deployment setting.", "key_idea": "The authors propose that the goals, assumptions and constraints of practitioners -- which inform what, when and how NLG systems are evaluated -- are often partially stated, implicitly stated or not stated at all.", "method": "A formative semi-structured interview study of NLG practitioners (N=18) was combined with a survey study of a broader sample of practitioners (N=61) to surface goals, community practices, assumptions and constraints that shape NLG evaluations.", "outcome": "Through the interview study and the survey, the authors detail how NLG practitioners' goals, community practices, assumptions, and constraints shape the evaluation of NLG systems, including their implications and how they embody ethical considerations.", "future_impact": "The authors' results prompting a rethinking of how the evaluation of NLG systems is conducted could potentially impact the development and deployment of more effective and ethical NLG systems in the future.", "venue": "NAACL", "year": 2022, "title": "Deconstructing NLG Evaluation: Evaluation Practices, Assumptions, and Their Implications"} +{"pid": "634d80f190e50fcafd4ef4cb", "context": "Deep Learning techniques have been increasingly used for Automatic Text Scoring in education but they often lack the ability to explain and justify their predictions, leading to a decrease in their trustworthiness.", "key_idea": "This study aims to explore whether Deep Learning-based graders identify the same important words as human graders when grading short answer questions.", "method": "A user study was conducted where human graders manually annotated important words in assessing answer quality, and the overlap with words identified by DL-based graders was measured. Additionally, a randomized controlled experiment was run to investigate the impact of DL-based graders highlighting important words on human grading.", "outcome": "Deep Learning-based graders showed a certain degree of alignment with human graders in identifying important words, irrespective of whether they agreed on the quality of an answer. Highlighting important words detected by Deep Learning-based graders was found to be a potential method of assisting human grading, although deeper investigation is required.", "future_impact": "The findings may pave the way for enhancements in Automatic Text Scoring systems, potentially making them more explainable and trustworthy. Further studies might also help understand how human graders utilize highlighted important words, which can optimize the collaboration between machine graders and humans.", "venue": "NAACL", "year": 2022, "title": "Do Deep Neural Nets Display Human-like Attention in Short Answer Scoring?"} +{"pid": "08efb45d-0a70-46ed-a4b0-ba4f17af5f74", "context": "Current systems in Natural Language Processing (NLP) heavily depend on manually annotated datasets, which are expensive to construct. Very little work adequately exploits unannotated data such as discourse markers between sentences because of data sparseness and ineffective extraction methods.", "key_idea": "The authors propose a new method to automatically discover sentence pairs with relevant discourse markers, allowing for exploiting unannotated data. This can be applied to massive amounts of data.", "method": "The proposed method was applied to create a dataset containing 174 discourse markers with at least 10,000 examples each. The data was then used for learning transferable sentence embeddings.", "outcome": "Even though sentence representation learning through prediction of discourse markers yields state of the art results across different transfer tasks, it\u2019s not clear that the models made use of the semantic relation between sentences.", "future_impact": "The results indicate that there is still room for improvements in learning semantic relations between sentences through discourse markers, pointing to potential further research in this area.", "venue": "NAACL", "year": 2019, "title": "Mining Discourse Markers for Unsupervised Sentence Representation Learning"} +{"pid": "a6e09940-5850-4f08-80c7-bbcdc58131ba", "context": "Multi-document Summarization (MDS) is of great value, with many scoring models proposed to select suitable sentences to form a summary. Among these, clustering-based methods are popular.", "key_idea": "This paper presents a new unified sentence scoring model for MDS that simultaneously measures representativeness and diversity in sentence selection.", "method": "They evaluate their proposed model using experiments on the DUC04 dataset, comparing it against the DUC04 best method and existing clustering-based methods.", "outcome": "Experimental results showed that the proposed MDS method outperformed the DUC04 best method and existing clustering-based methods, generating results closely comparable to state-of-the-art generic MDS methods.", "future_impact": "The new MDS method has two advantages: it is the first to use the density peaks clustering algorithm, which is effective and fast, and it does not require external resources like Wordnet and Wikipedia or complex language parsing algorithms. This makes it easy to reproduce and deploy in a real environment.", "venue": "NAACL", "year": 2015, "title": "Clustering Sentences with Density Peaks for Multi-document Summarization"} +{"pid": "b1f8b28b-4d71-4dae-89f7-6374feb54b92", "context": "The team participated in SemEval-2016 Task 4 (Sentiment Analysis in Twitter), specifically the Message Polarity Classification subtask.", "key_idea": "The team designed a system that is a weighted ensemble consisting of two subsystems: one using manually crafted features and the second using features based on word embeddings.", "method": "The proposed system was tested through participation in SemEval-2016 Task 4, where its performance was ranked against 34 other teams.", "outcome": "The proposed system performed well, being ranked 5th among the 34 participating teams in the sentiment analysis competition.", "future_impact": "The source code of the system is made publicly available, potentially enabling other researchers to build upon the methods used and improve further.", "venue": "NAACL", "year": 2016, "title": "aueb.twitter.sentiment at SemEval-2016 Task 4: A Weighted Ensemble of SVMs for Twitter Sentiment Analysis"} +{"pid": "5f64817891e011f934ad2493", "context": "Recent work has shown that short, uninterpretable input fragments can yield high confidence and accuracy in neural models. These fragments are referred to as Minimal Prediction Preserving Inputs (MPPIs). However, the reasons for the existence of MPPIs remain in question.", "key_idea": "This study sets out to investigate various hypotheses explaining the existence of MPPIs in question answering, such as poor calibration of neural models, lack of pretraining, and potential dataset biases.", "method": "The authors examine the impact of random training seed, model architecture, pretraining, and training domain on the invariance of MPPIs. Additionally, they assess the transferability of MPPIs across domains, as well as the effect of penalizing over-confidence on MPPIs.", "outcome": "The observed results show that MPPIs remain significantly invariant regardless of the considered factors. They also showcase remarkable transferability between domains, closing half the gap between the performance of models on short and original queries. Penalizing over-confidence on MPPIs does not enhance generalization or robustness against adversarial attacks.", "future_impact": "The authors anticipate their study will stimulate further systematic analysis of model behavior outside of the distribution of human-interpretable examples, as the interpretability of MPPIs is found to be insufficient in characterizing model generalization capacity.", "venue": "NAACL", "year": 2020, "title": "On the Transferability of Minimal Prediction Preserving Inputs in Question Answering"} +{"pid": "7277c1c5-1e6d-4337-afb2-2623f5153adf", "context": "When humans communicate via natural language, they often use metalanguage to clarify their meanings, one key aspect of which is the mention of words and phrases, as distinguished from their use. However, there is a gap in automatically recognizing this phenomenon.", "key_idea": "This study is an ongoing work on identifying and categorizing instances of language-mention, with the end-goal of building a system capable of automatic recognition.", "method": "The researchers define language-mention and gather a corpus of instances from Wikipedia as resources for developing their system.", "outcome": "The paper discusses the definition of language-mention and provides a corpus collected from Wikipedia, but it doesn't present any experimental outcomes or system performance yet as the work is ongoing.", "future_impact": "The future direction of this project will potentially contribute to the automation of language-mention recognition, leading to more effective communication and understanding of natural language.", "venue": "NAACL", "year": 2010, "title": "Distinguishing Use and Mention in Natural Language"} +{"pid": "60af771e9e795e6b8e55c7e0", "context": "Lemmatization aims to handle sparse inflection data and most prior work on ML based lemmatization has targeted high resource languages with readily available datasets. For languages with less linguistic information, particularly those with complex morphologies, machine learning based lemmatizers are the preferred approach. Yet, these languages often have fewer word forms available, making them 'low-resource'.", "key_idea": "This paper focuses on lemmatisation for low-resource, morphologically rich scheduled Indian languages using neural methods. The authors also investigate how to optimize lemmatization in this low resource context.", "method": "The authors conduct tests to analyse the variance in monolingual model performance while varying the corpus size and contextual morphological tag data for training.", "outcome": "The results show that monolingual approaches with data augmentation can offer competitive accuracy even in a low resource setting.", "future_impact": "These findings bode well for future natural language processing applications in low-resource settings.", "venue": "NAACL", "year": 2021, "title": "How low is too low? A monolingual take on lemmatisation in Indian languages"} +{"pid": "5bcfd9da-823e-43b3-9c9b-af798d534a49", "context": "Improving the accuracy of many natural language processing tasks often involves a reranking step, selecting a single output from a list of candidate outputs generated by a baseline system.", "key_idea": "The authors propose a novel family of reranking algorithms that learn separate low-dimensional embeddings of input and output spaces, with prediction becoming a low-dimensional nearest-neighbor search.", "method": "The proposed reranking algorithms were tested on a part-of-speech tagging task in four different languages.", "outcome": "Experiments show significant improvements over a baseline decoder and existing reranking approaches, proving the efficiency of the proposed algorithms.", "future_impact": "In the future, this effective alternative approach to reranking could improve a wide range of Natural Language Processing tasks, with its ability to automatically learn relationships between inputs and outputs.", "venue": "NAACL", "year": 2012, "title": "Low-Dimensional Discriminative Reranking"} +{"pid": "70987110-8131-4ef2-a85b-f964e990d6c2", "context": "Existing question-answering systems rely on fixed ontology or domain models that predetermine what users can say and do not learn from conversational dialogs.", "key_idea": "The authors propose a system that learns to relate concepts in science questions to propositions in a fact corpus from conversational dialogs, stores new concepts and relations in a knowledge graph (KG), and uses the KG to solve questions.", "method": "The authors develop strategies for dialog-based knowledge acquisition, which are compared against a baseline and general knowledge sources.", "outcome": "The relation-based strategies of the proposed system complete more successful dialogs than a query expansion baseline and are more effective for solving science questions than relations from general knowledge sources.", "future_impact": "The authors believe their method is practical enough to generalize to other domains, potentially expanding the applicability of dialog-based knowledge acquisition for question-answering systems.", "venue": "NAACL", "year": 2015, "title": "Learning Knowledge Graphs for Question Answering through Conversational Dialog"} +{"pid": "6269f8245aee126c0f049c21", "context": "Deep learning models in NLP often fail to generalize due to overfitting to low-level correlations between simple features and specific output labels. Previous efforts have attempted to mitigate this by balancing datasets, even going as far as to eliminate single-word correlations.", "key_idea": "The authors express the opinion that increasingly-powerful models continue to exploit smaller spurious correlations, rendering dataset balancing insufficient. They suggest that a perfectly balanced dataset might lose important signals of common sense and world knowledge.", "method": "As an opinion paper, there's no specific experiment or validation methodology mentioned in the abstract.", "outcome": "The authors argue that even balancing all single-word features is not enough to mitigate all spurious correlations found by increasingly powerful models.", "future_impact": "The authors highlight alternative solutions beyond dataset balancing that could potentially lead to more meaningful improvements in model generalization. These include enriching datasets with more context, allowing models to abstain and interact with users, and shifting from large-scale fine-tuning to zero- or few-shot setups.", "venue": "NAACL", "year": 2022, "title": "On the Limitations of Dataset Balancing: The Lost Battle Against Spurious Correlations."} +{"pid": "d646cb43-2f4f-4cb4-980e-5a95f0bdc46a", "context": "Current automatic machine translation evaluation methods do not take into account the predicted compositionality scores of multiword expressions.", "key_idea": "The authors introduce the first attempt to integrate predicted compositionality scores of multiword expressions into automatic machine translation evaluation, specifically English noun compounds into the TESLA machine translation evaluation metric.", "method": "The authors tested their idea by integrating compositionality scores for English noun compounds into the TESLA machine translation evaluation metric.", "outcome": "The attempt was marginally successful.", "future_impact": "The authors speculate that a larger-scale attempt to integrate compositionality scores into machine translation evaluation may likely have a greater impact.", "venue": "NAACL", "year": 2015, "title": "The Impact of Multiword Expression Compositionality on Machine Translation Evaluation"} +{"pid": "c3a66aa5-24e9-4333-93de-8255ba14a2ef", "context": "Arabic inflectional morphology requires infixation, prefixation and suffixation, leading to a large space of morphological variation, which causes complexities in Arabic morphology generation.", "key_idea": "The paper proposes a novel approach to reducing the complexity of Arabic morphology generation using discrimination trees and transformational rules, by decoupling the problem of stem changes from that of prefixes and suffixes.", "method": "While the paper focuses on the application of the proposed approach to hollow verbs, it also discusses its general application.", "outcome": "Implementing the proposed method resulted in a significant reduction in the number of rules required for Arabic morphology generation, as much as a factor of three for certain verb types.", "future_impact": "While the paper focuses on hollow verbs, the wider applicability of the approach discussed in the paper suggests potential for further research and advances in reducing the complexity of morphology generation for other languages and verb types.", "venue": "NAACL", "year": 2000, "title": "Arabic morphology generation using a concatenative strategy"} +{"pid": "60af76679e795e6b8e55c706", "context": "The Transformer translation model has been greatly explored using probing-based approaches, mostly focusing on using or probing source linguistic features in the encoder. However, the evolution of word translation in Transformer layers has not been deeply investigated and it's generally assumed that encoder layers capture source information while decoder layers handle translation.", "key_idea": "The paper challenges the traditional understanding of the Transformer model, suggesting that translation actually begins in the encoder layers and even in the input embeddings. It also suggests that some lower decoder layers might not significantly contribute to decoding.", "method": "The authors use a probing approach, projecting the representations of the layer being analyzed to the final trained and frozen classifier level of the Transformer decoder to measure word translation accuracy.", "outcome": "The findings support a change in the configuration of the Transformer model, where increasing the number of encoder layers and decreasing the number of decoder layers can enhance the decoding speed without a loss in translation quality. In particular, an 18-4 deep encoder configuration shows an improvement in translation quality by +1.42 BLEU (En-De) and a speed-up of 1.4.", "future_impact": "findings could inform about efficient configurations of Transformer models leading to enhanced translation quality and improved speed, enabling faster, more effective translation model development.", "venue": "NAACL", "year": 2021, "title": "Probing Word Translations in the Transformer and Trading Decoder for Encoder Layers"} +{"pid": "627332775aee126c0f18d649", "context": "Existing word embeddings, which are fundamental in natural language processing, are high-dimensional and consume significant computational resources.", "key_idea": "The authors propose WordTour, an unsupervised one-dimensional word embedding solution. They adopt an approach that decomposes the desiderata of word embeddings into two parts: completeness and soundness, and focus on soundness.", "method": "The effectiveness of the proposed method WordTour is evaluated through a user study and document classification.", "outcome": "The proposed one-dimensional word embeddings method, WordTour, has been confirmed to be effective based on the results from user study and document classification test.", "future_impact": "WordTour provides a minimal, extremely efficient means to handle word embeddings, which could help in improving efficiency in natural language processing tasks.", "venue": "NAACL", "year": 2022, "title": "Word Tour: One-dimensional Word Embeddings via the Traveling Salesman Problem"} +{"pid": "609a34e491e011a44725cbf0", "context": "Grammatical Error Correction (GEC) is aimed at correcting writing errors and ensuring accuracy in writing. However, existing GEC models often produce spurious corrections or fail to detect many errors. Even though well-trained GEC models can generate high-quality hypotheses, current models overlook possible GEC evidence from different hypotheses.", "key_idea": "This paper presents the Neural Verification Network (VERNet), a novel method for GEC quality estimation using multiple hypotheses. VERNet uses a reasoning graph to establish interactions among the hypotheses and employs two kinds of attention mechanisms to propagate GEC evidence and verify the quality of generated hypotheses.", "method": "The authors tested VERNet using experiments on four GEC datasets, in which the generated hypotheses were evaluated for their quality of grammatical error correction.", "outcome": "The results show that VERNet achieves state-of-the-art grammatical error detection performance as well as the best quality estimation results. It also significantly improves GEC performance by reranking hypotheses.", "future_impact": "The accuracy and effectiveness of VERNet in improving GEC models could potentially enhance these models for more reliable and efficient GEC tasks in future.", "venue": "NAACL", "year": 2021, "title": "Neural Quality Estimation with Multiple Hypotheses for Grammatical Error Correction"} +{"pid": "635bca9890e50fcafd33abc0", "context": "Word substitution-based adversarial attacks pose a threat to BERT language models, and improving their robustness against these attacks is currently a research challenge.", "key_idea": "The authors propose an approach that uses adversarial perturbations in a self-supervised contrastive learning context, generating word-level adversarial attacks as hard positives during this learning process.", "method": "The authors use contrastive learning to create word-level adversarial examples on-the-fly. This method is then tested against four different word substitution-based adversarial attacks, and in combination with adversarial training.", "outcome": "The proposed method improves the robustness of BERT against four different word substitution-based adversarial attacks. Combining this method with adversarial training yields higher robustness than adversarial training alone.", "future_impact": "Since this method improves BERT's robustness using only unlabeled data, it could enable the use of larger text datasets to train robust language models against word substitution-based adversarial attacks.", "venue": "NAACL", "year": 2022, "title": "Self-Supervised Contrastive Learning with Adversarial Perturbations for Defending Word Substitution-based Attacks."} +{"pid": "5f7ae4aa91e011983cc81da5", "context": "The desire to create agents that can both communicate and act in pursuit of a goal has led to the development of the LIGHT\u2014an extensive crowd-sourced fantasy text-game. However, this game hasn't adequately facilitated training agents that can speak and act to achieve a goal.", "key_idea": "The authors extend the LIGHT with a dataset of quests that contain natural language motivations combined with in-game goals and human demonstrations. They also propose a reinforcement learning system that incorporates large-scale language modeling and commonsense reasoning-based pre-training.", "method": "The authors use a reinforcement learning system with a factorized action space of action commands and dialogue. The model is initially pre-trained with large-scale language modeling and commonsense reasoning. Zero-shot evaluations are conducted using held-out human expert demonstrations.", "outcome": "The zero-shot evaluations demonstrated that the proposed agents can behave consistently and converse naturally aligning with their motivations.", "future_impact": "The system introduced could have implications for the development of more sophisticated AI agents that can both act and communicate effectively in complex, goal-driven scenarios.", "venue": "NAACL", "year": 2020, "title": "How to Motivate Your Dragon: Teaching Goal-Driven Agents to Speak and Act in Fantasy Worlds"} +{"pid": "608bdf3991e0112fc4e65b96", "context": "Currently, models in the NLP domain quickly achieve outstanding performance on benchmark tasks but fail on simple challenge examples and real-world applications.", "key_idea": "To address this issue, the authors introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. It supports human-and-model-in-the-loop dataset creation where annotators create examples that a model will misclassify, but another person will not.", "method": "The authors demonstrate the Dynabench platform on four initial NLP tasks and discuss potential objections to dynamic benchmarking as a new standard for the field.", "outcome": "The application of Dynabench has been demonstrated on four NLP tasks and with this, the authors illustrate the platform's potential and how dataset creation, model development, and model assessment can directly inform each other.", "future_impact": "The authors envision Dynabench to foster more robust and informative benchmarks in the future, potentially establishing dynamic benchmarking as a new standard for the NLP field.", "venue": "NAACL", "year": 2021, "title": "Dynabench: Rethinking Benchmarking in NLP"} +{"pid": "612857d75244ab9dcb5c7647", "context": "Emotion Recognition in Conversation (ERC) has gained importance with the growing use of interactive machines. However, many studies reflect the dialogue context for more accurate results because emotion recognition in conversation can be inaccurate if previous utterances are not taken into account.", "key_idea": "The authors introduce CoMPM, which is the combination of a context embedding module (CoM) and a pre-trained memory module (PM), which tracks memory of the speaker's past utterances within the context, in order to improve the accuracy of emotion recognition.", "method": "Experiments were conducted on multi-party datasets (MELD, EmoryNLP) as well as dyadic-party datasets (IEMOCAP, DailyDialog).", "outcome": "The results show that the approach of using a context embedding module combined with a tracking memory module achieves competitive performance on all tested datasets.", "future_impact": "The study advances the field of Emotion Recognition in Conversation, allowing for more human-like sympathetic conversations in interactive machines.", "venue": "NAACL", "year": 2022, "title": "CoMPM: Context Modeling with Speaker's Pre-trained Memory Tracking for Emotion Recognition in Conversation"} +{"pid": "ad668250-3861-4d00-8376-037117ccaf17", "context": "Tactile maps are crucial for visually impaired individuals, and their usability can be substantially enhanced by strategically employing spatial language in assisting utterances.", "key_idea": "This study suggests a system that generates assisting utterances based on a semantic conceptualization of map user's movements, aiming to improve the efficiency of tactile-map reading.", "method": "A prototype of the proposed system was implemented, and its effectiveness was demonstrated through a worked example using input derived from experimental data.", "outcome": "The prototype system was developed and its output displayed a plausible solution, indicating that the system could generate assisting utterances as required.", "future_impact": "The approach may further enhance the understanding and usability of tactile maps for blind and visually impaired individuals, though the paper does not explicitly forecast future research directions.", "venue": "NAACL", "year": 2012, "title": "Generating Situated Assisting Utterances to Facilitate Tactile-Map Understanding: A Prototype System"} +{"pid": "c8a7ba61-15cb-4c56-8686-32be872273b6", "context": "The challenge of linking English FrameNet lexical entries to the appropriate Chinese word senses has yet to be effectively addressed.", "key_idea": "The authors propose a method of automatically constructing an English-Chinese bilingual FrameNet using a bilingual lexicon, FrameNet frame context, and HowNet taxonomy structure, without needing any manual mapping between FrameNet and HowNet semantic roles.", "method": "The approach involves coercing the English FrameNet into Chinese using the bilingual lexicon, frame context in FrameNet, and the taxonomy structure in HowNet.", "outcome": "The authors' method achieves an 82% average F-measure for the most ambiguous lexical entries.", "future_impact": "The produced bilingual FrameNet can be used to improve Machine Translation and cross-lingual information retrieval systems.", "venue": "NAACL", "year": 2004, "title": "Automatic construction of an English-Chinese bilingual FrameNet"} +{"pid": "282a1899-9373-42ac-97a6-56437bff6792", "context": "Decoding for similarity-based loss functions is a challenging task, particularly with respect to efficient realization, as previously demonstrated by Tromble et al. (2008). Existing methods are either inefficient or approximate solutions.", "key_idea": "The paper introduces efficient algorithms for expected similarity maximization, a concept that coincides with minimum Bayes decoding for similarity-based loss functions. The algorithms are appropriate for sequence kernels in a general class of positive definite symmetric kernels.", "method": "The authors discuss both a general algorithm and another efficient algorithm that is applicable in a common unambiguous scenario. The proposed algorithms are applied to machine translation and tested on several data sets.", "outcome": "Empirical results reveal that the proposed methods present a substantial speed-up, up to two orders of magnitude, as compared to the method of Tromble et al. (2008), and even three times faster compared to an approximate algorithm designed for the same task.", "future_impact": "The speed improvements achieved with these efficient algorithms set the stage for future explorations of more suitable or optimal kernels for specific tasks.", "venue": "NAACL", "year": 2010, "title": "Expected Sequence Similarity Maximization"} +{"pid": "512a69db-7eb7-4f81-801a-40ad9991d4c1", "context": "The impact of corpus size on case frame acquisition for discourse analysis, particularly in Japanese, is not well-understood.", "key_idea": "The authors investigate the relationship between the size of a corpus and the performance of case frame construction within the context of discourse analysis in Japanese.", "method": "The authors constructed case frames from six corpora of different sizes, up to a maximum of 100 billion words, and applied these case frames to syntactic and case structure analysis, and zero anaphora resolution.", "outcome": "The research found that larger corpora yielded better results in the construction of case frames, and that the performance did not plateau even when the corpus size reached 100 billion words.", "future_impact": "This study could encourage further research into the effects of corpus size on various aspects of language analysis and processing.", "venue": "NAACL", "year": 2009, "title": "The Effect of Corpus Size on Case Frame Acquisition for Discourse Analysis"} +{"pid": "fb831744-a2a8-4c90-9ee8-1e73273998b4", "context": "Natural language parsing can be improved through the use of machine learning techniques, but traditional parsers did not capitalize on all available solutions.", "key_idea": "The authors apply bagging and boosting, two machine learning techniques, to natural language parsing with a trainable statistical parser.", "method": "The authors conducted experiments with techniques using a trainable statistical parser and conducted error analysis of the result of the boosting technique.", "outcome": "The best resulting system from applying bagging and boosting delivered an improvement in F-measure equivalent to that achieved by doubling the corpus size. In addition, their method revealed some inconsistent annotations in the Penn Treebank.", "future_impact": "The strategies employed in this work may also be useful for semi-automatically finding inconsistent annotations in treebanks.", "venue": "NAACL", "year": 2000, "title": "Bagging and boosting a treebank parser"} +{"pid": "7e191242-61b4-4f0c-818d-df5bfcf18694", "context": "The identification and classification of biomedical terms in research publications is a complex and crucial task.", "key_idea": "The authors propose a novel approach based on the Perceptron HMM algorithm to effectively identify and classify each important biomedical term into a biomedical concept class.", "method": "The authors implemented the proposed system and trained it on 2,000 Medline abstracts and tested it on 404 unseen Medline abstracts.", "outcome": "The proposed system achieves a 68.6% F-measure, performing close to the state-of-the-art with only a small feature set.", "future_impact": "The authors' approach proves the potential of the Perceptron HMM algorithm in dealing with complex, interdependent features which could be adopted in various similar tasks in the future.", "venue": "NAACL", "year": 2006, "title": "Biomedical Term Recognition with the Perceptron HMM Algorithm"} +{"pid": "6075758d91e0110f6fe683a8", "context": "While new words are constantly being introduced to communities, not all of these words persist in the community's lexicon. The role of social networks in contributing to lexical change has been understudied thus far.", "key_idea": "The authors have undertaken a large-scale analysis of over 80k neologisms in 4420 online communities across a decade, with the key focus being the influence of a community's network structure on lexical change.", "method": "The study uses Poisson regression and survival analysis to understand the relationships between a community's network structure and lexical change.", "outcome": "The study found that the community's network structure significantly contributes to lexical change. Apart from overall size, properties such as dense connections, the lack of local clusters, and more external contacts promote lexical innovation and retention. Topic-based communities do not experience strong lexical levelling, but accommodate more niche words.", "future_impact": "The findings of this work support the sociolinguistic hypothesis that lexical change is partially shaped by the structure of the underlying network and uncover new insights specific to online communities, which could inform future research in the field.", "venue": "NAACL", "year": 2021, "title": "The structure of online social networks modulates the rate of lexical change"} +{"pid": "e6cafd52-dc18-441a-bf86-e3bcfd4b3758", "context": "In text categorization, feature selection plays a critical role but often struggles with the issue of redundancy. Traditional methods used for feature selection like conventional information gain and greedy feature selection methods are not always optimal.", "key_idea": "The authors propose a new method for feature selection, the MMR-based (Maximal Marginal Relevance) feature selection, that aims at reducing redundancy between features while maintaining high information gain.", "method": "The effectiveness of the MMR-based feature selection method is evaluated by comparing it with Koller and Sahami's feature selection method and conventional information gain, commonly used methods for feature selection in text categorization.", "outcome": "According to empirical results, the MMR-based feature selection method proves to be more effective than the other compared methods, and in some cases, it even improves the performance of conventional machine learning algorithms over SVM, which is known for its classification accuracy.", "future_impact": "The MMR-based feature selection might significantly impact the approach to text categorization by addressing the challenge of feature redundancy and in some situations, could also potentially enhance the performance of conventional machine learning algorithms.", "venue": "NAACL", "year": 2004, "title": "MMR-based feature selection for text categorization"} +{"pid": "634d80f190e50fcafd4ef48b", "context": "Event Detection in non-English languages has been significantly less explored compared to English. Current multilingual Event Detection datasets are often small, not publicly accessible, and cover only a few popular languages.", "key_idea": "The authors introduce a new large-scale multilingual dataset for Event Detection called MINION, which consistently annotates events for eight different languages, including five languages not supported by existing datasets.", "method": "The authors perform extensive experiments and analysis on the MINION dataset to demonstrate the challenges and transferability of Event Detection across different languages.", "outcome": "The provided analysis and experimentations on the MINION dataset showcase the challenges and transferability aspects of Event Detection in different languages.", "future_impact": "This paper calls for more research effort in the area of Event Detection for non-English languages, a topic that could be facilitated by the introduced MINION dataset.", "venue": "NAACL", "year": 2022, "title": "MINION: a Large-Scale and Diverse Dataset for Multilingual Event Detection"} +{"pid": "61b80b6c5244ab9dcbf48c97", "context": "State-of-the-art dialogue models often have factual accuracy and self-contradiction issues. They have also been observed to fail in maintaining character identity throughout discourse and may take on the role of their interlocutor.", "key_idea": "This paper formalizes and quantifies the issue of dialogue models failing to consistently maintain character identity, and also proposes discriminative models to recognize who is speaking as a potential solution.", "method": "The authors conduct experiments involving human evaluations to assess the problem. They also use discriminative models trained specifically to recognize the speaker and measure their effectiveness. A variety of mitigation methods, including changes to model architecture, training protocol, and decoding strategy were evaluated.", "outcome": "The results show that the problem occurs as initially suspected. The models that were trained specifically to recognize the speaker performed well and they reduced the mistaken identity issues by nearly 65% according to human annotators while improving engagement.", "future_impact": "Despite the significant reduction in mistaken identity issues, maintaining character identity in dialogue models remains a challenging problem, indicating the need for further research in this area.", "venue": "NAACL", "year": 2021, "title": "Am I Me or You? State-of-the-Art Dialogue Models Cannot Maintain an Identity."} +{"pid": "634d80f190e50fcafd4ef328", "context": "Humans resolve referents in their environment using natural language, vision, and context. Challenges arise with underspecified language or the presence of multiple possible referents. Current coreference tools in natural language processing struggle to handle such ambiguous cases.", "key_idea": "The study examines how pragmatic modulators, specifically social norms applicable in the given context, are important for correctly interpreting referents in scenarios with potential ambiguity.", "method": "The authors conduct a human subjects experiment to demonstrate how the interpretation of referring expressions is influenced by social norms within the given context.", "outcome": "The study confirms that social norms applicable in the given context significantly influence the interpretation of referring expressions.", "future_impact": "Implications of this work suggest improvements in the capabilities of assistive robots, as they will regularly need to resolve referents in their environment.", "venue": "NAACL", "year": 2022, "title": "Social Norms Guide Reference Resolution"} +{"pid": "36655bd6-9fa5-4cad-a126-a58d2a8cafd5", "context": "Bullying, in both physical and cyber worlds, is a recognized national health issue among adolescents, but previous studies have been hindered by data scarcity. In addition, computational methodologies for studying bullying have mostly been limited to cyberbullying.", "key_idea": "The authors propose that social media is a valuable and rich data source for studying bullying in both physical and cyberspace, and present several key NLP tasks as potential methodologies, including text classification, role labeling, sentiment analysis, and topic modeling.", "method": "The authors employ off-the-shelf NLP solutions to establish baseline results on the proposed tasks in order to explore the application of social media data in studying bullying.", "outcome": "The authors demonstrate that by applying basic NLP techniques, social media can indeed be a useful data source for both the physical and cyber examination of bullying, although specific results are not detailed in the abstract.", "future_impact": "The study opens up this line of investigation to the NLP community, encouraging improved models and techniques for studying bullying through social media data in the future.", "venue": "NAACL", "year": 2012, "title": "Learning from Bullying Traces in Social Media"} +{"pid": "ccf1eeb2-c9c0-41bd-b584-cffee9dde777", "context": "The Whiteboard project was initiated to integrate deep and shallow natural language processing components to take advantage of their synergies. This was done to increase robustness, guide the search space, and reduce processing time.", "key_idea": "The authors developed a central integration tool known as the Whiteboard Annotation Transformer (WHAT) to facilitate the integration of natural language processing components.", "method": "The WHAT is an XSLT-based tool, and its benefits and practical applications are discussed using examples of XSL transformation of shallow and deep annotations in integrated architectures.", "outcome": "The Whiteboard Annotation Transformer has been shown to be effective in enabling NLP component integration, with its open, portable infrastructure being well-suited for and applicable to developing hybrid NLP architectures and applications.", "future_impact": "The open and portable nature of WHAT could enable wider use and experimentation in hybrid NLP architectures and applications.", "venue": "NAACL", "year": 2003, "title": "WHAT: an XSLT-based infrastructure for the integration of natural language processing components"} +{"pid": "ab913b55-340c-431a-ae53-15416c73c272", "context": "Currently, the scoring of the Evidence dimension in a writing assessment is done based on information provided by experts, which is labor-intensive and limits scalability.", "key_idea": "The study proposes an automatic method to extract multiword topical components from text, which aims to reduce expert effort and enhance the scalability of an automatic scoring system.", "method": "The authors conduct experiments to automatically extract data-driven topical components and compare the scoring performance using these components.", "outcome": "Experimental results demonstrate that using automatically extracted data-driven topical components for scoring is promising.", "future_impact": "The proposed approach can potentially improve the efficiency and scalability of automatic scoring systems in response-to-text writing assessments.", "venue": "NAACL", "year": 2016, "title": "Automatically Extracting Topical Components for a Response-to-Text Writing Assessment"} +{"pid": "c0d3a9bb-1d66-4e3c-aae4-35b3395a6df2", "context": "Improving science knowledge in students can involve the identification of misconceptions and remediation with relevant science webpages. However, the automation of this process in an interactive essay writing tutor was not adequately addressed.", "key_idea": "The researchers propose an interactive essay writing tutor that uses student essays to identify misconceptions and recommend appropriate science webpages to correct those misconceptions. The system involves five components: identifying core science concepts, determining pedagogical sequences, identifying student misconceptions, aligning misconceptions to science concepts, and recommending webpages.", "method": "The researchers developed and evaluated models for each component of their proposed system.", "outcome": "Initial models for each component of the tutor were provided and evaluated, but specific results or measurements were not stated in the abstract.", "future_impact": "The proposed system could revolutionize how misconceptions in student essays are identified and corrected, leading to better science knowledge acquisition.", "venue": "NAACL", "year": 2012, "title": "Identifying science concepts and student misconceptions in an interactive essay writing tutor"} +{"pid": "cc4d8504-c855-45c5-9a63-62502a218e16", "context": "Statistical machine translation usually does not consider the loss of translation errors in the decoding process.", "key_idea": "The paper introduces the concept of Minimum Bayes-Risk (MBR) decoding for statistical machine translation, which aims to minimize the expected loss of translation errors under various loss functions measuring translation performance.", "method": "The paper proposes a hierarchy of loss functions that utilize different levels of linguistic data, from word strings and word-to-word alignments from a machine translation system to syntactic structure from source and target language sentences. The performance of the MBR decoders was then evaluated on a Chinese-to-English translation task.", "outcome": "It was found that MBR decoding can effectively be used to adjust statistical machine translation performance in accordance with specific loss functions.", "future_impact": "The proposed MBR decoding may lead to customized translation systems that can be optimized for specific loss functions, thus enhancing the efficiency and usefulness of statistical machine translation.", "venue": "NAACL", "year": 2004, "title": "Minimum Bayes-risk decoding for statistical machine translation"} +{"pid": "8fa33a17-1e2a-4017-ade2-4d452a582c77", "context": "Previously, the task of question answering (QA) over Freebase on the WEBQUESTIONS dataset was complex and time-consuming.", "key_idea": "The authors found that 85% of all questions can be directly answered via a single binary relation. Therefore, they transform the task into slot-filling for tuples: predicting relations to get answers given a question\u2019s topic.", "method": "The authors design efficient data structures to identify question topics from 46 million Freebase topic names, without using any NLP processing tools. A new lean QA system is proposed for real-time responses.", "outcome": "The proposed system performs exceptionally, answering two thousand questions in 51 seconds on a laptop during offline batch testing. It also outperforms the previous state-of-the-art by achieving a 7.8% better F1 score.", "future_impact": "The proposed lean QA system could greatly speed up the question answering process using Freebase and prompt further efficiency improvements in similar work.", "venue": "NAACL", "year": 2015, "title": "Lean Question Answering over Freebase from Scratch"} +{"pid": "607d4e8c91e011bf62020907", "context": "Natural language processing (NLP) research combines basic science and applied science. However, the exchange process between basic NLP and applications often emerges naturally, resulting in many innovations going unapplied and many important questions left unstudied.", "key_idea": "The authors propose a new paradigm of Translational NLP, aiming to structure and facilitate the processes by which basic and applied NLP research inform one another.", "method": "The authors discuss how many significant advances in NLP research have emerged from the intersection of basic principles with application needs, and present a conceptual framework outlining the stakeholders and key questions in translational research.", "outcome": "The authors present a conceptual framework outlining the stakeholders and key questions in translational NLP research, serving as a roadmap for developing Translational NLP as a dedicated research area.", "future_impact": "The presented framework identifies general translational principles to facilitate exchange between basic and applied research, implying potential improvements in applying innovations and addressing understudied questions in NLP.", "venue": "NAACL", "year": 2021, "title": "Translational NLP: A New Paradigm and General Principles for Natural Language Processing Research"} +{"pid": "de5dc692-ef19-4031-a292-ce84995f73aa", "context": "Multiword expressions, including nominal compounds, collocational forms, and idioms, are increasingly being recognized for their potential applications in educational technology.", "key_idea": "The paper focuses on how multiword expression identification can be leveraged in different kinds of educational applications, such as automated essay evaluation, and teacher professional development in curriculum development for English language learners.", "method": "The authors detail the applications of multiword expressions in the context of educational technology and discuss how polarity for noun-noun compounds is resolved in a sentiment system which is designed to handle evaluation of argumentation in testtaker writing.", "outcome": "Different applications for multiword expression identification in educational technology were identified and used successfully, the results of which will be further discussed in the full paper.", "future_impact": "The use of multiword expression identification could influence and enhance the development of future educational technologies and may also contribute towards improving automated essay evaluation and teacher professional development.", "venue": "NAACL", "year": 2013, "title": "The Far Reach of Multiword Expressions in Educational Technology"} +{"pid": "1d8a8a59-ccfd-436b-8c74-13b6ef24987d", "context": "Relative clauses have been considered to contain background information, not directly relevant or highly useful in semantic analysis, and are often removed in text summarization as they are viewed to contain non-central content to the topic or discourse.", "key_idea": "The study proposes the importance of distinguishing the Presentational Relative Clause (PRC) construction from other relative clause types, highlighting that the relative clause in PRC contains the assertion of the utterance, contrary to common understanding.", "method": "The method used in this study is linguistic analysis to suggest informative features that may be used in the automatic extraction of PRC constructions.", "outcome": "The study suggests informative features that can be used in automatically extracting PRC constructions.", "future_impact": "Identifying the Presentational Relative Clause construction is believed to be useful for discriminating central information from peripheral, which can be beneficial for applications in text summarization and semantic analysis.", "venue": "NAACL", "year": 2010, "title": "Identifying Assertions in Text and Discourse: The Presentational Relative Clause Construction"} +{"pid": "c1f80d43-de8d-4b86-bb92-8287f0478c10", "context": "As linguistic models become more nuanced and complex, standard inference techniques struggle to keep up. Many of these models are tightly bound and resist common dynamic programming strategies.", "key_idea": "The paper proposes the use of Sequential Monte Carlo (SMC) approaches, specifically particle filters, to approximate complex linguistic models by sampling either sentences or word types and integrating them into a Gibbs sampler for part-of-speech (PoS) inference.", "method": "The authors implement two particle filters and analyze their behavior in comparison with a block sentence sampler, a local token sampler, and a heuristic sampler, which limits inference to a single PoS per word type.", "outcome": "The experiment indicates that particle filters can efficiently approximate a challenging or otherwise intractable sampler. However, a higher posterior likelihood does not necessarily mean better Many-to-One accuracy.", "future_impact": "The outcome suggests that the proposed approach shows potential and that utilizing more advanced particle filters could result in improved performance.", "venue": "NAACL", "year": 2012, "title": "Unsupervised Part of Speech Inference with Particle Filters"} +{"pid": "b97e1428-3a97-4415-b191-d9467606b9e4", "context": "Several areas of biomedical information extraction, such as drug development and cancer genomics, can be quite challenging due to the complex structure and relationships between entities.", "key_idea": "The authors propose a comprehensive framework for biomedical information extraction, which integrates multiple levels of corpus annotation: a Treebank for syntactic structure, a Propbank for predicate-argument structure, and annotation of entities and their interrelations.", "method": "The authors are training statistical taggers using the multi-level annotation for extracting biomedical information, and also using them for improving the annotation process.", "outcome": "The abstract does not provide specific results or outcomes from the methodology applied.", "future_impact": "The authors anticipate that this integrated annotation framework would improve the efficiency and accuracy of biomedical information extraction, although the specific impact is not explicitly stated in the abstract.", "venue": "NAACL", "year": 2004, "title": "Integrated Annotation for Biomedical Information Extraction"} +{"pid": "4ee938d1-41cf-4a57-bff9-18a12c646234", "context": "To avoid non-understandings in Spoken Dialogue Systems (SDS), wizards directly query the application database with the Automated Speech Recognition (ASR) hypothesis (voice search).", "key_idea": "The authors aim to learn how to avoid misunderstandings by investigating how wizards deal with uncertainty in voice search results.", "method": "A Wizard-of-Oz experiment was conducted with multiple wizard subjects, each viewing ASR results for utterances whose interpretation is critical to task success: requests for books by title from a library database.", "outcome": "Wizards were found to be quite successful at selecting the correct title from query results and the most successful wizard could also tell when the query results did not contain the requested title.", "future_impact": "The learned models of the best wizard's behavior, which combine features available to wizards with some that are not, such as recognition confidence and acoustic model scores, may inform future development of voice search capabilities in SDS.", "venue": "NAACL", "year": 2010, "title": "Learning about Voice Search for Spoken Dialogue Systems"} +{"pid": "9e9a7fce-e70d-4309-8e65-fedf9726c380", "context": "Named entity recognition (NER) systems trained on newswire perform poorly when tested on Twitter due to the different nature of the language used.", "key_idea": "The solution proposed involves the construction of special models, leveraging unlabeled tweets to build Brown clusters and word vectors and using annotated newswire data with an importance weighting scheme.", "method": "The authors use well-known techniques to improve Twitter NER performance when given a small set of annotated training tweets and establish new benchmarks on two common test sets.", "outcome": "By adding these constructed word representations to the system, the authors report an improvement of up to 20 F1 in Twitter NER.", "future_impact": "The approach of applying well-understood techniques, namely Brown clusters and word vectors, for other settings like Twitter where the language used differs significantly from newswire, could inspire future research.", "venue": "NAACL", "year": 2015, "title": "The Unreasonable Effectiveness of Word Representations for Twitter Named Entity Recognition"} +{"pid": "62563f855aee126c0f6f3184", "context": "Multimodal data can provide more robust features for sentiment detection than unimodal data. However, most previous research work rarely explores the token-level feature fusion of multimodal data and learning the common features related to sentiment expressed in it.", "key_idea": "The authors propose a technique called Contrastive Learning and Multi-Layer Fusion (CLMLF) that employs token-level feature fusion between text and image data for sentiment detection, in addition to designing two contrastive learning tasks for common sentiment-related feature learning in multimodal data.", "method": "The authors first encode text and images to obtain hidden representations, then use a multi-layer fusion module to align and combine token-level features of text and images. The method is evaluated on three publicly available multimodal datasets.", "outcome": "The proposed CLMLF method demonstrates its effectiveness in multimodal sentiment detection and outperforms existing techniques on the experimented multimodal datasets.", "future_impact": "The developed method could pave the way for more sophisticated sentiment detection models that fully utilize all available multimodal data for better understanding underlying sentiments.", "venue": "NAACL", "year": 2022, "title": "CLMLF: A Contrastive Learning and Multi-Layer Fusion Method for Multimodal Sentiment Detection."} +{"pid": "3845f166-8e35-4414-97ab-4e9e2cf282b4", "context": "Previous methods for semantic orientation classification failed to classify phrases composed of unseen words.", "key_idea": "The authors present a novel method for extracting semantic orientations of adjective-noun pairs, classifying them as positive, negative, or neutral. It constructs a lexical network of similar/related words and uses the Potts model to estimate the semantic orientations of the phrases.", "method": "The model's practical effectiveness is assessed through empirical evaluation, with an emphasis on its ability to classify phrases composed of unseen words.", "outcome": "Empirical evaluations showcase the effectiveness of the proposed method. The model's capability to classify phrases consisting of unseen words is evidence.", "future_impact": "The method's capability to process and assess phrases containing previously unseen words potentially paves the way for more comprehensive and versatile semantic orientation solutions.", "venue": "NAACL", "year": 2007, "title": "Extracting Semantic Orientations of Phrases from Dictionary"} +{"pid": "978ab58a-26ff-4519-94fc-f9cd1f0c3b90", "context": "Existing translation-based knowledge graph embeddings do not accurately represent transitive and symmetric relations as they do not take into account the role of entities in triples.", "key_idea": "The paper introduces a novel approach involving role-specific projection where entities are mapped to distinct vectors based on their role in a triple resulting in logical property preserving embeddings.", "method": "The proposed approach is applied to TransE, TransR, and TransD to produce lppTransE, lppTransR, and lppTransD, respectively. The performance of these extensions is evaluated on link prediction and triple classification tasks.", "outcome": "According to experimental results, the proposed logical property preserving embeddings showed state-of-the-art performance on both link prediction and triple classification tasks.", "future_impact": "The results underscore the importance of preserving logical properties of relations while embedding knowledge graphs, and thus illuminates a pathway for developing improved embeddings in the future.", "venue": "NAACL", "year": 2016, "title": "A Translation-Based Knowledge Graph Embedding Preserving Logical Property of Relations"} +{"pid": "6153e0205244ab9dcb39a47b", "context": "Going from sentence-level to document-level relation extraction has posed challenges due to increasing text length and more complicated entity interactions. Current methods only implicitly model critical information sources such as relevant contexts and entity types, leading to the problems of ineffective supervision and uninterpretable model predictions.", "key_idea": "The authors propose an approach called Supervising and Augmenting Intermediate Steps (SAIS) that explicitly captures key information sources like relevant contexts and entity types by introducing intermediate supervision and augmentation during the relation extraction process.", "method": "SAIS is based on a variety of carefully designed tasks. The method is evaluated on three benchmarks for document-level relation extraction: DocRED, CDR, and GDA. They also assess model uncertainty to further enhance the performance with evidence-based data augmentation and ensemble inference.", "outcome": "SAIS results in state-of-the-art relation extraction performance on three benchmarks (DocRED, CDR, and GDA) and achieves a 5.04% relative increase in F1 score compared to the runner-up in evidence retrieval on DocRED.", "future_impact": "The method can potentially improve extraction of relationships in documents and has potential for increased interpretability due to the retrieval of supporting evidence, although the authors do not explicitly mention this in the abstract.", "venue": "NAACL", "year": 2022, "title": "SAIS: Supervising and Augmenting Intermediate Steps for Document-Level Relation Extraction"} +{"pid": "1d07d6d4-525d-42e7-ac97-b1359260d72b", "context": "There has been active development of corpora and annotations in the BioNLP community. As these resources accumulate, the issue of reusability arises.", "key_idea": "To improve the reusability of corpora and annotations, the authors propose PubAnnotation, a persistent and sharable repository where various corpora and annotations can be stored together in a stable and comparable way.", "method": "The authors present a prototype repository as a proof-of-concept.", "outcome": "PubAnnotation, a persistent and shareable repository where various corpora and annotations can be stably and comparably stored, is proposed and prototyped by the authors.", "future_impact": "The authors anticipate that PubAnnotation will improve the reusability of corpora and annotations in the BioNLP community.", "venue": "NAACL", "year": 2012, "title": "PubAnnotation - a persistent and sharable corpus and annotation repository"} +{"pid": "60af77bb9e795e6b8e55c891", "context": "The task of implicit discourse relation recognition (IDRR) involves identifying logical relations between two nearby sentences in a piece of text. Existing solutions fail to fully utilize the contextual information which is crucial in interpreting each individual sentence.", "key_idea": "The authors propose a novel graph-based Context Tracking Network (CT-Net) that recalibrates the discourse context for improved IDRR. They adopt a new structural approach, converting the discourse into a paragraph association graph (PAG), in which each sentence follows related context from the intricate discourse via various kinds of edges.", "method": "The CT-Net extracts contextual representation from the PAG using a uniquely designed cross-grained updating mechanism that can effectively combine both sentence-level and token-level contextual semantics. They compare the performance of this method with other models using the PDTB 2.0 dataset for experiments.", "outcome": "The experiments showed that the CT-Net performed better than models that just approximated the context.", "future_impact": "The introduction of a novel approach to model context within discourse relation recognition could help in achieving improved performance for related tasks, although the specific potential improvements are not identified in the abstract.", "venue": "NAACL", "year": 2021, "title": "Context Tracking Network: Graph-based Context Modeling for Implicit Discourse Relation Recognition"} +{"pid": "88d7dd28-58f3-401a-a48a-acbbf0073dc9", "context": "Integrating neural language models in translation systems is crucial for real-world applications, but scaling these models is a challenging task.", "key_idea": "The authors investigate and evaluate the impact of new and existing scaling techniques on the quality of end-to-end Machine Translation (MT) and explore the tradeoffs between neural models and back-off ngram models.", "method": "The paper involves the evaluation of scaling techniques on MT quality. It also investigates the usage of noise contrastive estimation and diagonal contexts for further speed improvements.", "outcome": "The results show that neural models are strong candidates for natural language applications in memory-constrained environments but still lag behind traditional models in raw translation quality.", "future_impact": "The study concludes with recommendations for building a scalable neural language model for MT, which could inform future research and developments in the field.", "venue": "NAACL", "year": 2015, "title": "Pragmatic Neural Language Modelling in Machine Translation"} +{"pid": "f49c99ad-2f12-44a3-af18-250176c276c9", "context": "The current efforts in automated story telling and poetry generation involve modeling various human abilities.", "key_idea": "The author makes an analogy between bird-flight and jet planes, exploring how computational models used in automated storytelling and poetry generation relate to human linguistic creativity.", "method": "The method involves reviewing a number of efforts in automated storytelling and poetry generation, identifying the human abilities being modeled in each case, and drawing comparisons with the case for flight.", "outcome": "The paper identified what the similarities are between the case for linguistic creativity and the case for flight.", "future_impact": "The comparison and exploration might provide insights into artificial linguistic creativity if the analogy proves to be valid.", "venue": "NAACL", "year": 2010, "title": "Engineering Linguistic Creativity: Bird Flight and Jet Planes"} +{"pid": "2a61c2b1-f102-4920-917b-dce57e32f566", "context": "Existing word embedding algorithms are not typically tailored to task-specific bilingual settings.", "key_idea": "A simple wrapper method is introduced which adapts off-the-shelf word embedding algorithms to learn task-specific bilingual word embeddings by using a small dictionary of task-specific word equivalence classes.", "method": "The prevalent method involves creating mixed context-target pairs using the task-specific equivalence classes to train existing embedding models, which was then evaluated on unsupervised cross-language part-of-speech (POS) tagging and semi-supervised cross-language super sense (SuS) tagging tasks.", "outcome": "The proposed method outperforms off-the-shelf bilingual embeddings in both the unsupervised cross-language part-of-speech (POS) tagging task and the semi-supervised cross-language super sense (SuS) tagging task.", "future_impact": "The proposed method can be easily adapted to specific tasks, offering significant utility without the need for parallel data and allowing flexibility in the choice of embedding algorithm.", "venue": "NAACL", "year": 2015, "title": "Simple task-specific bilingual word embeddings"} +{"pid": "fbe113d4-456d-4797-8cc7-88ab04b386e2", "context": "Topical poetry segmentation is an area that needs basic research.", "key_idea": "The authors conduct a pilot study to test initial assumptions and methodologies on topical poetry segmentation, focusing specifically on the poem Kubla Khan.", "method": "Nine segmentations of the poem titled Kubla Khan were collected and analyzed to produce inter-coder agreement.", "outcome": "The study found low but comparable inter-coder agreement BSe in the segmentations.", "future_impact": "The discussions and analyses of these codings can focus on improving agreement and offer initial results on the nature of topics in this poem, potentially influencing further research on topical poetry segmentation.", "venue": "NAACL", "year": 2013, "title": "An initial study of topical poetry segmentation"} +{"pid": "f44d6c1e-7ba7-4819-a6d2-2599dfe7f170", "context": "Functional meta-data about authors and important information regarding study parameters such as longitudinal variables and medical intervention methods are currently manually extracted from clinical research articles. This is often cumbersome and time-consuming.", "key_idea": "The authors propose using conditional random fields (CRFs) to automatically extract this metadata from the title page and body text of clinical research articles.", "method": "The paper explores the usage of CRFs for extraction of formulaic metadata from the title page and free text fields from the body text of the articles, but does not provide specifics of the dataset or experimental setup.", "outcome": "The preliminary results show an acceptable level of performance for formulaic metadata extraction and high precision for those found in the free text.", "future_impact": "Automated extraction of such information can benefit readers by allowing deep semantic search of articles and aid policy makers and sociologists in tracking macro level trends in research.", "venue": "NAACL", "year": 2010, "title": "Extracting Formulaic and Free Text Clinical Research Articles Metadata using Conditional Random Fields"} +{"pid": "3eb61c9d-f581-42bc-a052-c31fff96537c", "context": "Facial expressions are a key nonverbal communication cue and they can provide useful feedback in conversations, including in human-robot interaction. However, interpreting these expressions in the context of success or failure feedback has been a challenge, especially in comparison to human recognition performance.", "key_idea": "The key idea of this paper is evaluating three standard pattern recognition techniques -- active appearance models, gabor energy filters, and raw images -- for facial feedback interpretation in terms of valence (success and failure), and comparing these results with human performance.", "method": "The authors conducted an evaluation using a database that contains videos of people interacting with a robot by teaching it the names of several objects. The robot's task was to identify the objects correctly and the researchers assessed the subjects' spontaneous facial expressions in response to the robot's answers.", "outcome": "Their results showed that an automatic classification of facial expressions in terms of valence using standard pattern recognition techniques is possible, with an accuracy that is comparable to the average human classification rate. However, there was a high variance between different subjects, similar to human performance.", "future_impact": "The results point to the potential of using standard pattern recognition techniques in interpreting facial feedback, which could be applied in enhancing human-robot interactions or other applications that involve automatic recognition and response based on facial cues.", "venue": "CVPR", "year": 2010, "title": "Facial expressions as feedback cue in human-robot interaction\u2014a comparison between human and automatic recognition performances"} +{"pid": "701aaedd-0d1f-46d3-b06f-bf3d8bf8914c", "context": "Background detection is a common problem in computer vision and has traditionally used models that simply record each pixel's historical background values.", "key_idea": "This paper presents a new background detection algorithm that not only records each pixel's historical background values, but also estimates the efficacy of these values based on occurrence statistics, allowing for selective adaptation to background changes with different timescales and minimizing the generation of ghosts.", "method": "The authors carried out evaluations of their proposed algorithm using the ChangeDetection.net database.", "outcome": "The results show that the proposed algorithm outperforms most earlier state-of-the-art algorithms not only in terms of accuracy, but also in terms of processing speed.", "future_impact": "The algorithm, with its ability to quickly adapt to different timescales and minimize ghosts, could shape future approaches to pixel-level adapting background detection.", "venue": "CVPR", "year": 2014, "title": "A Fast Self-Tuning Background Subtraction Algorithm"} +{"pid": "84a56aa0-d7a1-4c15-b470-abb5a1e65bfb", "context": "Detecting and tracking multiple moving targets from airborne videos and audio data are crucial, but require complex architecture and computations often beyond the capabilities of a single machine.", "key_idea": "This paper introduces a Cloud-based Audio-Video (CAV) fusion system. This system innovatively combines a user-based voice-to-text color feature descriptor track matching with automated hue feature extraction from image pixels.", "method": "The authors developed a method that employs Cloud computing to allow for real-time performance. The method involves testing with the tracking of multiple moving targets using airborne videos and assessing frame rate, detection, tracking, and classification performance under realistic conditions.", "outcome": "The results demonstrate that the CAV approach provides improved frame rate, enhanced detection, real-time tracking, and classification performance under realistic conditions, compared to a single machine workflow.", "future_impact": "The CAV system is general-purpose, potentially allowing for the detection and tracking of different valuable targets' movement and suspicious behavior recognition through multi-intelligence data fusion.", "venue": "CVPR", "year": 2015, "title": "A cloud infrastructure for target detection and tracking using audio and video fusion"} +{"pid": "95fcc6f7-80f6-42bc-b950-4eb6aa06af17", "context": "Prior work has not been able to estimate a human's anthropometric measurements and pose from a single image simultaneously.", "key_idea": "The novel approach of the study is the four-step technique that combines feature selection, statistical human body information, joint limit constraints, and optimization to estimate anthropometry and pose from an image.", "method": "The authors select image points as projections of selected landmarks and generate plausible length estimates. They then generate plausible poses using geometric methods and perform optimization under certain constraints. They apply this method to both synthetic and real input data.", "outcome": "The results showcase the accuracy, advantages, and limitations of the novel method in estimating anthropometry and pose from a single image.", "future_impact": "The study opens up the ability to estimate both anthropometry and pose from a single image, which may have broad applications in fields reliant on image-based human body analysis.", "venue": "CVPR", "year": 2000, "title": "Estimating anthropometry and pose from a single image"} +{"pid": "7df348eb-ec7d-4d5f-beb9-432a6a4b9e1f", "context": "Vision and sound are basic sources through which humans understand the world and there is no existing method to generate sound given visual input.", "key_idea": "This work introduces the task of generating sound based on visual input. This could be used for creating sounds in virtual reality or providing additional accessibility for people with visual impairments.", "method": "The authors apply learning-based methods to generate raw waveform samples given input video frames and evaluate on a dataset of videos containing a variety of sounds.", "outcome": "The experiments demonstrate that the generated sounds are fairly realistic and have good temporal synchronization with the visual inputs.", "future_impact": "This research could lead to applications in virtual reality where sound is automatically generated for virtual scenes. Also, it could be used to provide additional accessibility to images or videos for people with visual impairments.", "venue": "CVPR", "year": 2018, "title": "Visual to Sound: Generating Natural Sound for Videos in the Wild"} +{"pid": "5eb67f3591e011ab59e82a98", "context": "This paper reviews a second challenge on spectral reconstruction from RGB images which involves the recovery of whole-scene hyperspectral (HS) information from a 3-channel RGB image. There has been a previous challenge in this area.", "key_idea": "The challenge provides two tracks - a 'Clean' track, which involves estimation of hyperspectral images from noise-free RGBs, and a 'Real World' track, which involves recovery of hyperspectral images from noisy JPEG-compressed RGB images.", "method": "The challenge uses a natural hyperspectral image dataset containing a total of 510 HS images. The Clean and Real World tracks had 103 and 78 registered participants respectively, with 14 teams competing in the final testing phase.", "outcome": "The paper does not detail specific outcomes, only mentioning that the challenge scores of proposed methods were evaluated and a state-of-the-art gauge in spectral reconstruction from an RGB image was established.", "future_impact": "The creation of new benchmarks for spectral reconstruction from RGB images, and insights from the challenge can lead to improved methods in this field.", "venue": "CVPR", "year": 2020, "title": "NTIRE 2020 Challenge on Spectral Reconstruction from an RGB Image"} +{"pid": "603e1e58-6414-4d1c-8690-dae7a6c814ca", "context": "Matching and registration of shapes is a crucial area within Computer Vision, Pattern Recognition, and Medical Image Analysis.", "key_idea": "The paper presents a shape representation framework that is based on Gaussian curvature and Markov random fields (MRFs) for shape matching.", "method": "The method projects a surface mesh model in R3 into a two-dimensional space, modeling it as an extended boundary closed Markov random field. Correspondence between two surface meshes is then established by performing probabilistic inference on the MRF via Gibbs sampling.", "outcome": "The technique successfully combines geometric, topological, and probabilistic information, producing an effective representation for shapes in three dimensional space.", "future_impact": "The representation is generalizable to higher dimensional spaces and can be utilized for shape matching, registration, and statistical shape analysis in future studies.", "venue": "CVPR", "year": 2007, "title": "An MRF and Gaussian Curvature Based Shape Representation for Shape Matching"} +{"pid": "5eccb534e06a4c1b26a83840", "context": "Large detection datasets often have a 'long tail' of lowshot classes with very few bounding box annotations. There is a need for more effective detection for lowshot classes, ideally that can leverage weakly labelled web-scale datasets that only have image-level labels.", "key_idea": "The authors propose a modification to the FRCNN model that can automatically infer label assignment for object proposals from weakly labelled images during training. This label assignment is posed as a Linear Program with constraints on the number and overlap of object instances in an image.", "method": "The authors compare their proposed model to training with just a few annotated examples, and examine the effect of the amount of weakly labelled and fully labelled data needed to train the detection model. They test their model on the Low Visibility (LVIS) dataset and different lowshot variants of the COCO dataset.", "outcome": "Their framework significantly improves performance when augmenting with weakly labelled examples. It achieves a 3.5 gain in average precision (AP) on the LVIS dataset, outperforming self-supervised baselines like omni-supervision for lowshot classes.", "future_impact": "This work introduces a detection framework for lowshot classes using weakly labelled web-scale data, leading to potential improvements in object detection, especially for lowshot classes with limited bounding box annotated images.", "venue": "CVPR", "year": 2020, "title": "DLWL: Improving Detection for Lowshot Classes With Weakly Labelled Data"} +{"pid": "622abdd05aee126c0f56ba22", "context": "Temporal Action Localization (TAL) has progressed significantly under the supervised learning paradigm, but current methods function under the closed set assumption, which is inadequate for dealing with unknown actions in open-world scenarios.", "key_idea": "The paper introduces the Open Set TAL (OSTAL) problem and proposes a general framework, OpenTAL, based on Evidential Deep Learning (EDL) which can manage uncertainty-aware action classification, actionness prediction, and temporal location regression.", "method": "The framework, OpenTAL, is developed using an importance-balanced EDL method for learning classification uncertainty and it also incorporates positive-unlabeled learning for actionness. The classification uncertainty is further calibrated with guidance from the temporal localization quality.", "outcome": "Experiments on THUMOS14 and ActivityNet1.3 benchmarks demonstrate the effectiveness of the OpenTAL method.", "future_impact": "OpenTAL is designed to be a general framework that can equip existing TAL models for tackling open set scenarios, potentially broadening their applicability and utility.", "venue": "CVPR", "year": 2022, "title": "OpenTAL: Towards Open Set Temporal Action Localization"} +{"pid": "43f0ce24-28c9-4e48-9c9d-5d2a1428fe53", "context": "Calibrating a small field of view central perspective non-frontal camera is a difficult task. Existing techniques suffer from over-parameterization issues, needing heuristic initialisation of a few calibration parameters, and have limitations in handling arbitrary lens-sensor tilt. Approaches based on rotation parameterization of sensor non-frontalness/tilt suffer from these drawbacks and approaches based on decentering distortion modeling can only handle small tilts and cannot explicitly estimate the sensor tilt.", "key_idea": "The paper proposes a new rotation based non-frontal camera calibration method that focuses on improving the limits of existing techniques. It includes a generalized pupil-centric imaging model for arbitrary axis lens-sensor tilt and an analytical solution that estimates calibration parameters, including sensor rotation angles and pupil-centric parameters.", "method": "The authors present an analytical technique and computational calibration parameter estimates. They initialize non-linear calibration optimization for a set of synthetic and real data captured from a non-frontal camera.", "outcome": "The proposed method shows reduced pixel re-projection and undistortion errors compared to state of the art techniques in rotation and decentering based approaches to non-frontal camera calibration.", "future_impact": "The successful implementation of this method has the potential to improve non-frontal camera calibration processes significantly, reducing errors and handling arbitrary lens-sensor tilt.", "venue": "CVPR", "year": 2014, "title": "Generalized Pupil-centric Imaging and Analytical Calibration for a Non-frontal Camera"} +{"pid": "68597f59-6dfa-4401-b737-971e31e5ca76", "context": "Browsing activities is important for various applications and consumes a large amount of viewing time for longer videos, making summarization of these videos crucial.", "key_idea": "The authors propose a method for the selection of salient activities in longer videos based on the motion of feature points, where the saliency of a frame is dependent on the total motion and specified time for summarization, and the motion information is modeled as a Gaussian mixture model (GMM).", "method": "The proposed method utilizes GMM to model the motion information in a video in order to estimate key motion frames. The salient frames are then detected based on the strength of the motion and user-specified time, contributing to the summarization whilst keeping the chronology of activities.", "outcome": "Applications of the proposed method in the summarization of surveillance videos, movies, TV serials etc. are demonstrated and yield comparable results with the stroboscopic approach and maintaining an average retention ratio of 95%.", "future_impact": "The proposed method could have a wide impact on applications that require browsing activities in long-duration videos, from surveillance use-cases to entertainment like TV serials and movies.", "venue": "CVPR", "year": 2013, "title": "Time driven video summarization using GMM"} +{"pid": "62c2a8cd5aee126c0fcebdbe", "context": "Existing visual grounding (VG) methods that align the correct regions of an image with a natural language query about the image are constrained by a single-stage process, with performance depending on the density and quality of candidate regions. They also lack the ability to optimize located regions continuously.", "key_idea": "The paper proposes a new approach to the VG challenge that models the process as a progressively optimized visual semantic alignment operation. They introduce a new model, Multi-modal Dynamic Graph Transformer (M-DGT), that builds upon a dynamic graph structure with regions as nodes and their semantic relations as edges.", "method": "The M-DGT makes sustainable adjustments to the nodes and edges of the graph based on multi-modal information and the graph feature. This approach efficiently shrinks the graph to approach ground truth regions. The model's performance was tested on the Flickr30k Entities and RefCOCO datasets.", "outcome": "Experiments show that M-DGT achieves superior results over existing state-of-the-art methods using an average of 48 boxes as initialization, it outperforms existing methods in terms of both accuracy and Intersect over Union (IOU) scores. Notably, incorporating M-DGT to optimize the predicted regions of existing methods improved their performance even further.", "future_impact": "The model may pave the way for future advancements in visual grounding and related areas of computer vision and natural language processing, and impact the performance of existing models when used for optimization.", "venue": "CVPR", "year": 2022, "title": "Multi-Modal Dynamic Graph Transformer for Visual Grounding"} +{"pid": "5dcd263a3a55ac58039516c5", "context": "Contrastive learning for unsupervised visual representation learning has traditionally relied on static dictionaries, which may not facilitate efficient learning due to their static nature.", "key_idea": "The paper introduces Momentum Contrast (MoCo) for unsupervised visual representation learning that uses a dynamic dictionary with a queue and a moving-averaged encoder to enable efficient contrastive unsupervised learning.", "method": "The authors tested MoCo under the common linear protocol on ImageNet classification, and on several downstream detection/segmentation tasks on the PASCAL VOC, COCO, and other datasets.", "outcome": "MoCo provided competitive results on ImageNet classification and outperformed supervised pre-training in 7 detection/segmentation tasks on various datasets, sometimes by large margins.", "future_impact": "These findings suggest that the gap between unsupervised and supervised representation learning has been largely closed in many visual tasks, paving the way for more effective and efficient unsupervised learning in future.", "venue": "CVPR", "year": 2020, "title": "Momentum Contrast for Unsupervised Visual Representation Learning"} +{"pid": "933bcc37-2595-47b8-9b78-49e300de8355", "context": "The digitization of printed documents is commonplace due to technological advancements, but it comes with an increased risk of fraudulent and criminal activities due to the ease of manipulation of these digital scans.", "key_idea": "The authors propose a source identification method for scanned documents that utilizes local tetra patterns to capture unique, device-specific signatures from images of the printed documents.", "method": "The proposed method uses all characters to train a single classifier, reducing the amount of training data required. The method is tested for font size independence on an existing scanner dataset and for font shape independence on a smartphone dataset of comparable size.", "outcome": "The authors indicated the proposed method depicts font size independence when tested on an existing scanner dataset and a novel step towards font shape independence when tested on a smartphone dataset.", "future_impact": "The suggested method can provide critical information for investigation agencies and help in building secure communication systems, helping to combat fraudulent and criminal activities.", "venue": "CVPR", "year": 2017, "title": "Source Classification Using Document Images from Smartphones and Flatbed Scanners"} +{"pid": "7cd410ea-1d89-44e2-9e46-ee4055c9d0b8", "context": "The currently available optical flow datasets are restricted in size and not diverse enough, this limits the testing of performance of methods on non-photorealistic images.", "key_idea": "The team presents the Creative Flow+ Dataset, a diverse multi-style artistic video dataset which is richly labeled with various parameters such as per-pixel optical flow, occlusions, correspondences, segmentation labels, normals, and depth, aimed at testing optical flow methods on non-photorealistic images.", "method": "The dataset includes 3000 animated sequences rendered using styles randomly selected from 40 textured line styles and 38 shading styles, spanning the range between flat cartoon fill and wildly sketchy shading.", "outcome": "The Creative Flow+ dataset poses a new challenge in generalizing real-world Computer Vision techniques to stylized content. Learning-based optical flow methods have been shown to fail to generalize to this data and struggle to compete with classical approaches.", "future_impact": "This dataset can stimulate new research in the area of learning-based optical flow methods and their application to stylized data. Additionally, they release the complete dataset creation pipeline, allowing the community to generate and stylize their own data on demand.", "venue": "CVPR", "year": 2019, "title": "Creative Flow+ Dataset"} +{"pid": "deecb079-ed80-48c3-82ae-dff4de8bbff9", "context": "Mobile food recognition systems require sending images to a server for recognition, which can result in delays and potential privacy concerns.", "key_idea": "The authors propose a real-time mobile food recognition system, conducted on a smartphone, that estimates both the calorie and nutritious content of the food, as well as records the user's eating habits. This is used to recognize food items within user-indicated bounding boxes.", "method": "For food recognition, the system segments each food item region by using GrabCut, then extracts a color histogram and SURF-based bag-of-features, and classifies it into one of fifty food categories using linear SVM and fast 2 kernel. Also, a system was implemented as an Android smartphone application leveraging multiple CPU cores for real-time recognition.", "outcome": "The proposed system achieved an 81.55% classification rate for the top five category candidates when the ground-truth bounding boxes were present. User study indicated positive evaluation over food recording systems without object recognition.", "future_impact": "This real-time mobile food recognition system has the potential to further enhance nutritional tracking by combining easy to use interface with efficient food recognition mechanisms and calorie estimation.", "venue": "CVPR", "year": 2013, "title": "Real-Time Mobile Food Recognition System"} +{"pid": "a8bc664c-6302-4d24-87d2-c8b4e9da04a9", "context": "Computer vision algorithms often make mistakes in human-centric applications, and some errors are more frustrating for users than others. An annoyance or cost matrix that quantifies the frustration of each type of mistake is not readily accessible.", "key_idea": "The authors propose an approach to predict the annoyance created by unseen mistakes by learning from example mistakes and their relative annoyance. This approach involves the use of attribute-based representations to transfer the knowledge of annoyance.", "method": "The authors tested this approach on faces and scenes, comparing this method with different baselines and applying it to a real-world image retrieval application.", "outcome": "The experimental results confirmed that the proposed approach can predict annoyance more accurately than baselines and results in less annoying mistakes in a real-world image retrieval application.", "future_impact": "This new approach can aid in designing computer vision algorithms that minimize user annoyance, particularly handy for a wide spectrum of human-centric applications where annoyance is closely linked to human perception.", "venue": "CVPR", "year": 2014, "title": "Predicting User Annoyance Using Visual Attributes"} +{"pid": "ad7da8c8-42e7-4c93-bb39-e80a06e97e75", "context": "In visual surveillance, event rules are often manually defined for trajectory analysis, which involves analysing patterns of moving objects. This manual rule definition poses challenges in terms of scalability, efficiency and human effort.", "key_idea": "The paper proposes a generic rule induction framework based on trajectory series analysis to automate the learning of event rules. It involves mapping trajectories into primitive events and using a minimum description length principle based grammar induction algorithm to infer meaningful rules.", "method": "The authors carry out experiments in a traffic crossroad scenario to validate the effectiveness of the proposed method.", "outcome": "Experimental results show that the grammar rules obtained by the proposed algorithm are consistent with the actual traffic events in the crossroad. Additionally, the algorithm is capable of correctly learning the traffic lights rule in the crossroad by eliminating irrelevant trajectories.", "future_impact": "This study introduces an automatic event rule learning method that can reduce human effort, improve scalability and efficiency in visual surveillance systems that are reliant on trajectory analysis.", "venue": "CVPR", "year": 2007, "title": "Trajectory Series Analysis based Event Rule Induction for Visual Surveillance"} +{"pid": "47170e9d-e1cf-4182-aba6-f2e2d41e9a35", "context": "Handheld projectors are a new type of display modality that provide a larger display area than handheld devices with fixed screens, but they are not tied to a physical screen or a fixed projection area.", "key_idea": "The authors review a prototype handheld projector and introduce interaction using a cursor that can be tracked across the projection, aiming to support web-browsing applications and explore potential novel applications not available with a physical screen or a fixed projection.", "method": "The authors present their prototype handheld projector and discuss its interaction design. They also show examples of potential applications, illustrating both conventional uses such as web browsing, and novel uses that interact with the physical world.", "outcome": "The authors successfully demonstrate that their handheld projector can be used for typical applications, like web-browsing, and potentially novel applications that provide a more natural interaction with the physical world.", "future_impact": "The introduction of handheld projector and its interaction method may inspire new types of applications that aren't possible with regular screens or fixed projections, creating more natural ways to interact with the digital world.", "venue": "CVPR", "year": 2005, "title": "Handheld Projectors for Mixing Physical and Digital Textures"} +{"pid": "5de4e0bf3a55ac2224ba54a3", "context": "Existing domain adaptation methods aim at learning features that can be generalized among domains, often requiring an update to the source classifier to adapt to the target domain and struggling to properly balance performance between source and target domains.", "key_idea": "Instead of training a classifier to adapt to the target domain, the study proposes using a separable component called a data calibrator to help the fixed source classifier recover discrimination power in the target domain, whilst preserving the source domain's performance.", "method": "The proposed method is evaluated in situations where the difference between two domains is small, and in cases where it is larger, it leverages synthetic images generated by GANs.", "outcome": "When domain differences are small, the source classifier's representation performs well in the target domain, outperforming GAN-based methods in digits. For larger domain differences, the proposed method boosts performance and hits state-of-the-art performance in digits datasets and driving scene semantic segmentation.", "future_impact": "Empirical results from this study reveal that certain intriguing hints, which can be mitigated by adversarial attack to domain discriminators, are sources of performance degradation under domain shift, and possibly providing a new research direction to alleviate this impact.", "venue": "CVPR", "year": 2020, "title": "Light-weight Calibrator: a Separable Component for Unsupervised Domain Adaptation"} +{"pid": "fdeff5be-b9bf-440e-805f-6fab335ee173", "context": "Detecting and tracking multiple people in spaces such as stores is a necessary component of effective surveillance in shopping environments. However, identifying and grouping individuals who are part of a shopping group and monitoring specific activities, such as transactions at a checkout, within these environments remain challenging.", "key_idea": "The authors present a monocular real-time computer vision system capable of identifying shopping groups by accurately detecting and tracking multiple individuals as they wait in line at a checkout or service counter area.", "method": "The proposed system segments each frame into foreground regions containing multiple individuals, leveraging temporal segmentation of foreground and motion cues to further segment these regions into individual persons. An appearance model based on color and edge density is used to track each person's movements, then analysing interbody distances to group people into shopping groups. The system also monitors the activities of the cashier to determine when shopping transactions start and end.", "outcome": "The system successfully demonstrated real-time tracking and group detection, illustrating its robust performance.", "future_impact": "The proposed system holds potential for delivering improved surveillance in shopping environments, thereby leading to enhanced monitoring and management of such retail spaces.", "venue": "CVPR", "year": 2001, "title": "Detection and tracking of shopping groups in stores"} +{"pid": "61a596595244ab9dcbdfd667", "context": "Domain generalization algorithms are designed to use data from multiple domains to learn models that generalize well to unseen domains. Current evaluation methods cannot effectively expose the impact of various factors contributing to the lackluster performance of existing algorithms.", "key_idea": "The authors propose an evaluation framework that decomposes the error into components capturing distinct generalization aspects. This work specifically focuses on assessing domain-invariant representation learning and identifying potential failure modes.", "method": "The proposed evaluation framework is used to assess the performance of different domain generalization algorithms on Colored MNIST and Camelyon-17 datasets.", "outcome": "The authors found that the most significant contributor to generalization error could change depending on method, dataset, regularization strength, and training length. On Colored MNIST, domain generalization algorithms failed because they reached domain-invariance only on the training domains, while on Camelyon-17, domain-invariance degraded the quality of representations on unseen domains.", "future_impact": "This study hints towards shifting the focus to tuning the classifier on top of a rich representation as a potentially promising direction in domain generalization.", "venue": "CVPR", "year": 2022, "title": "Failure Modes of Domain Generalization Algorithms."} +{"pid": "467b5a75-72fd-4436-94e1-d87e43131b5b", "context": "The performance of local appearance-based face recognition schemes can be affected by various factors, but the effects of feature selection and feature normalization have not been assessed thoroughly in past research.", "key_idea": "The paper investigates how feature selection and feature normalization are affecting the performance of a local appearance-based face recognition schema which uses local features extracted using block-based discrete cosine transform.", "method": "Three different feature sets are extracted from local vectors and then normalized in two ways: making them unit norm and dividing each coefficient by its standard deviation learned from the training set. The processed test face images are classified using four different distance measures: L1 norm, L2 norm, cosine angle, and covariance between feature vectors, on the AR and CMU PIE face databases.", "outcome": "The experiments demonstrated the importance of using appropriate feature sets and properly normalizing the feature vector in order to enhance the performance of appearance-based face recognition schemes.", "future_impact": "The findings could guide the development of more effective face recognition schemes that take into account the critical factors of feature selection and normalization.", "venue": "CVPR", "year": 2006, "title": "Analysis of Local Appearance-Based Face Recognition: Effects of Feature Selection and Feature Normalization"} +{"pid": "e5049522-e7af-478d-81ae-5062e585c014", "context": "The Naive Bayes Nearest Neighbor (NBNN) is an image classification algorithm, but its accuracy and scalability to large numbers of object classes can be improved.", "key_idea": "The authors propose an improvement to the NBNN algorithm, termed Local Naive Bayes Nearest Neighbor (Local NBNN), which maintains all of the reference data in one search structure and bases posterior probability estimates on the local neighborhood of a descriptor, thereby ignoring distant classes.", "method": "The authors tested their algorithm, comparing it to the original NBNN as well as other spatial pyramid methods, using datasets like Caltech 256.", "outcome": "The Local NBNN provided a 100 times speed-up over the original NBNN on the Caltech 256 dataset and outperformed all previous NBNN-based methods and the original spatial pyramid model. However, it did not outperform state-of-the-art spatial pyramid methods that use local soft assignment and max-pooling.", "future_impact": "While not explicitly mentioned in the abstract, an implied future impact could be revisions to the Local NBNN algorithm to potentially outperform state-of-the-art spatial pyramid methods using local soft assignment and max-pooling.", "venue": "CVPR", "year": 2012, "title": "Local Naive Bayes Nearest Neighbor for image classification"} +{"pid": "da59d5da-86f8-4774-a8e3-4a7f1d25a942", "context": "Existing video co-segmentation methods struggle to separate complex foregrounds and backgrounds using low-level feature representations.", "key_idea": "The authors introduce a category-independent object proposal based video co-segmentation method called ObMiC, which can extract multiple foreground objects in a video set, overcoming the limitations of separating foregrounds and backgrounds using low-level representations.", "method": "The authors formulate object-based co-segmentation as a co-selection graph, favoring regions with foreground-like characteristics, while accounting for intra-video and inter-video foreground coherence. To handle multiple foreground objects, they propose a multi-state selection graph model (MSG) that optimizes the segmentations of different objects jointly.", "outcome": "Experiments demonstrate that the object-based multiple foreground video co-segmentation method (ObMiC) compares favorably with related techniques on both single and multiple foreground cases.", "future_impact": "The proposed MSG model could potentially be applied to standard graph models to turn them into a multi-state selection solution optimizable by existing energy minimization techniques.", "venue": "CVPR", "year": 2014, "title": "Object-Based Multiple Foreground Video Co-segmentation"} +{"pid": "3e7c83e5-9b50-4a26-82e7-3fd3a8a3df92", "context": "Existing models for image segmentation struggle with fitting diverse region shapes and are sensitive to initial shape and position.", "key_idea": "The authors propose fully 3D active surface models for image segmentation that are capable of fitting a diverse range of region shapes and have low sensitivity to initial shape and position. The models are designed with self-inflation/deflation forces that permit the active surface to travel long distances without any external forces.", "method": "The proposed models are tested using synthetic and real images, taking into account accuracy, robustness to image noise and imperfect image data, and the ability to converge to the correct boundary even with incorrect initial estimates.", "outcome": "The proposed models demonstrated accurate segmentation in tests with synthetic and real images, with results indicating they are robust to noise and imperfect data and can converge to the correct boundary, even when the initial estimate is off.", "future_impact": "The authors anticipate that the proposed models, with their robustness and versatile fitting capabilities, can significantly improve image segmentation tasks with a wide range of region shapes.", "venue": "CVPR", "year": 1997, "title": "Fully 3D active surface models with self-inflation and self-deflation forces"} +{"pid": "5e982cc591e0119e8a952424", "context": "Current action recognition techniques have achieved great success on public benchmarks, but their performance in real-world applications requiring fine-grained action understanding, such as sport analysis, is unsatisfactory.", "key_idea": "The authors propose the FineGym dataset developed from gymnastic videos to take action recognition to a new level. Unlike existing datasets, FineGym provides temporal annotations at both action and sub-action levels with a three-level semantic hierarchy.", "method": "The authors use the FineGym dataset to systematically investigate representative methods of action recognition.", "outcome": "Investigation on the FineGym dataset results in a number of interesting findings, though specific details are not provided in the abstract.", "future_impact": "The authors hope that the FineGym dataset could advance research towards action understanding.", "venue": "CVPR", "year": 2020, "title": "Finegym: A Hierarchical Video Dataset For Fine-Grained Action Understanding"} +{"pid": "c8bff909-f08d-47d9-b45c-5eb73a1239b4", "context": "The problem of semantic segmentation based on RGB-D data, focusing on cluttered indoor scenes containing many visual categories and instances, is a challenging task. The effectiveness of automatic descriptors obtained using deep learning on this problem is still in question.", "key_idea": "The authors propose a method combining a generalization of parametric max flow figure-ground proposal methodology using intensity and depth information, new region description methods with second-order pooling, and a search-based structured prediction inference and learning process that can resolve conflicts in overlapping spatial partitions.", "method": "The proposed method was extensively evaluated on the challenging NYU Depth Dataset V2, extended for the RMRC 2013 and RMRC 2014 Indoor Segmentation Challenges, and a scene classification benchmark. Different descriptors, both pre-designed and automatically obtained using deep learning, were used in the experiments.", "outcome": "The proposed model ranks first in the RMRC 2014 Indoor Segmentation Challenges. With second-order and deep learning features combined, there can be over a 15% relative accuracy improvement. The method further improves the state of the art by 24% on a scene classification benchmark.", "future_impact": "The proposed parametric figureground intensity and depth-constrained proposal process and the principled search-based structured prediction inference and learning process have the potential for advancing semantic segmentation based on RGB-D data.", "venue": "CVPR", "year": 2015, "title": "Second-order constrained parametric proposals and sequential search-based structured prediction for semantic segmentation in RGB-D images"} +{"pid": "5d835fd63a55ac583ecde695", "context": "Image style transfer is an underdetermined problem with large number of potential solutions, and most current methods cannot produce diverse outputs. Recently, diversity losses were introduced to train feed-forward networks that can produce diverse outputs, but these still face issues.", "key_idea": "The authors propose a new method for diversified style transfer by incorporating the whitening and coloring transforms (WCT) with a novel deep feature perturbation (DFP) process. The DFP uses an orthogonal random noise matrix to perturb the deep image features while keeping the original style information intact.", "method": "The authors perform experiments to test the proposed method and its ability to increase diversity in style transfer outcomes. The method is also evaluated in user studies, where users assess the results obtained through diversified approaches based on the new method.", "outcome": "The results of the experiments indicate that the new method enhances the diversity while maintaining the quality of stylization. User studies also reveal that users find results obtained using the new method more satisfying.", "future_impact": "The authors anticipate that this learning-free method, which can be easily integrated into existing WCT-based methods, will make these current methods more capable of generating diverse results.", "venue": "CVPR", "year": 2020, "title": "Diversified Arbitrary Style Transfer via Deep Feature Perturbation"} +{"pid": "7afa267d-ea9a-4673-913b-b4eb8d4d9d54", "context": "Computerized sketch-face recognition is a crucial tool for law enforcement and has been intensively studied in recent years. The problem is that the existing CUFS face-sketch dataset is not realistic as the sketches are drawn while looking at the photograph to be matched later, leading to nearly perfect results and obscuring the true challenges of the task.", "key_idea": "This study poses doubts about the validity of the popular CUFS face-sketch dataset and introduces an off-the-shelf face recognition system with simple shape and edge features for photo-sketch and sketch-photo matching without requiring training data.", "method": "The authors compare the performance of their system to existing state-of-the-art approaches examining how different features (such as the hair region) contribute to recognition rates.", "outcome": "The authors find that their off-the-shelf system outperforms more sophisticated methods, even without using training data, achieving a recognition rate of 85.22% using only the hair region.", "future_impact": "The authors suggest that the existing dataset in this field is not appropriate for research and needs to be replaced with a more realistic one to further advance sketch-face recognition.", "venue": "CVPR", "year": 2012, "title": "Data insufficiency in sketch versus photo face recognition"} +{"pid": "2a07300b-6bf3-4e66-8a50-d9502977bbe7", "context": "Synthesizing complex images from text descriptions using existing models often results in poor quality synthesized images.", "key_idea": "The authors propose the Object-driven Attentive Generative Adversarial Networks (Obj-GANs), which allow attention-driven, multi-stage refinement for synthesizing complex images from text descriptions.", "method": "The authors implement the Obj-GAN with a novel object-driven attentive generative network and an object-wise discriminator based on the Fast R-CNN model. They then test its performance using the large-scale MS-COCO benchmark.", "outcome": "The proposed Obj-GAN significantly outperforms the previous state of the art in various metrics, increasing the inception score by 27% and decreasing the FID score by 11% on the MS-COCO benchmark.", "future_impact": "The insights provided through a comparison between the classic grid attention and the new object-driven attention could shed light on how future models may generate complex scenes with higher quality.", "venue": "CVPR", "year": 2019, "title": "Object-Driven Text-To-Image Synthesis via Adversarial Training"} +{"pid": "e8f08de6-6597-4fc4-a3c4-a641269d95ba", "context": "The problem of reconstructing an image from a bag of square, non-overlapping image patches, the jigsaw puzzle problem, is challenging even for humans and is known to be NP-complete. Prior approaches treated this problem as a constraint satisfaction problem.", "key_idea": "This paper proposes a departure from previous methods by developing a graphical model where each patch location is a node and each patch is a label at nodes in the graph. This model requires a pairwise compatibility term, which measures an affinity between two neighboring patches, and a local evidence term.", "method": "The paper discusses ways to obtain these terms for the jigsaw puzzle problem, evaluates several patch compatibility metrics, compares two forms of local evidence for the graphical model and estimates low resolution images from bags of patches.", "outcome": "It is demonstrated that the dissimilarity-based compatibility gives the best results. It is shown that sparse-and-accurate evidence, fixing as few as 4 \u2013 6 patches at their correct locations, is enough to reconstruct images consisting of over 400 patches, which is reportedly the largest puzzle solved in the literature to date. Additionally, coarse estimation of the low resolution image from a bag of patches is made possible.", "future_impact": "The work may pave the way for further innovations in image reconstruction by suggesting that a bag of image patches encodes some geometric information about the original image.", "venue": "CVPR", "year": 2010, "title": "A probabilistic image jigsaw puzzle solver"} +{"pid": "ce5bfbca-5ce4-43f7-bf80-cebc3eb9350b", "context": "Current methods for learning visual categories depend on a large amount of labeled data and can face severe difficulties when the number of labeled examples is limited.", "key_idea": "The paper proposes a method for learning representations from large amounts of unlabeled images which have associated captions with the aim to improve future image classification tasks.", "method": "The effectiveness of the proposed method was evaluated and compared to a fully-supervised baseline model, a model that does not utilize the captions and learns a visual representation by performing PCA on the unlabeled images alone, and a model that uses the output of word classifiers trained using captions and unlabeled data.", "outcome": "The experiments show that the proposed method significantly outperforms the fully-supervised baseline model, the model that ignores the captions, and the model that uses the output of word classifiers trained on captions and unlabeled data.", "future_impact": "The study opens up the opportunity to use other types of meta-data for learning visual representations besides captions.", "venue": "CVPR", "year": 2007, "title": "Learning Visual Representations using Images with Captions"} +{"pid": "2b93c6dd-4a08-42c0-b67d-7fa9bc58cecd", "context": "Correctly matching feature points in a pair of images is an important preprocessing step for many computer vision applications, which usually requires explicitly computing them.", "key_idea": "The authors propose an innovative method to estimate the number of correct feature matches and the overlapping region between images without explicitly computing them. This method leverages the spatial order of the features as projected onto the image's x-axis, transforming the match analysis into a permutation analysis between sequences.", "method": "The proposal is based on the use of the Kendall distance metric between permutations, along with assumptions on the correct and incorrect matches distribution. This method is validated through its application in otwo scenarios: a new halting condition for RANSAC-based epipolar geometry estimation methods, and discarding spatially unrelated image pairs in the Structure-from-Motion pipeline.", "outcome": "The experiments with both synthetic and real data demonstrate the proposed method's effectiveness, significantly reducing the running time of the image matching stage in the Structure-from-Motion pipeline by about 99% while preserving approximately 80% of the correctly matched feature points.", "future_impact": "The proposed method can be applied to other computer vision applications, impacting efficiency in these areas. Furthermore, it may fundamentally shift approaches toward feature matching, as it allows for the calculation of the likelihood that a given match is correct based on its estimated number and the rank of the features within the sequences.", "venue": "CVPR", "year": 2016, "title": "Using Spatial Order to Boost the Elimination of Incorrect Feature Matches"} +{"pid": "639e555f-1758-4a27-b458-63c87cf81d74", "context": "The problem of pose transfer, i.e., transferring the pose of a given person to a target one, has been previously approached using various generative adversarial networks (GANs). However, maintaining appearance consistency and shape consistency is a challenge.", "key_idea": "This paper proposes a new GAN composed of a sequence of Pose-Attentional Transfer Blocks that each transfers certain regions it attends to, while producing the person image progressively.", "method": "The proposed network's efficiency and efficacy are validated both qualitatively and quantitatively on the Market-1501 and DeepFashion datasets, and the generated person images are compared with results from previous studies.", "outcome": "The generated images possess better appearance consistency and shape consistency with the input images, making them appear more realistic compared to those produced by previous methods. The proposed network also demonstrates the capability to generate training images for person re-identification.", "future_impact": "The proposed network architecture can be used to generate training images for person re-identification activities, helping to mitigate issues around data insufficiency in this field.", "venue": "CVPR", "year": 2019, "title": "Progressive Pose Attention Transfer for Person Image Generation"} +{"pid": "93cb5b58-4e63-4f57-9019-dc5942497ef1", "context": "Generative models of 3D human motion are often restricted to a small number of activities and therefore do not generalize well to novel movements or applications. Most work on deep learning for sequence prediction focuses on video and speech, not taking into account the structure of skeletal data.", "key_idea": "The authors propose a deep learning framework for human motion capture data that learns a generic representation from a large corpus of motion capture data and can generalize well to new and unseen movements.", "method": "The authors use an encoding-decoding network that learns to predict future 3D poses from the most recent past, then extract a feature representation of human motion. They present and evaluate different network architectures that make different assumptions about time dependencies and limb correlations. The learned features are quantified, used for action classification and the receptive fields of the network units are visualized.", "outcome": "The proposed method outperforms the recent state of the art in skeletal motion prediction, even those that use action-specific training data. They demonstrate that deep feedforward networks trained from a generic motion capture (mocap) database can be used for feature extraction from human motion data.", "future_impact": "The learned representation can be used as a foundation for classification and prediction in the application of human motion data.", "venue": "CVPR", "year": 2017, "title": "Deep Representation Learning for Human Motion Prediction and Classification"} +{"pid": "ad9d412e-1616-4d27-88c3-79e52578463d", "context": "The existing approaches for object recognition through the matching of invariant local image features are based on matching to individual training images.", "key_idea": "This paper presents a new method for combining multiple images of a 3D object into a single model representation, capable of recognizing 3D objects from any viewpoint and generalizing models to non-rigid changes.", "method": "The authors devise a system to cluster training images into an existing view representation or treat it as a new view based on the geometric accuracy of the match to previous model views. They also develop a new probabilistic model to reduce the false positive matches.", "outcome": "The proposed system was able to robustly recognize 3D objects in cluttered natural images in sub-second times.", "future_impact": "The new method could lead to improved robustness in object recognition by combining features acquired under a range of imaging conditions.", "venue": "CVPR", "year": 2001, "title": "Local feature view clustering for 3D object recognition"} +{"pid": "5dd6604a3a55ac78684ace9a", "context": "In late fusion, each modality is separately processed in a unimodal Convolutional Neural Network (CNN) stream and the scores of each modality are fused at the end. Due to its simplicity, late fusion is the predominant approach in many state-of-the-art multimodal applications.", "key_idea": "The authors introduce the Multimodal Transfer Module (MMTM), a neural network module that leverages knowledge from multiple modalities in convolutional neural networks and permits slow modality fusion.", "method": "Using squeeze and excitation operations, the authors use the MMTM to recalibrate the channel-wise features in each CNN stream. This method is tested on four datasets spanning dynamic hand gesture recognition, speech enhancement, and action recognition with RGB and body joints.", "outcome": "Experimental results show that the proposed framework improves the recognition accuracy of well-known multimodal networks, demonstrating state-of-the-art or competitive performance on the tested datasets.", "future_impact": "The authors propose that the MMTM could be added among unimodal branches with minimum changes in their network architectures, allowing each branch to be initialized with existing pretrained weights. This implies potential for versatile application in multimodal research.", "venue": "CVPR", "year": 2020, "title": "MMTM: Multimodal Transfer Module for CNN Fusion"} +{"pid": "c0be18fa-4466-47a7-9458-1139cbd2dffa", "context": "There is a lack of theoretical analysis for characterizing the shadows cast by a point light source in relation to its position to the camera and the complexities of handling unusual camera-light configurations.", "key_idea": "This paper provides a mathematical characterization of the shadows cast by a point light source considering the epipolar geometry of camera-light pairs and including unusual and convenient camera-light configurations.", "method": "The developed theory is applied to compute a lower bound on the number of lights needed to extract all depth discontinuities from a general scene using a multiflash camera. Experiments with depth edges computed using two-flash setups, a four-flash setup, and a configuration with a flash at the camera's center of projection are used to illustrate the theory.", "outcome": "Theoretical analysis successfully provides a characterization of which discontinuities are missed and which are accurately detected by the algorithm. Experimental results validate the theory and its applicability for choosing an optimal light placement.", "future_impact": "The theoretical groundwork presented will provide a solution for some degenerate cases, and it could potentially be used to guide optimal light placement in the context of multiflash cameras and similar applications.", "venue": "CVPR", "year": 2008, "title": "Characterizing the shadow space of camera-light pairs"} +{"pid": "a4167022-b7af-48fc-8af5-92125fdf3173", "context": "Many computer vision challenges require continuous outputs, but are solved by discrete classification due to its natural containment within a probability n-simplex as defined by the softmax activation function. Regular regression lacks this closed geometry, leading to unstable training and suboptimal local minima.", "key_idea": "The authors propose a new continuous output method, spherical regression. This leverages the fact that many continuous output problems in computer vision are naturally contained in closed geometric manifolds, such as n-spheres. The authors introduce a spherical exponential mapping on n-spheres at the regression output to obtain well-behaved gradients for stable training.", "method": "The authors experimented on several computer vision challenges, including viewpoint estimation, surface normal estimation and 3D rotation estimation to validate the spherical regression method.", "outcome": "The experiments demonstrated the benefits of spherical regression in all tested scenarios - viewpoint estimation, surface normal estimation, and 3D rotation estimation.", "future_impact": "The spherical regression method, by providing stable training and solving continuous output problems in a geometrically intuitive way, could offer significant improvements to a variety of computer vision tasks.", "venue": "CVPR", "year": 2019, "title": "Spherical Regression: Learning Viewpoints, Surface Normals and 3D Rotations on N-Spheres"} +{"pid": "62c2a55b5aee126c0fcc8d40", "context": "Metrics for evaluating generative models measure the discrepancy between real and generated images. The commonly-used Fr\u00e9chet Inception Distance (FID) metric, extracts \u201chigh-level\u201d features using a deep network from both sets. However, variations induced by low-level preprocessing including image resizing and compression have not been properly investigated.", "key_idea": "The paper shows that commonly overlooked low-level image processing steps, such as image resizing with a fixed-width prefilter and lossy compression like JPEG, can lead to variations and unforeseen consequences in feature extraction and algorithm performance.", "method": "The authors investigate the effect of image resizing methods and lossy compression on the performance of generative models and on the FID metric used to evaluate them. They also release a reference implementation to facilitate future comparisons.", "outcome": "The paper finds that improperly implemented resizing can introduce aliasing artefacts, affecting downstream feature extraction. It also shows that if compression is used on training images, FID scores might actually improve if the generated images are also subsequently compressed.", "future_impact": "This paper aims to draw attention to the underappreciated role of low-level image processing in generative modeling. By characterizing variations in generative modeling development pipelines and providing recommendations based on signal processing principles, future work can better account for these subtleties and improve comparative analyses.", "venue": "CVPR", "year": 2022, "title": "On Aliased Resizing and Surprising Subtleties in GAN Evaluation"} +{"pid": "d59fbddb-6cef-490a-97a5-8f7897372b26", "context": "Cross-country navigation presents significant perception challenges due to the variable terrain.", "key_idea": "The authors propose an operational perception system for cross-country navigation that transforms laser range data into a Cartesian Elevation Map (CEM), which a detailed vehicle model uses to produce traversability information along selected trajectories.", "method": "The authors demonstrate their approach by running it on the autonomous land vehicle at Martin Marietta Aerospace Corporation in Denver.", "outcome": "The authors successfully demonstrated their obstacle detection and avoidance algorithms in both simulated and real-world environments.", "future_impact": "The authors suggest that enhancements to their current system can be made, but they do not specify what these enhancements may be.", "venue": "CVPR", "year": 1988, "title": "An operational perception system for cross-country navigation"} +{"pid": "550f11f3-a7c4-4b39-8163-7f729f362243", "context": "The status quo in human motion prediction includes models that specialize in either short-term or long-term prediction, usually involving significant computation.", "key_idea": "The study introduces novel neural temporal models for predicting and synthesizing human motion featuring a two-level processing architecture, easily computable features with derivative information, and a new multi-objective loss function. It also presents a new metric, Normalized Power Spectrum Similarity (NPSS), for evaluating the long-term predictive ability of motion synthesis models.", "method": "The proposed system was tested for its ability to model long-term motion trajectories. Also, a user study was conducted to compare NPSS's correlation with human evaluation of long-term motion to that of the mean-squared error measure.", "outcome": "The proposed system demonstrated improved modeling of long-term motion trajectories. The user study found that the NPSS metric correlated more strongly with human evaluation of long-term motion than the mean-squared error measure.", "future_impact": "The proposed model with the new NPSS metric might set a new standard for evaluating the long-term predictive ability of motion synthesis models.", "venue": "CVPR", "year": 2019, "title": "A Neural Temporal Model for Human Motion Prediction"} +{"pid": "61a444b35244ab9dcb6e227d", "context": "The conventional approach to video captioning is to learn from offline-extracted dense video features usually obtained at a fixed frame rate and often trained on image/video understanding tasks without adaptation to video captioning data.", "key_idea": "The authors propose SwinBERT, an end-to-end transformer-based model for video captioning, which accepts video frame patches directly as inputs and outputs a natural language description. The model uses a video transformer to encode spatial-temporal representations that can adapt to variable lengths of video input without designed for different frame rates. The authors also propose adaptively learning a sparse attention mask to reduce redundancy in consecutive video frames and improve task-specific performance.", "method": "The authors conducted extensive experiments on five video captioning datasets to test the SwinBERT model and its effectiveness when using a learned sparse attention mask.", "outcome": "SwinBERT demonstrated significant across-the-board performance improvements over previous methods, often by a large margin. The learned sparse attention masks additionally helped achieve new state-of-the-art results and showed adaptability between different video lengths and datasets.", "future_impact": "The proposed model and its inherent adaptive learning of a sparse attention mask could redefine the state-of-the-art standard for video captioning tasks, and the masks' adaptability may enhance transfer learning viability across different video lengths and datasets in future applications.", "venue": "CVPR", "year": 2022, "title": "SwinBERT: End-to-End Transformers with Sparse Attention for Video Captioning"} +{"pid": "5ae11fa8-bffe-47e6-9f69-95d673e47739", "context": "For intelligent transportation systems to operate in complex physical and social environments, they need to understand and learn how humans drive and interact with traffic scenes. However, sufficient real-life data for such learning is lacking.", "key_idea": "The paper introduces the Honda Research Institute Driving Dataset (HDD), a dataset consisting of 104 hours of real human driving in the San Francisco Bay Area, collected for the purpose of studying driver behavior in real-life environments.", "method": "An instrumented vehicle equipped with various sensors is used to collect data, which is then analyzed in comparison with other driving datasets. Additionally, the authors introduce a novel annotation methodology for untrimmed data sequences and employ baseline algorithms for driver behavior detection.", "outcome": "The paper provides a detailed analysis of the HDD dataset in comparison to other driving datasets. The feasibility of using the HDD dataset for studying driver behavior is established via testing baseline algorithms.", "future_impact": "The provided dataset and the new annotation methodology are expected to stimulate further research in understanding driver behavior from real-world driving situations, contributing to advancements in intelligent transportation systems.", "venue": "CVPR", "year": 2018, "title": "Toward Driving Scene Understanding: A Dataset for Learning Driver Behavior and Causal Reasoning"} +{"pid": "c77a456d-ddf5-48bd-b462-4b2a2fd3addd", "context": "The problem of estimating structure and motion from long continuous image sequences has not been effectively addressed in prior research.", "key_idea": "The authors propose an approach to estimate structure and motion from long continuous image sequences, by applying the expectation maximization algorithm based on an extended Kalman smoother to impose time-continuity of the motion parameters.", "method": "The authors' solution involves repeatedly estimating the state transition matrix of the dynamic equation and the parameters of noise processes in dynamic and measurement equations for a real image sequence.", "outcome": "This approach gives maximum likelihood estimates of the motion and structure parameters for a long video-rate image sequence with partially unknown system equation and noise.", "future_impact": "This research is vital for dealing with long video-rate image sequences with partially unknown system equations and noise.", "venue": "CVPR", "year": 2001, "title": "Structure and motion estimation with expectation maximization and extended Kalman smoother for continuous image sequences"} +{"pid": "804983ef-00e2-4579-ba33-600c2e5110a0", "context": "Sparse coding, which decomposes a vector using a few basis elements, is a common practice in machine learning and image processing. Its effectiveness is largely determined by the basis set or dictionary which is typically learned from specific data. Traditional dictionaries are unstructured flat sets of atoms, with no particular structure connecting the atoms.", "key_idea": "The proposed idea is to leverage structured dictionaries obtained from an epitome, meaning that the atoms are all patches of a chosen size within this epitome image. This leads to a substantial reduction in the number of parameters that need to be learned, as well as providing sparse image decompositions with shift-invariance properties.", "method": "The authors introduced a new formulation and an algorithm for learning structured dictionaries associated with epitomes and validated it in the context of image de-noising tasks.", "outcome": "The proposed method enables a considerable reduction in the number of parameters that need to be learned and allowed to de-noise images effectively.", "future_impact": "The impact envisioned by the authors is the use of epitomes and the corresponding structured dictionaries in various image processing tasks such as image de-noising.", "venue": "CVPR", "year": 2011, "title": "Sparse image representation with epitomes"} +{"pid": "615d13fe5244ab9dcb637c37", "context": "There is a growing interest in understanding document content with novel deep learning architectures. However, document understanding tasks require dense information annotations, which are costly to scale and generalize. Active learning techniques have been proposed to reduce the overall cost of annotation while maintaining the performance of the underlying deep learning model.", "key_idea": "OPAD, a novel framework using reinforcement policy for active learning in content detection tasks for documents is proposed. This framework learns the acquisition function to pick samples, while also optimizing performance metrics typically associated with the tasks.", "method": "The OPAD framework is applied to various tasks related to document understanding like layout parsing, object detection and named entity recognition. Additionally, weak labeling scenarios are considered, along with novel rewards to account for class imbalance and user feedback in the annotation interface. Ablation studies for human feedback and class imbalance rewards are presented.", "outcome": "The proposed OPAD framework showed superior performance for active learning for various tasks related to document understanding. A comparison of annotation times for different approaches is also provided.", "future_impact": "The OPAD framework can potentially reduce the cost of annotation significantly, while improving the active learning method in document understanding tasks.", "venue": "CVPR", "year": 2022, "title": "OPAD: An Optimized Policy-based Active Learning Framework for Document Content Analysis"} +{"pid": "61c2974b5244ab9dcbcf1ead", "context": "In medical image segmentation, models trained on specific types of images (e.x., MR images) often perform poorly when asked to segment different types of images (e.x., CT images). The cross-modality segmentation problem represents a significant challenge in domain adaptation.", "key_idea": "The authors propose a dual-normalization model for generalizable cross-modality image segmentation which leverages source-similar and source-dissimilar images augmented through nonlinear transformation in unknown target domains.", "method": "The proposed method utilizes independent batch normalization layers for separate normalization and a style-based selection scheme to choose the right processing path in the test stage. The approach was validated through experiments on three publicly available datasets: BraTS, Cross-Modality Cardiac, and Abdominal Multi-Organ datasets.", "outcome": "Experiments on multiple datasets show that this method outperforms other state-of-the-art domain generalization methods in cross-modality image segmentation tasks.", "future_impact": "This novel dual-normalization model could make significant strides in cross-modality image segmentation, showing vast clinical potential for improving patient care and outcomes.", "venue": "CVPR", "year": 2022, "title": "Generalizable Cross-modality Medical Image Segmentation via Style Augmentation and Dual Normalization"} +{"pid": "2dec67d0-1fa7-4f8a-911a-a4f54ff8a338", "context": "Recent research is focusing on automatic 3D reconstructions of urban environments from video acquired at street level. These systems record vast amounts of video, requiring a stereo matcher capable of processing data at speeds comparable to the recording frame rate. Urban environments exhibit mostly planar surfaces, often imaged at oblique angles, and pose a challenge for many window-based stereo matchers that perform poorly with slanted surfaces.", "key_idea": "The authors propose a multi-view plane-sweep-based stereo algorithm that correctly handles slanted surfaces. The designed algorithm identifies the scene's principle plane orientations, estimates depth for each direction, and combines the results of each sweep.", "method": "The proposed algorithm incorporates priors on the locations of planes in the scene to enhance the quality of reconstruction and reduce computation time, especially for uniform textureless surfaces. The team demonstrate the algorithms efficacy on a variety of scenes.", "outcome": "The paper shows improved accuracy in handling slanted surfaces by demonstrating the effectiveness of the proposed algorithm in a variety of scenes.", "future_impact": "The ability to handle slanted surfaces correctly and in real-time could improve 3D reconstruction of urban environments, specifically benefiting applications needing rapid processing of massive video data.", "venue": "CVPR", "year": 2007, "title": "Real-Time Plane-Sweeping Stereo with Multiple Sweeping Directions"} +{"pid": "9625a497-2a28-4e6a-8056-7fd9663c4d88", "context": "Interacting naturally with a computer-generated display through the use of gestures is a growing field. An issue to this mode of interaction, particularly in unrestricted setups, is tracking the human arm robust in video sequences captured via cameras.", "key_idea": "This study suggests a new approach integrating a multi-cue based localization scheme with a tracking framework to reliably track the human arm dynamics in unrestricted environments. The cues fused are motion, shape, and color for identifying key image features.", "method": "The authors test their system using an Extended Kalman Filter, modified to utilize the articulated structure of the arm. They propose an interaction scheme between tracking and localization to improve the estimation process while reducing computational requirements.", "outcome": "The proposed tracking system shows improved performance, as validated through extensive experiments and simulations.", "future_impact": "The system proposed in this paper could bring significant improvements in natural human-computer interaction through realistic and efficient gesture recognition in unrestricted environments.", "venue": "CVPR", "year": 1998, "title": "Reliable tracking of human arm dynamics by multiple cue integration and constraint fusion"} +{"pid": "33bfabcf-96ba-4fba-943a-e0840e00c58e", "context": "Analysis of human eye fixations on natural images revealed that human attention is sparsely distributed and usually deployed to locations with abundant structural information. However, the current framework for modeling saccadic eye movements and visual saliency fails to adequately represent these findings.", "key_idea": "This study proposes a unified statistical framework inspired by novel observations on human attention distribution for modeling both saccadic eye movements and visual saliency based on Super Gaussian Component (SGC) analysis.", "method": "The model sequentially obtains SGC using projection pursuit and generates eye-movements by selecting the location with maximum SGC response. Its effectiveness and robustness over state-of-the-arts is demonstrated by carrying out dense experiments on psychological patterns and human eye fixation benchmarks.", "outcome": "The results demonstrate the superior effectiveness and robustness of the proposed model over existing state-of-the-art techniques in saccadic behavior simulation and stimulation of human eye fixations.", "future_impact": "The results show promising potential for the use of statistical approaches in human behavior research.", "venue": "CVPR", "year": 2012, "title": "What are we looking for: Towards statistical modeling of saccadic eye movements and visual saliency"} +{"pid": "bc6d66d3-7f37-4669-a6e2-1ab713b83423", "context": "Traditional image search methods rely on the user providing binary relevance feedback, which may not always accurately reflect their preferences or provide an efficient search.", "key_idea": "A novel feedback mode for image search is introduced, where the user can adjust properties of exemplar images according to their needs. An offline approach is implemented that learns a set of ranking functions, each predicting the relative strength of a nameable attribute ('sportiness', 'furriness', etc.) in an image.", "method": "The system presents an initial set of reference images at query time, and the user provides relative attribute feedback. This feedback is used to update the relevance function, re-rank the pool of images, and iterate until the top ranked images closely match the user's envisioned target.", "outcome": "The proposed technique performs well in refining image searches for people, products, and scenes and outperforms binary relevance feedback in terms of search speed and accuracy.", "future_impact": "The method introduced allows users to 'whittle away' irrelevant portions of the visual feature space and use semantic language to communicate more accurately and efficiently with the search system, thereby improving the precision of search results.", "venue": "CVPR", "year": 2012, "title": "WhittleSearch: Image search with relative attribute feedback"} +{"pid": "7190ccc5-bf40-4be2-8cd2-6cf11fbec324", "context": "Recent successes in single image super-resolution (SISR) have been driven by deep convolutional neural networks (CNNs). However, existing CNN-based SISR methods make assumptions about the degradation process of the low-resolution image, leading to poor performance when these assumptions are inaccurate, and they lack scalability to handle multiple degradations.", "key_idea": "The authors introduce a general framework with a dimensionality stretching strategy that allows a single convolutional super-resolution network to account for the blur kernel and noise level in the SISR degradation process, which enables it to handle various degradations.", "method": "The authors conduct experiments on synthetic and real low-resolution images using the proposed convolutional super-resolution network that intakes two key factors of the SISR degradation process as input.", "outcome": "The proposed super-resolution network shows favorable results on multiple degradations and proves to be computationally efficient.", "future_impact": "The proposed method could provide a highly effective and scalable solution for practical SISR applications.", "venue": "CVPR", "year": 2018, "title": "Learning a Single Convolutional Super-Resolution Network for Multiple Degradations"} +{"pid": "3d35f821-6a2e-43bf-b83b-14988f53fef6", "context": "The classical problem in computer vision is determining the pose of a calibrated perspective camera given the image projections of three known world points, assuming a central perspective projection, for which up to four pairs of solutions exist.", "key_idea": "The authors tackle a generalized variant of this problem, which doesn't assume a central perspective projection, instead dealing with rays sampled from the camera in an arbitrary, known fashion, with the task of finding the camera pose making these rays meet three known world points.", "method": "The authors propose an efficient numerical algorithm, which reduces the problem to finding the intersections between a circle and a ruled quartic surface (corresponding to the up to eight possible solutions), expressed as the roots of an octic polynomial.", "outcome": "The proposed algorithm has been designed to solve the pose of any type of calibrated camera or camera rig, in the classical case allowing up to four pairs of solutions to be obtained in closed form by exploiting a reflection symmetry that causes the odd order terms of the octic polynomial to vanish.", "future_impact": "The proposed algorithm is intended to be used in a hypothesize-and-test architecture, expanding its applicability beyond the traditional central perspective projection scenarios.", "venue": "CVPR", "year": 2004, "title": "A minimal solution to the generalised 3-point pose problem"} +{"pid": "21f1d913-1c64-4da2-96c8-38205aa68495", "context": "Existing text detection systems in natural scenes have limitations in accuracy, relying mainly on a single type of feature.", "key_idea": "The authors propose a novel, data-driven system that incorporates cascaded classification using coarse-to-fine detection of character pixels with convolutional features (Text-Conv), connected component extraction using edge and color features, and a graph-based segmentation of connected components into words (Word-Graph).", "method": "The authors use the ICDAR 2015 Robust Reading Focused Scene Text dataset to train their system. The detection stages are trained using different definitions for foreground text regions, some based on bounding box intersection, and others on bounding box and pixel intersection.", "outcome": "The proposed system achieves pixel, character, and word detection f-measures of 93.14%, 90.26% and 86.77% respectively on the ICDAR 2015 Robust Reading Focused Scene Text dataset, outperforming state-of-the-art systems.", "future_impact": "The proposed approach has the potential to be applied to other detection targets with homogenous color in natural scenes.", "venue": "CVPR", "year": 2016, "title": "A Text Detection System for Natural Scenes with Convolutional Feature Learning and Cascaded Classification"} +{"pid": "a6eedddd-f59b-427f-8642-d41771aebac1", "context": "The problem of recovering motion parameters and point correspondences in computer vision, specifically within nonrigid motion analysis, is an area that has only recently begun to be addressed, despite its wide application potential.", "key_idea": "A novel method is proposed for estimating motion parameters and point correspondences between surfaces under small nonrigid deformations. The method utilizes the changes in differential geometric properties of a surface undergoing motion.", "method": "Simulations are performed using nonrigid motion on ellipsoidal data to test the proposed algorithms, with further testing conducted on a sequence of facial range images. The algorithms are also applied to detect abnormalities in cardiac images.", "outcome": "The simulations demonstrated the effectiveness and accuracy of the proposed algorithms in estimating motion parameters and point correspondences under small nonrigid deformations.", "future_impact": "The proposed method could see applications in a wide range of areas requiring nonrigid motion analysis, including the detection of abnormalities in cardiac images.", "venue": "CVPR", "year": 1994, "title": "Determination of motion parameters and estimation of point correspondences in small nonrigid deformations"} +{"pid": "2ccfacfd-b100-4f81-90cf-68c50ee60f30", "context": "Face verification methods utilize deep convolutional neural networks (DCNNs) to extract low dimensional, discriminative, and high-level abstracted features. However, the training of DCNNs can be a challenge when there is insufficient training data.", "key_idea": "A new face verification method is proposed that employs multiple DCNNs and a deep ensemble, trained with triplet-based loss function using sets of reference, positive, and negative images. Each DCNN extracts two types of features from stacked multi-scale convolutional layer blocks, which are combined into a descriptor and used for training using the joint Bayesian learning method and its transfer learning variant.", "method": "The method is tested on the LFW face dataset, with four different types of networks, and using only 198,018 images. An additional check is performed by combining the proposed method with a high-dimensional LBP-based joint Bayesian method.", "outcome": "The proposed method, using the joint Bayesian learning and its transfer learning method, achieves an accuracy of 98.33% on the LFW. By augmenting this with the high-dimensional L\u03b2P-based joint Bayesian method, the accuracy reached 99.08%.", "future_impact": "The proposed methodology demonstrates the potential to significantly improve the accuracy of face verification, especially when there is insufficient training data to train DCNNs.", "venue": "CVPR", "year": 2017, "title": "Deep Convolutional Neural Network Using Triplets of Faces, Deep Ensemble, and Score-Level Fusion for Face Recognition"} +{"pid": "6232a74d5aee126c0fe13f27", "context": "Deep Metric Learning (DML) methods learn metric spaces but only solve binary class assignments and thus ignore higher-level semantic relations. As a result, these methods often encode incomplete semantic context and misrepresent the semantic relation between classes, impacting the model's ability to generalize.", "key_idea": "The authors propose a language guidance objective for visual similarity learning. By leveraging language embeddings of expert- and pseudo-classnames, they aim to provide more meaningful language semantics to enhance the semantic consistency of the visual representation space.", "method": "The authors conduct extensive experiments and ablations to validate their approach of integrating language guidance into deep metric learning.", "outcome": "Language guidance in Deep Metric Learning results in significant, model-agnostic improvements and achieves competitive and state-of-the-art results on all benchmarks.", "future_impact": "The use of language guidance for enhancing visual similarity learning opens up new possibilities for improving the contextual representation and generalizability of DML models. Its model-agnostic nature means it can be applied to various DML models for better performance.", "venue": "CVPR", "year": 2022, "title": "Integrating Language Guidance into Vision-based Deep Metric Learning"} +{"pid": "c4ed05df-10ae-4b4a-9a69-f6750453a1b5", "context": "The merit of online ensemble learning for classification problems is under exploration, and the importance of variables is usually estimated using batch training methods such as Gini index, ReliefF, and gain ratio.", "key_idea": "The authors propose an online random forests algorithm that estimates variable importance incrementally based on correlation ranking. They further introduce an 'incremental hill climbing' algorithm, where features are added in a 'forward step (FS)' and removed in a 'backward step (BE)', combined in implementations called CorrFS and CorrBE respectively.", "method": "The authors test their method using UCI public databases and compare their performance with classifiers constructed from batch training.", "outcome": "The CorrFS and CorrBE implementations have shown to achieve comparable performance to classifiers constructed from batch training.", "future_impact": "This online random forests algorithm that uses CorrFS and CorrBE can be used for online ensemble learning for classification problems, providing a new perspective on feature selection approaches in comparison to batch mode methods.", "venue": "CVPR", "year": 2008, "title": "Online random forests based on CorrFS and CorrBE"} +{"pid": "f475ad7e-60a5-4e16-8af0-67e96642450b", "context": "Multilinear algebra is a powerful theoretical tool for visual geometry. However, the widespread use of traditional typographical notation often makes its inherent simplicity and conceptual elegance less accessible.", "key_idea": "The authors propose to use standard tensor diagrammatic techniques adapted to the specific requirements of visual geometry, in order to provide a simpler and more intuitive understanding of geometric relations, which are represented by circuits that can be manipulated using simple rules.", "method": "The proposed approach is illustrated through several constructions, including straightforward derivations of the standard multiview relations (Fundamental Matrix, Trifocal and Quadrifocal Tensors), and nearly mechanical procedures for camera extraction.", "outcome": "The authors show that through the use of tensor diagrammatic techniques, multiview geometric relations can be represented more intuitively and manipulated more easily, enhancing their understanding and ease of use.", "future_impact": "This new approach in handling multilinear algebra and visual geometry carries the potential to facilitate further research in this field or similar fields that require understanding and manipulation of complex geometric relations.", "venue": "ICCV", "year": 2009, "title": "Diagram techniques for multiple view geometry"} +{"pid": "3910cb0d-3665-450a-a54e-e138b282ba38", "context": "The process of accurate indoor 3D reconstruction involves scene layout estimation and global registration problems. Traditional methods tackle these challenges separately, possibly losing certain inter-dependent information which could be valuable for improving reconstruction results.", "key_idea": "The paper proposes a novel method that jointly solves scene layout estimation and global registration problems, which works by alternating between layout estimation and layout-based global registration in an iterative fashion.", "method": "The authors experimentally test their method with both synthetic and real-world datasets. They build scene fragments using KinectFusion, register them through pose graph optimization, extract the scene layout through hierarchical agglomerative clustering and energy-based multi-model fitting, and register all the range data through the global iterative closest point algorithm.", "outcome": "The authors' method demonstrated successful indoor 3D reconstruction performance, but specific quantitative or qualitative results are not detailed in the abstract.", "future_impact": "The proposed method could potentially improve the accuracy of indoor 3D reconstruction tasks and inspire integrated approaches in solving inter-dependent problems in computer vision.", "venue": "ICCV", "year": 2017, "title": "Joint Layout Estimation and Global Multi-view Registration for Indoor Reconstruction"} +{"pid": "e66989a1-25b0-4074-86c1-5acc4078468b", "context": "Articulated hand pose estimation is traditionally challenged by noisy data and occlusions. Existing approaches that primarily use synthetic data in training also face performance issues due to discrepancies between realistic and synthetic pose data.", "key_idea": "This study presents the first semi-supervised transductive algorithm for real-time articulated hand pose estimation, known as Semi-supervised Transductive Regression (STR) forest, which learns the connection between a small labelled realistic dataset and a larger synthetic dataset, and a new data-driven pseudo-kinematic technique is proposed for refining noisy or occluded joints.", "method": "The authors create the STR forest and use it to learn relationships between datasets. They also implement a pseudo-kinematic technique to refine articulations, specifically targeting noisy and occluded joints.", "outcome": "Experimental results suggest that the proposed STR forest and pseudo-kinematic technique outperform state-of-the-art approaches for hand pose estimation in terms of accuracy, robustness, and speed. It also effectively deals with noise and occlusions.", "future_impact": "This work suggests potential for improving real-time articulated hand pose estimation through the use of transductive learning, unlabelled data, and pseudo-kinematic techniques for refining articulations.", "venue": "ICCV", "year": 2013, "title": "Real-Time Articulated Hand Pose Estimation Using Semi-supervised Transductive Regression Forests"} +{"pid": "6528c1f1939a5f4082995b5e", "context": "Frame quality deterioration is a challenge in video understanding and to alleviate the information loss due to deteriorated frames, current techniques use complex transformer-based integration modules. However, these models are specifically tailored towards a single task, making it difficult to generalize.", "key_idea": "This paper introduces a new framework dubbed Spatio-Temporal Prompting Network (STPN) that dynamically adjusts the input features in the backbone network to efficiently extract robust and accurate video features by predicting video prompts that contain spatio-temporal information.", "method": "The STPN prepends predicted video prompts to the patch embeddings of the current frame as the updated input for video feature extraction. Its robustness and accuracy is validated on three varied video understanding tasks: video object detection, video instance segmentation, and visual object tracking.", "outcome": "Without making any application-specific modifications, STPN demonstrated state-of-the-art performance on ImageNetVID for video object detection, YouTubeVIS for video instance segmentation, and GOT-10k for visual object tracking.", "future_impact": "The introduced Spatio-Temporal Prompting Network, being not task-specific, has the potential to be effectively applied to a variety of video-based tasks, simplifying the feature extraction process in video understanding.", "venue": "ICCV", "year": 2023, "title": "Spatio-temporal Prompting Network for Robust Video Feature Extraction."} +{"pid": "6064451791e011538305ceff", "context": "Semantic labelling is highly associated with geometry and radiance reconstruction, with similar shapes and appearances more likely to derive from similar classes. However, self-supervised approaches are not applicable to semantics due to the requirement of human-defined labels despite recent advances in implicit neural reconstruction techniques.", "key_idea": "The authors expand neural radiance fields (NeRF) to jointly encode semantics with appearance and geometry, allowing accurate 2D semantic labels to be obtained using minimal in-place annotations specific to the scene.", "method": "The authors use the intrinsic multi-view consistency and smoothness of NeRF to allow sparse labels to efficiently propagate. They test this approach in scenarios where labels are either sparse or very noisy in room-scale scenes.", "outcome": "The novel approach proves advantageous in various applications. These include efficient scene labelling tools, novel semantic view synthesis, label denoising, super-resolution, label interpolation and multi-view semantic label fusion in visual semantic mapping systems.", "future_impact": "Given its versatility across multiple applications, this approach promises to contribute significantly to the field of semantic labelling and present a variety of future research opportunities in areas requiring efficient and accurate semantic labelling.", "venue": "ICCV", "year": 2021, "title": "In-Place Scene Labelling and Understanding with Implicit Scene Representation."} +{"pid": "2b8c6432-6445-4bdf-a60b-2f26ead53c4e", "context": "While large-scale Deep Neural Networks (DNNs) have proven successful in various computer vision tasks, their high computational intensity presents challenges for deployment on resource-limited systems. Existing low-rank basis approximations of filters have been used to quicken testing without sacrificing performance. However, the effective training of DNNs toward lower-rank space for more efficient DNNs remains an open area of investigation.", "key_idea": "To address these challenges, the authors propose Force Regularization. This technique enforces filters to coordinate more weight information into a lower-rank space, theoretically and practically demonstrating that standard Low-Rank Approximations (LRAs) can reconstruct filters using much lower bases, resulting in faster DNNs.", "method": "The authors comprehensively evaluate the effectiveness of their approach in ResNets, AlexNet, and GoogLeNet. They also assess Force Regularization in terms of both the speed of the DNNs and the stability of the fine-tuning process.", "outcome": "Force Regularization results in ResNets, AlexNet, and GoogLeNet seeing significant speed improvements, for example, a 2\u00d7 speedup on modern GPUs without accuracy loss and 4:05\u00d7 speedup on CPU, albeit with minor accuracy degradation. Also, Force Regularization improves the initialization of low-rank DNNs, allowing for faster convergence toward higher accuracy.", "future_impact": "The lower-rank DNNs obtained through this method can be further sparsified, thereby proving that Force Regularization can be integrated with state-of-the-art sparsity-based acceleration methods, potentially opening up new avenues for future research.", "venue": "ICCV", "year": 2017, "title": "Coordinating Filters for Faster Deep Neural Networks"} +{"pid": "ee45cc7b-9508-458c-9cf0-31a75a958c61", "context": "In many computer vision applications, modeling and synthesizing image noise is crucial. The traditional additive white Gaussian and heteroscedastic (signal-dependent) noise models offer only a rough approximation of actual sensor noise.", "key_idea": "The paper introduces Noise Flow, a powerful and precise noise model based on recent normalizing flow architectures. Noise Flow merges basic parametric noise models with the versatility and expressiveness of normalizing flow networks.", "method": "The authors use a normalizing flow network to create a compact, comprehensive noise model capable of representing multiple cameras and gain factors.", "outcome": "Noise Flow significantly outperforms existing noise models, improving the likelihood of sampled noise by 52% and providing a 0.42 nats/pixel improvement over the camera-calibrated noise level functions.", "future_impact": "Noise Flow's first application of deep learning and data-driven noise distributions to go beyond simple parametric models may pave the way for more advanced noise models in the future.", "venue": "ICCV", "year": 2019, "title": "Noise Flow: Noise Modeling With Conditional Normalizing Flows"} +{"pid": "610cb85c5244ab9dcb20c0cf", "context": "Conventional unsupervised feature representation methods assign high similarity to close instance pairs in the feature space, typically causing incorrect pairwise relationships for large neighborhoods because the Euclidean distance doesn't effectively represent true semantic similarity on the feature manifold.", "key_idea": "The authors propose an instance similarity learning (ISL) method that mines the feature manifold in an unsupervised way, discovering the semantic similarity among instances to gain discriminative representations. Generative Adversarial Networks (GAN) are specifically employed to mine the underlying feature manifold.", "method": "Generated features are used as proxies to progressively explore the feature manifold, allowing the acquisition of semantic similarity among instances as reliable pseudo supervision. This method's effectiveness is evaluated through extensive experiments on image classification.", "outcome": "The experimental results demonstrate that the proposed method outperforms the state-of-the-art methods in unsupervised feature representation for image classification.", "future_impact": "With the code available for use, this work could enable further advancements in developing discriminative feature representations for various unsupervised learning tasks.", "venue": "ICCV", "year": 2021, "title": "Instance Similarity Learning for Unsupervised Feature Representation."} +{"pid": "60d15b3891e011c16f0cb4ad", "context": "RANSAC-type robust estimators and their variants have been primarily used for hypothesis verification in two-view geometry but have limitations in efficiency and accuracy, including a high incidence of incorrect models and significant computational overhead.", "key_idea": "The paper intoduces VSAC, a novel RANSAC-type robust estimator which distinguishes by its unique concept of independent inliers for more effective dominant plane handling and error-free rejection of incorrect models, and its sophisticated local optimization process that runs on average only once.", "method": "The efficacy, speed, and accuracy performance of VSAC were validated through experiments on four standard datasets: EVD, HPatches, PhotoTourism, and Kusvod2.", "outcome": "VSAC performed significantly faster than existing estimators, with an average speed of 1-2ms on a CPU, two orders of magnitude faster than MAGSAC++, the most accurate estimator of two-view geometry to date. It also demonstrated near error-free performance with no failures in repeated runs across tested datasets.", "future_impact": "Notably exhibiting great efficiency and precision, VSAC could provide significant improvements in applications relating to two-view geometry estimations.", "venue": "ICCV", "year": 2021, "title": "VSAC - Efficient and Accurate Estimator for H and F."} +{"pid": "6528c1f8939a5f4082996fcf", "context": "Vision Transformers (ViTs), especially the Fully Attentional Network (FAN), have shown robustness against out-of-distribution scenarios and have achieved state-of-the-art robustness.", "key_idea": "The authors propose an improved pre-training method for FAN models using a Self-emerging Token Labeling (STL) framework, which includes a two-stage training process. The process first generates semantically meaningful patch token labels using a FAN token labeler, followed by a student model training stage that uses both token labels and original class labels.", "method": "A FAN model based on FAN-L-Hybrid was trained using the STL framework. The performance of this model was tested on ImageNet-1K, ImageNet-C, and ImageNet-A datasets and compared with the original FAN model. Additionally, the performance of the model was evaluated on downstream tasks such as semantic segmentation.", "outcome": "With the STL framework, the FAN-L-Hybrid model achieved a new state-of-the-art performance for ImageNet-A, significantly outperforming the original FAN counterpart. The framework also improved performance on downstream tasks such as semantic segmentation, with up to a 1.7 margin compared to the counterpart model.", "future_impact": "The proposed STL framework could potentially lead to an improvement in the generalization and robustness of Vision Transformers, particularly in out-of-distribution scenarios and semantic segmentation tasks.", "venue": "ICCV", "year": 2024, "title": "Fully Attentional Networks with Self-emerging Token Labeling"} +{"pid": "bd304930-0622-4bc8-90c6-1e0f953b9852", "context": "The problem of tracking a face and its facial features in an input video sequence necessitates an effective solution.", "key_idea": "The authors propose a face and facial feature tracking system using the active appearance algorithm that adapts a wireframe model to the face in each frame.", "method": "The adaptation parameters derived from the wireframe model are converted to MPEG-4 facial animation parameters in the proposed system.", "outcome": "The early results of implementing this system are promising.", "future_impact": "The authors suggest that this work should be continued to aim for creating a real-time model-based coder.", "venue": "ICCV", "year": 2001, "title": "Using the active appearance algorithm for face and facial feature tracking"} +{"pid": "239ada84-ac30-4aac-b680-83846b169f06", "context": "Existing methods to generate textual descriptions of videos are typically limited to specific categories of events and actions, and there's a need for automatically describing a wider range of video events.", "key_idea": "A novel method is proposed comprising of an action classifier and a description generator. The action classifier detects and classifies actions in videos to be used as verbs in descriptions, while the description generator finds the actors and generates sentences based on the verb, subject, and direct and indirect objects.", "method": "The proposed method exploits the discriminative power of a bag-of-features action detector coupled with the generative power of a rule-based action descriptor. It is then compared against a homogeneous setup with a rule-based action detector and action descriptor.", "outcome": "The proposed hybrid method, that integrates a discriminative action detector and a generative descriptor, outperforms a homogeneous setup that uses rule-based action detector and descriptor.", "future_impact": "This work opens up possibilities for generating textual descriptions for a wider range of video content based on detected human actions, resulting in substantial improvements to video explanation, video search, and accessibility applications.", "venue": "ICCV", "year": 2012, "title": "Automated textual descriptions for a wide range of video events with 48 human actions"} +{"pid": "64d9a6873fda6d7f061d37ac", "context": "Explainable artificial intelligence (XAI) seeks to understand the inner workings of complex deep neural models. However, XAI lacks ground-truth explanations, making its automatic evaluation a challenging task.", "key_idea": "The authors introduce a novel synthetic vision dataset, FunnyBirds, and corresponding automated evaluation protocols. This dataset facilitates semantically meaningful image interventions, such as the removal of individual object parts.", "method": "Using their proposed synthetic dataset and protocols, they carry out semantically meaningful image interventions, such as removing specific object parts, and evaluate the explanations of 24 different combinations of neural models and XAI methods.", "outcome": "Through the use of their tools, the authors demonstrate the strengths and weaknesses of the evaluated methods comprehensively and automatically, marking a significant step forward in the automatic evaluation of XAI methods.", "future_impact": "This new dataset and accompanying evaluation protocols will facilitate a better understanding of explainability methods on a deep level, potentially leading to improved transparency in machine learning models.", "venue": "ICCV", "year": 2023, "title": "FunnyBirds: A Synthetic Vision Dataset for a Part-Based Analysis of Explainable AI Methods"} +{"pid": "f5cf7d72-1efd-45d3-9f9f-f928f6665630", "context": "The application of modern deep networks in medical imaging is still rudimentary primarily due to the large size of three-dimensional images, requiring huge convolutional or fully connected layers. The situation worsens when dealing with longitudinal analysis of 3D image volumes through recurrent structures and when a reliability measure, rather than a point estimate of model parameters, is necessary.", "key_idea": "The authors propose the use of the tensor train decomposition, specifically the 'orthogonal' tensor train, for constructing networks with significantly fewer parameters. This approach allows for effective training of powerful recurrent networks on whole brain image volume sequences.", "method": "The authors utilize the 'orthogonal' tensor train to express a standard network layer both theoretically and empirically. They test its ability to effectively reconstruct whole brain volumes and compare its performance, in terms of faster convergence and stronger confidence intervals, to the standard tensor train decomposition.", "outcome": "The 'orthogonal' tensor train demonstrates the ability to effectively reconstruct whole brain volumes, showing faster convergence and stronger confidence intervals than the standard tensor train decomposition.", "future_impact": "The proposed 'orthogonal' tensor train method can enhance the application of deep learning in medical imaging by enabling effective training of powerful recurrent networks on whole brain image volume sequences. In addition, the method provides a direction for further research on how to handle voluminous data in the realm of scientific applications.", "venue": "ICCV", "year": 2019, "title": "Scaling Recurrent Models via Orthogonal Approximations in Tensor Trains"} +{"pid": "baa26caf-e853-4912-af5f-4f67d4805c1f", "context": "Nearest Neighbor (NN) search is important in computer vision algorithms. However, conducting NN search on the immense amount of image data available on the Internet is becoming an issue because the main memory of a single PC becomes insufficient.", "key_idea": "The authors propose an Approximate Nearest Neighbor (ANN) search on Hard Disk Drive (HDD) in order to handle large amounts of data, based on recently proposed Principal Component Hashing (PCH). In this solution, the hash bins are represented by the leaf nodes of a B+ tree to allow dynamic addition and deletion of data.", "method": "The authors validate the method through experiments using a standard PC and 10,000 stored images, measuring memory consumption, the first search time including the cold start-up time and successive search times.", "outcome": "The memory consumption is 42 times smaller, the first search time including the cold start-up time is 4.3 times faster and the mean search time decreases from 7.4 seconds to 0.64 seconds for successive searches. The authors also demonstrate the ability of the algorithm to perform NN search on 1 million image datasets with only 193MB memory consumption.", "future_impact": "The properties of this algorithm indicate it would be suitable for non-time-critical NN search applications and NN search engines used by web servers, where the search engine starts in response to occasional queries.", "venue": "ICCV", "year": 2009, "title": "Approximate nearest neighbor search on HDD"} +{"pid": "714f0ec9-acd4-4481-bc70-880f01db676b", "context": "Recent years have seen the use of contextual models that exploit maps for various recognition and localization tasks. However, these models could be further improved with additional information.", "key_idea": "The paper proposes to enhance freely available world maps using aerial images, formulating the problem as inference in a Markov random field parameterized in terms of the location of the road-segment centerlines and their width.", "method": "The authors use OpenStreetMap for the experiments and the efficiency of their inference method enables them to segment all OSM roads in the entire world in a single day using a small cluster of 10 computers.", "outcome": "Their approach, which was trained using only 1.5 km2 aerial imagery, was able to produce accurate results across the globe and outperformed the state-of-the-art in two new benchmarks collected by the authors.", "future_impact": "The enhanced maps produced by this approach can be of high utility in the future for semantic segmentation of ground images, suggesting further applications in related areas.", "venue": "ICCV", "year": 2015, "title": "Enhancing Road Maps by Parsing Aerial Images Around the World"} +{"pid": "7a587764-bcca-454a-9bd7-d6dc50d129ce", "context": "2D electrophoresis gel image processing is a widely used and affordable technology in biomedical research. However, there are challenges in the current methods leading to a need for improvements in the final software applications.", "key_idea": "The authors propose a comprehensive algorithm for gel image registration that covers all the steps, including robust and meaningful detection of spots, improvements on the computation of the distance between spots using shape contexts, and an iterative random sampling process for dealing with spot differences between images.", "method": "The algorithm for gel image registration is formulated and implemented, following the aforementioned steps, and the results are evaluated in the context of their meaningfulness.", "outcome": "The proposed algorithm successfully covers all the steps for gel image registration, providing a meaningful detection of spots, improved computation of distances between spots using shape contexts, and an effective iterative random sampling process for handling spot differences across images.", "future_impact": "The advanced image processing algorithm developed in this study can significantly improve the process of 2D electrophoresis gel image processing, making it easier and more cost-effective for researchers in biomedical sciences, particularly those operating under budget constraints.", "venue": "ICCV", "year": 2007, "title": "Processing of 2D Electrophoresis Gels"} +{"pid": "6528c207939a5f4082999ff5", "context": "The accuracy of the visual tasks for top-view fisheye cameras is currently limited due to the spatial distortion created by poses of objects", "key_idea": "The authors propose a Deformable Convolution based on the Poincar\u00e9 Ball (DCPB), an analogy between the fisheye model and the Poincar\u00e9 ball, which helps to alleviate the issue of spatial distortion in images.", "method": "The authors use DCPB to conduct Graph Convolutional Networks (GCN) in the Poincar\u00e9 ball, calculating the geodesic distances to Poincar\u00e9 hyperplanes as offsets and modulation scalars. Experiments are conducted using a public dataset called THEODORE.", "outcome": "The experimental results show that DCPB markedly improves the performance of the neural network and obtains higher accuracy.", "future_impact": "The efficiency of DCPB demonstrates potential for using temporal information in fisheye videos, possibly informing further research and advances in visual tasks involving distorted images.", "venue": "ICCV", "year": 2023, "title": "DCPB: Deformable Convolution based on the Poincar\u00e9 Ball for Top-view Fisheye Cameras."} +{"pid": "1f9bea69-f09e-44a3-9a4b-c93670b2f44b", "context": "Camera relocalization, i.e, determining the position and orientation of a camera within a reference frame given an image taken by the camera, has been a challenging problem in computer vision. Existing methods often require scene-specific training, which limits their scalability and ability to generalize to unseen scenes.", "key_idea": "The authors propose a new deep learning approach for camera relocalization. It involves using a convolutional neural network (CNN) to retrieve similar database images and predict the relative pose between the query and the database images. It then obtains the camera location for the query image via triangulation from two relative translation estimates using a RANSAC based approach.", "method": "The proposed system was trained using training image pairs. Evaluation of the approach was done using the authors' own dataset and the standard 7 Scenes benchmark.", "outcome": "The results show that the proposed approach generalizes well to previously unseen scenes and compares favorably to other recent CNN-based methods.", "future_impact": "This work could improve scalability of camera relocalization methods and their ability to extend to scenes not available during the training of the network. It would be interesting to see further applications of the proposed method in other types of visual localization tasks.", "venue": "ICCV", "year": 2017, "title": "Camera Relocalization by Computing Pairwise Relative Poses Using Convolutional Neural Network"} +{"pid": "94e54d29-9494-4d0a-9e0d-b3a055d1c856", "context": "Falls are a major risk for the elderly, especially those suffering from dementia wherein they are not able to react properly to such emergency situations. Current state-of-the-art fall detection approaches need to be improved.", "key_idea": "The authors suggest a vision-based approach for fall detection which incorporates 3D data obtained from a Kinect device and fuzzy logic to improve the robustness and effectiveness of detection.", "method": "The proposed approach is evaluated using 72 video sequences which consist of 40 falls and 32 activities of daily living.", "outcome": "The proposed method outperforms current state-of-the-art algorithms in terms of fall detection.", "future_impact": "The study contributes to improving elderly care by providing a more robust and effective automated fall detection system.", "venue": "ICCV", "year": 2012, "title": "Robust fall detection by combining 3d data and fuzzy logic"} +{"pid": "33384c05-ff5a-4a0d-be3c-f5cbdf0fc186", "context": "Conventionally, the motion of a driving car is highly constrained, and no efficient predictors existed that could 'learn' typical egomotion statistics to support tasks such as feature matching, tracking, and egomotion estimation.", "key_idea": "The authors propose a powerful predictor that learns the typical egomotion statistics based on the 'ground truth' data from KITTI odometry benchmark sequences and suggests that a coordinated turn motion model with moderate vibrations is realistic.", "method": "The authors analyze the statistics of the 'ground truth' data in the KITTI odometry benchmark sequences and develop a predictor that reduces uncertainties about relative motion when a new image frame comes in, effectively steering the matching process.", "outcome": "The proposed predictor significantly reduces uncertainty about the relative motion when a new image frame comes in and can also be used to detect outliers in the temporal sequence of egomotion parameters.", "future_impact": "Predictors like the one proposed in the study can be used to steer the process of matching from one frame to the next, bringing about a potential revolution in feature matching, tracking, and egomotion estimation.", "venue": "ICCV", "year": 2015, "title": "The Statistics of Driving Sequences -- And What We Can Learn from Them"} +{"pid": "a85b8f8a-7521-4a65-9ab7-ce1ab21cf0a5", "context": "Previous work on Structure from Motion (SFM) has investigated parameter uncertainties by imposing gauge constraints on the solution before considering covariance estimation.", "key_idea": "This paper explores how covariance-based uncertainty can be represented under inherent indeterminacies in SFM, termed gauge freedoms, and examines the effect of selecting a particular gauge on parameter uncertainties.", "method": "The authors derive a Geometric Equivalence Relationship to enable comparisons of covariances under different parametrizations and gauges based on their geometric uncertainty, and propose a fast method for covariance estimation.", "outcome": "They demonstrate that gauge choice can dramatically affect parameter uncertainties, but the inherent geometric uncertainty remains the same. They show that the uncertainty of gauge invariants accurately captures the geometric uncertainty of the solution.", "future_impact": "The findings and methods proposed offer useful measures for evaluating solution uncertainty in SFM, which may guide future research and applications in this field.", "venue": "ICCV", "year": 1999, "title": "Uncertainty Modeling for Optimal Structure from Motion"} +{"pid": "bb9e685e-c8aa-49e2-835c-890b45b0841f", "context": "Most previous visual recognition systems assume ideal inputs without real-world degradations, such as low resolution, motion blur, and out-of-focus blur. Conventional approach first resorts to blind image restoration and then feeds the restored image into a classifier, leading to defective output due to the ill-posed nature of blind image restoration.", "key_idea": "This paper presents a joint blind image restoration and recognition method based on the sparse representation prior. It handles the challenging problem of face recognition from low-quality images, where the degradation model is unknown. The method states that the degraded input image, if correctly restored, will have a sparse representation in terms of the training set, indicating the identity of the test image.", "method": "The proposed algorithm is evaluated by solving the blind image restoration iteratively in pursuit of the sparsest representation for recognition. Extensive experiments are carried out on face datasets under various degradations.", "outcome": "The joint model demonstrated significant improvements over conventional methods that treat the two tasks - image restoration and recognition, independently.", "future_impact": "The study indicates that the image restoration task and the recognition task can benefit greatly from each other, which might lead to the development of enhanced visual recognition systems exploiting this interdependence.", "venue": "ICCV", "year": 2011, "title": "Close the loop: Joint blind image restoration and recognition with sparse representation prior"} +{"pid": "25c41f22-43a7-4020-8e2e-df2864e371ce", "context": "The effects of errors in estimates of the image center and focal length on pose refinement and other related three-dimensional (3D) inference from two-dimensional (2D) images problems have not been thoroughly explored.", "key_idea": "The authors examine the effect of incorrect knowledge of the camera center and incorrect estimates of the focal length on the determination of the location of the camera and other 3D inference tasks, and consider the impact of these errors on the pose refinement algorithm.", "method": "They use experiments with synthetic and real data to evaluate the effect of errors in the estimates of the image center and focal length on pose refinement.", "outcome": "The study finds that incorrect camera center knowledge does not significantly impact camera location determination for 'small' field of view systems. However, incorrect focal length estimates considerably affect the determination of the z-component of translation in camera coordinates. New points are located to an average accuracy of 1.5 mm and 0.3 ft for the two real image sequences respectively.", "future_impact": "This study improves understanding of the sensitivity of pose refinement to camera parameter estimation errors, which may motivate future research on improving these estimations or developing methods robust to such errors.", "venue": "ICCV", "year": 1990, "title": "Sensitivity of the pose refinement problem to accurate estimation of camera parameters"} +{"pid": "605b1ff691e0119ebe7e5dba", "context": "The light transport matrix (LTM) is an instrumental tool in line-of-sight (LOS) imaging. It describes how light interacts with the scene and enables applications such as relighting or separation of illumination components.", "key_idea": "The authors introduce a framework for estimating the LTM in non-line-of-sight (NLOS) scenarios. The authors couple recent virtual forward light propagation models with the LOS light transport equation and introduce specific illumination functions to overcome the challenging wide-aperture conditions of NLOS setups.", "method": "The authors design computational projector-camera setups, and use these virtual imaging systems to estimate the transport matrix of hidden scenes.", "outcome": "The introduced NLOS light transport matrix allows to (re)illuminate specific locations of a hidden scene, and separate direct, first-order indirect, and higher-order indirect illumination of complex cluttered hidden scenes.", "future_impact": "The proposed technique could extend the applications of NLOS scenarios by enabling the separation of different types of illuminations and enabling relighting of specific scene locations in ways similar to existing line-of-sight techniques.", "venue": "ICCV", "year": 2021, "title": "Virtual light transport matrices for non-line-of-sight imaging."} +{"pid": "8908bea1-b634-41e9-a5da-a2daea0e7482", "context": "High-quality computer vision models typically address the problem of understanding the general distribution of real-world images. However, most cameras observe only a very small fraction of this distribution, creating a possibility for more efficient inference by specializing compact, low-cost models to specific frames observed by a single camera.", "key_idea": "The paper proposes to use model distillation to specialize accurate, low-cost semantic segmentation models to a target video stream, training the \u2018student\u2019 model in an online fashion on the live video.", "method": "The approach intermittently runs a high-cost 'teacher' model to provide a learning target during student model training. The model performance was compared with solutions based on flow or video object segmentation, and temporal stability with the original teacher model. A new video dataset was produced for evaluating the efficiency of inference over long-running video streams.", "outcome": "Online model distillation yields semantic segmentation models that closely approximate their Mask R-CNN teacher with 7 to 17 times lower inference runtime cost or 11 to 26 times lower in FLOPs, even when the target video's distribution is non-stationary.", "future_impact": "The method could improve efficiency in video segmentation tasks because it does not require offline pre-training on the target video stream, and can potentially offer better temporal stability than the original teacher model.", "venue": "ICCV", "year": 2019, "title": "Online Model Distillation for Efficient Video Inference"} +{"pid": "fdaa549c-812a-4c6d-a7ce-ebe61169967a", "context": "Visual categorization problems such as object classification or action recognition traditionally train subwindow classifiers on a large collection of manually annotated examples with masks or bounding boxes. This approach relies on time-consuming human labeling and is limited to very few categories. Also, the human selection of masks introduces potentially suboptimal biases in terms of window size and location.", "key_idea": "A novel method is proposed for learning a discriminative subwindow classifier from examples annotated with binary labels indicating the presence of an object or action of interest, but not its location. During training, the method simultaneously localizes the instances of the positive class and learns a subwindow SVM to recognize them.", "method": "The authors present an algorithm that extends this method to the classification of time series by localizing the most discriminative set of temporal segments in the signal. The method is evaluated on several datasets for object and action recognition.", "outcome": "The proposed method achieves results similar to, and in many cases superior to, those obtained with full supervision on several datasets of object and action recognition.", "future_impact": "The proposed method presents a new direction in machine learning, with potential implications in visual categorization problems. It could lead to decrease dependence on extensive manual labelling and help to address a wider range of categories.", "venue": "ICCV", "year": 2009, "title": "Weakly supervised discriminative localization and classification: a joint learning process"} +{"pid": "bbb37dfa-c5eb-4437-83ae-b50d9eea4b66", "context": "Human action can be recognised from a single still image by modelling Human-object interaction (HOI), which infers the mutual spatial structure information between human and object as well as their appearance. Existing approaches are sensitive to large variations of human poses, occlusion, and unsatisfactory detection of small objects, due to the reliance on accurate detection of human and object, and estimation of human pose.", "key_idea": "A novel exemplar based approach is proposed to overcome the limitation of existing approaches. This approach learns a set of spatial pose-object interaction exemplars, which are density functions that describe how a person is interacting with a manipulated object for different activities spatially in a probabilistic way.", "method": "The authors create a new framework which includes its proposed exemplar based HOI descriptor and an activity-specific matching model that learns the parameters. They validate this by carrying out experiments on two benchmark activity datasets.", "outcome": "The proposed exemplar based approach was able to obtain state-of-the-art performance in experiments conducted on two benchmark activity datasets.", "future_impact": "The new approach to human activity recognition based on Human-object interaction exemplars has the potential to be robust to errors in human/object detection and pose estimation.", "venue": "ICCV", "year": 2013, "title": "Recognising Human-Object Interaction via Exemplar Based Modelling"} +{"pid": "a3268a04-725a-42f4-899e-fe12ae85b245", "context": "Previous methods for autonomous mobile robot navigation involved using images captured by a monocular omnidirectional camera system. These methods, however, have not fully utilized the geometric properties of the images from the camera system.", "key_idea": "The paper introduces a novel algorithm for circle-marker detection in spherical images captured by an omnidirectional camera system, and leverages the geometrical properties of the images, utilizing them for the positioning and navigation of autonomous mobile robots.", "method": "The authors develop a voting method for the extraction of images of planar circle markers on a spherical image, using a spatial-quadric detection strategy. They also derive a positioning algorithm based on the assumption of the geometrical configuration of the camera system and the circle markers on the plane on which the robot moves.", "outcome": "The paper includes numerical examples, both for synthetic and real images, demonstrating the application of the proposed circle-marker detection method and positioning algorithm.", "future_impact": "The proposed method enabling robot navigation using the circle-detection algorithm, when mounted on a monocular omnidirectional camera system, could facilitate more efficient autonomous mobile robot operations in the future.", "venue": "ICCV", "year": 2007, "title": "Circle-Marker Detection Method for Omnidirectional Images and its Application to Robot Positioning"} +{"pid": "6099028591e011aa8bcb6e2c", "context": "In video recognition, the most informative region in each frame is usually a small image patch, which shifts smoothly across frames. However, recognising this informative patch efficiently is a challenging problem.", "key_idea": "The authors propose a reinforcement learning based approach, AdaFocus, for efficient spatially adaptive video recognition. The model uses a light-weighted ConvNet to process the full video sequence and a recurrent policy network to localize the task-relevant regions.", "method": "Experiments are conducted on five benchmark datasets, i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, to evaluate the performance and efficiency of the proposed model.", "outcome": "The proposed model proves to be significantly more efficient than the competitive baselines.", "future_impact": "AdaFocus' ability to improve computational efficiency opens up the possibility of its application in real-time video recognition tasks. Additionally, its extendability to consider temporal redundancy may inspire future research on dynamically skipping less valuable frames for video recognition.", "venue": "ICCV", "year": 2021, "title": "Adaptive Focus for Efficient Video Recognition."} +{"pid": "05d4568a-0a01-4f14-b115-258e7835216e", "context": "Autism Spectrum Disorder (ASD) is prevalent, affecting 1.5% of people in the US. The lack of clinical resources for early diagnoses has been a long-lasting issue.", "key_idea": "The research introduces a novel, data-driven method for diagnosing ASD. This approach combines eye tracking and deep neural networks to analyse differences in eye movement patterns between healthy individuals and those with ASD.", "method": "The authors apply an image selection method based on Fisher scores for feature learning with the most discriminative contents. They leverage the recent advances in deep neural networks for both prediction and visualization.", "outcome": "Experimental results show the superior performance of the proposed method in terms of multiple evaluation metrics used in diagnostic tests.", "future_impact": "The data-driven aspect of this method, which is free of assumptions, could be important for new discoveries in understanding ASD as well as other neurodevelopmental disorders.", "venue": "ICCV", "year": 2017, "title": "Learning Visual Attention to Identify People with Autism Spectrum Disorder"} +{"pid": "df7c4ef9-30c0-4507-aa36-72b60bfa50fc", "context": "Deep CNNs have made great progress in image semantic segmentation, but they require numerous densely-annotated images for training and struggle to generalize to unseen object categories. Few-shot segmentation has been developed to perform segmentation from only a few annotated examples.", "key_idea": "The authors propose a novel prototype alignment network called PANet, which learns class-specific prototype representations from a few support images within an embedding space and then performs segmentation over the query images by matching each pixel to the learned prototypes.", "method": "The authors approach the few-shot segmentation problem from a metric learning perspective. Their PANet employs non-parametric metric learning to generate high-quality prototypes that are both class representative and discriminative for different classes. They further introduce a prototype alignment regularization between the support and the query in PANet.", "outcome": "The proposed PANet achieves the mIoU score of 48.1% and 55.7% on PASCAL-5i for 1-shot and 5-shot settings respectively, surpassing the state-of-the-art method by 1.8% and 8.6%.", "future_impact": "The authors suggest that PANet fully harnesses knowledge from the support and provides better generalization on few-shot segmentation, leading to potential improvements in this field.", "venue": "ICCV", "year": 2019, "title": "PANet: Few-Shot Image Semantic Segmentation With Prototype Alignment"} +{"pid": "64c88ca43fda6d7f06268983", "context": "Deep Neural Networks for 3D point cloud recognition are vulnerable to adversarial examples, making them challenging to defend against due to the diverse ways attackers can generate these adversarial examples such as by adding, shifting, or removing points.", "key_idea": "The authors propose to establish a comprehensive point cloud adversarial robustness benchmark and propose a hybrid training augmentation method that considers various types of point cloud adversarial examples to adversarial training.", "method": "The authors conduct extensive and systematic experiments to evaluate adversarial robustness and to identify effective combinations of existing defense tricks in point cloud adversarial defenses.", "outcome": "By combining these tricks, a more robust defense framework is constructed with an average accuracy of 83.45% against various attacks, demonstrating its capability to enable robust learners.", "future_impact": "This work establishes a benchmark and proposes a method that could potentially improve robustness against adversarial examples in 3D point cloud recognition.", "venue": "ICCV", "year": 2023, "title": "Benchmarking and Analyzing Robust Point Cloud Recognition: Bag of Tricks for Defending Adversarial Examples"} +{"pid": "619655626750f87a92a93c75", "context": "Low-cost monocular 3D object detection is crucial in autonomous driving, and yet the accuracy of current solutions is still not satisfactory.", "key_idea": "The paper proposes the Dynamic Feature Reflecting Network (DFR-Net) for 3D object detection by reformulating the task into sub-tasks of object localization and appearance perception, and leveraging the reciprocal information between these sub-tasks.", "method": "The DFR-Net comprises two novel modules: (i) the Appearance-Localization Feature Reflecting module (ALFR) that separates task-specific features and reflects the reciprocal features; (ii) the Dynamic Intra-Trading module (DIT) that adaptively adjusts the training processes of various sub-tasks. This method was tested on the KITTI dataset.", "outcome": "Extensive experiments showed the effectiveness and generalization of DFR-Net, ranking first among all monocular 3D object detectors in the KITTI dataset as of March 16th, 2021.", "future_impact": "The proposed method is easily integrated into many cutting-edge 3D detection frameworks with little cost to boost performance, therefore it could contribute to advancements in the field of 3D object detection.", "venue": "ICCV", "year": 2021, "title": "The Devil is in the Task - Exploiting Reciprocal Appearance-Localization Features for Monocular 3D Object Detection."} +{"pid": "6528c1fd939a5f4082997e0f", "context": "Anomaly detection in videos is a critical research area with real-world applications. Current video anomaly detection (VAD) methods are constrained due to the lack of large-scale annotated anomaly events and typically focus on learning the distribution of normal samples to spot anomalies. They often rely on auxiliary networks to extract foreground object or action information, yet performance of these methods is significantly affected by the capability of the semantic models used.", "key_idea": "The authors propose a novel method based on the Diffusion Model (DM) to predict the features of video frames for anomaly detection. The method is designed to learn the distribution of normal samples without the need for any additional high-level semantic feature extraction models.", "method": "The authors develop two denoising diffusion implicit modules to predict and refine the features: the first module is dedicated to learning feature motion, whilst the latter focuses on feature appearance learning. The performance of the proposed method is then assessed by conducting extensive experiments.", "outcome": "The proposed method shows significantly improved performance in accurately predicting normal features over non-DM-based feature prediction-based VAD methods and is demonstrated to substantially outperform state-of-the-art competing methods in experiments.", "future_impact": "The proposed approach, being the first of its kind to use DM for feature prediction in VAD, could drive new advancements in the field of video anomaly detection. More effective and accurate anomaly detection in videos could have wide-reaching implications in real-world applications.", "venue": "ICCV", "year": 2023, "title": "Feature Prediction Diffusion Model for Video Anomaly Detection."} +{"pid": "6528c224939a5f408299ec67", "context": "Human pose estimation (HPE) is integral to scene understanding in numerous safety-critical domains involving human-machine interaction. Many models in this domain often fall prey to overconfidence issues common in deep probabilistic models, presenting the problem of inaccurate prediction.", "key_idea": "The paper proposes a human pose regression framework to examine the behavior of two established methods for simultaneous aleatoric and epistemic uncertainty estimation: maximum a-posteriori (MAP) estimation with Monte-Carlo variational inference and deep evidential regression (DER). They also introduce additional recalibration step to extract reliable confidence intervals.", "method": "They evaluate both approaches on the quality of their predicted variances and whether these truly capture the expected model error. They specifically take a closer look at deep evidential regression, which is applied comprehensively for the first time to the HPE problem.", "outcome": "The experimental results indicate that DER behaves as expected in challenging and adverse conditions, and that the predicted uncertainties match their purported aleatoric and epistemic sources. Notably, DER achieves smooth uncertainty estimates without the need for a costly sampling step.", "future_impact": "The findings show that DER could be a suitable candidate for uncertainty estimation in resource-limited platforms, thus benefiting the domain of human pose estimation.", "venue": "ICCV", "year": 2023, "title": "Plausible Uncertainties for Human Pose Regression."} +{"pid": "9dd6cda0-7cdd-4ade-80e7-ecfcc8d5eb5e", "context": "Updating a global 3D model with live RGB-D measurements, such as through a Truncated Signed Distance Function (TSDF) volumetric model and a fusion algorithm, has been successful for 3D reconstruction of indoor scenes. However, this method is expensive in memory, limiting its scalability for large scenes.", "key_idea": "This paper proposes a new flexible 3D scene representation which uses a set of planes, aiming to decrease memory usage while maintaining accurate reconstruction of indoor scenes from RGB-D image sequences.", "method": "The authors project the scene onto different planes to reduce the size of the scene representation, and this updated model was tested with sets of RGB-D measurements.", "outcome": "Experimental results show that the proposed flexible 3D scene representation method achieves accurate reconstruction, while supporting scalability for large indoor scenes.", "future_impact": "The proposed method's lower memory requirement could allow global textured 3D models to be updated more easily with live RGB-D measurements, possibly impacting the field with its increased capacity for large-scale indoor 3D reconstructions.", "venue": "ICCV", "year": 2013, "title": "A Flexible Scene Representation for 3D Reconstruction Using an RGB-D Camera"} +{"pid": "72453fa5-b181-4cae-989e-4b7e828d586e", "context": "Current deep learning architectures, like Deep CNNs, rely heavily on learnable operators and layers, yet, they may not always provide the best results nor the most efficient use of resources.", "key_idea": "This study proposes the use of scattering networks as the first layers of a supervised hybrid deep network, with the idea that the early layers don't necessarily need to be learned and can leverage predefined representations.", "method": "To validate this key idea, the authors experimented with a shallow cascade of 1 \u00d7 1 convolutions, and combining scattering networks with a modern ResNet. These networks were tested on large-scale image recognition tasks like the ImageNet ILSVRC2012 dataset, and smaller datasets like subsets of the CIFAR-10 and the STL-10 dataset.", "outcome": "The hybrid scattering deep networks achieved AlexNet accuracy on ImageNet ILSVRC2012, and a single-crop top 5 error of 11.4% on the same data set, comparable to the ResNet-18 architecture. The hybrid models also demonstrated excellent performance in small sample regimes, surpassing their end-to-end counterparts demonstrated on subsets of the CIFAR-10 dataset and on the STL-10 dataset.", "future_impact": "This work implies that learning invariance and geometrical priors in the architecture of machine learning models is beneficial and that incorporating scattering networks as predefined first layers in deep learning models could be a promising direction for future performance improvements and resource efficiency.", "venue": "ICCV", "year": 2017, "title": "Scaling the Scattering Transform: Deep Hybrid Networks"} +{"pid": "21490c18-d30d-45f8-83fe-40e7895d2554", "context": "The generation of event models typically requires tedious manual work based on the analysis of visual input such as video frames. Current systems apply traditional machine learning techniques to attain a pseudo-automatic approach.", "key_idea": "The researchers propose a technique that fully automates the generation of event models using qualitative reasoning and statistical analysis of video input. The technique leverages a tracking program, semantically partitions the view from a fixed camera based on moving objects' paths, and builds event models describing the behaviour of object pairs.", "method": "The view from a fixed camera is partitioned into semantically relevant regions based on the paths of moving objects. Using a notion of proximity and qualitative spatial reasoning techniques, event models are built using statistical methods. The system has been tested in a traffic domain.", "outcome": "The automated system was successful in learning various event models, expressed in the qualitative calculus, that represent human observable events. Furthermore, the system is capable of recognizing subsequent selected event occurrences or unusual behaviours.", "future_impact": "The proposed technique for automatic event model generation based on qualitative reasoning and statistical analysis of video input could be used to automatically recognize subsequent selected event occurrences or unusual behaviours in a variety of domains.", "venue": "ICCV", "year": 1998, "title": "Building qualitative event models automatically from visual input"} +{"pid": "51b1def9-d01b-444d-86e7-dc2c98a6d55c", "context": "Vergence methods in computer vision could not overcome problems created by occlusion and specular highlights, issues that are inherent in the visual process.", "key_idea": "The authors describe an active exploratory vergence method based on concurrent cross-correlation of multi-scale stereo images, aiming to concurrently detect occlusion and specular highlights during the vergence process.", "method": "An efficient parallel implementation of the multi-scale cross-correlation algorithm is used in the proposed active exploratory vergence method.", "outcome": "The proposed method demonstrated its robustness against occlusion and specular highlights.", "future_impact": "The proposed method's ability to efficiently detect occlusion and specular highlights concurrently during the vergence process could lead to the exploration of a better viewing direction and position in computer vision.", "venue": "ICCV", "year": 1993, "title": "Robust vergence with concurrent detection of occlusion and specular highlights"} +{"pid": "342dcd13-48c1-4760-9ced-e2e98c51533a", "context": "The relationship between the appearance of an outdoor scene and the ambient temperature has not been extensively explored until now.", "key_idea": "The authors identify two interactions between image sequences from outdoor cameras and temperature measurements, namely, semantically meaningful regions such as foliage and reflective oriented surfaces and small camera motions, both being indicative of the temperature.", "method": "The authors study statistical correlations between image sequences from outdoor cameras and temperature measurements. They propose simple scene-specific temperature prediction algorithms which turn a camera into a crude temperature sensor.", "outcome": "It was found that simple features like local pixel intensities outperform sophisticated, global features from semantically-trained convolutional neural networks for temperature estimation.", "future_impact": "The findings and proposed simple scene-specific temperature prediction algorithms can make cameras function as simple temperature sensors, which can have distinct applications in various fields.", "venue": "ICCV", "year": 2015, "title": "Hot or Not: Exploring Correlations between Appearance and Temperature"} +{"pid": "181c0b62-21c7-4fe1-9ec1-92e5dd944890", "context": "Existing architectures for semantic segmentation rely on a single encoder-decoder structure, but can be computationally burdensome and are unable to balance speed and high accuracy efficiently.", "key_idea": "The authors propose ShelfNet, an architecture with multiple encoder-decoder branch pairs and skip connections at each spatial level. This architecture appears as a shelf with multiple columns and serves as an ensemble of multiple deep and shallow paths, aiming to improve accuracy while reducing computational burden.", "method": "The authors compared ShelfNet with popular non real-time methods like PSPNet and real-time segmentation models such as BiSeNet, on the PASCAL VOC and Cityscapes datasets.", "outcome": "ShelfNet demonstrated 4x faster inference speed than PSPNet with similar accuracy on the PASCAL VOC dataset. On the Cityscapes Dataset, the proposed model achieves higher accuracy at similar speed as BiSeNet. Lastly, ShelfNet with ResNet34 backbone achieves 79.0% mIoU on Cityscapes Dataset, outperforming both PSPNet and BiSeNet with larger backbones like ResNet101.", "future_impact": "With its balance of speed and accuracy, ShelfNet has potential applications in tasks which require speed, such as street-scene understanding for autonomous driving.", "venue": "ICCV", "year": 2019, "title": "ShelfNet for Fast Semantic Segmentation"} +{"pid": "35ad8f20-4e0b-47f6-9f4f-2c49a54dcc38", "context": "Current approaches to the detection and 3D pose estimation of objects in color images rely heavily on extensive training phases and require large datasets, often needing hours or training time and hundreds of registered images for new objects.", "key_idea": "The authors propose a novel methodology for object detection and 3D pose estimation that eliminates the need for training data for new objects. This method focuses on objects with prominent corners and relies solely on their geometries. It uses object corners of various shapes and predicts their 3D poses using a small set of training images from existing objects.", "method": "The authors utilize a RANSAC-like algorithm that robustly detects and estimates the object's 3D pose by matching its corners on the CAD model with their detected counterparts in the image. A final check exploiting the full 3D geometry of the objects is also introduced to handle cases where multiple objects have the same corner spatial arrangement.", "outcome": "The proposed methodology is demonstrated to be effective on the challenging T-LESS dataset.", "future_impact": "The paper's methodology could be particularly valuable in industrial contexts where there is a constant introduction of new objects and retraining models frequently is not feasible.", "venue": "ICCV", "year": 2019, "title": "CorNet: Generic 3D Corners for 6D Pose Estimation of New Objects without Retraining"} +{"pid": "4c22735f-bfe5-4962-85e7-abfdddf1d4f5", "context": "Quantizing deep neural network models for efficient inference on modern deep learning hardware typically leads to either significant performance reduction or increased engineering time spent on training a network to be suitable for quantization.", "key_idea": "The authors propose a data-free quantization method for deep neural networks that does not require fine-tuning or hyperparameter selection. This is obtained by equalizing the weight ranges in the network using a scale-equivariance property of activation functions and correcting biases in the error introduced during quantization.", "method": "The authors apply their proposed weight equalization and bias correction method to common computer vision architectures and tasks such as the MobileNet family, semantic segmentation and object detection,", "outcome": "The proposed method achieves near-original model performance and for common architectures like the MobileNet family, it even achieves state-of-the-art quantized model performance.", "future_impact": "The introduced method, which improves quantization accuracy performance and can be implemented using a simple API call, can be applied to a wide variety of computer vision architectures and tasks enhancing their efficiency and performance without requiring fine-tuning or hyperparameter selection.", "venue": "ICCV", "year": 2019, "title": "Data-Free Quantization Through Weight Equalization and Bias Correction"} +{"pid": "a3e32822-1854-4333-8d73-5029ae57894d", "context": "Computing features of closed planar boundaries of 2-D digital images, or closed curves, is a task that has been previously tackled in computational vision research, but still poses challenges.", "key_idea": "The authors propose a novel method called distance accumulation that computes the distance from a point in the boundary to a chord specified by moving end points.", "method": "The robustness of the distance accumulation method was tested through experimental results using both simulated and real images.", "outcome": "The experimental results showed that the distance accumulation method was robust and its analysis was closely matched to planar curvature measurements.", "future_impact": "The successful implementation and results of the distance accumulation method may lead to further research and improvements in feature extraction from 2-D digital images.", "venue": "ICCV", "year": 1993, "title": "Distance accumulation and planar curvature"} +{"pid": "5b63d4c7-affd-4825-9067-2fc0aaae359e", "context": "Traditional methods struggle to efficiently track point features experiencing affine deformations and changes in illumination.", "key_idea": "The authors develop an efficient algorithm for tracking point features that are affected by both affine deformations and changes in illumination, utilizing a combined model of geometry and photometry.", "method": "They use the combined model of geometry and photometry to track features and detect outliers within a hypothesis testing framework.", "outcome": "The proposed algorithm can track features in real-time on a personal computer.", "future_impact": "The authors released their algorithm to the public, implying it could be incorporated into other systems that require real-time feature tracking.", "venue": "ICCV", "year": 2001, "title": "Real-time feature tracking and outlier rejection with changes in illumination"} +{"pid": "20cfd06e-15ce-4560-a9d3-bad7b310787a", "context": "Existing methods for image content authentication have issues with false positive and false negative probabilities.", "key_idea": "The authors introduce a hierarchical representation of image structure and utilize it for image content authentication. They achieve this through a Bayesian framework, modelling the image with the Markov pixon random field and defining an undirected graph - Bayesian structural content abstraction (BaSCA).", "method": "An efficient optimization scheme is implemented to iteratively minimize the distance (or learning error) to all content-identical image samples generated by an acceptable operation set defined by the user. The authors also use the regional pixon map to remove spurious vertices and establish a BaSCA hierarchy.", "outcome": "Experimental results indicate that the proposed approach has significantly less false positive and a comparable false negative probability compared with existing methods.", "future_impact": "The proposed representation captures the spatial topology information of homogeneous regions as well as their finest scale and interactions. The development of such a hierarchy can help in improving the robustness of image content authentication.", "venue": "ICCV", "year": 2005, "title": "Bayesian structural content abstraction for region-level image authentication"} +{"pid": "f3cf1d2d-df1b-4d39-87ca-104b854a4ab0", "context": "Extending state-of-the-art object detectors from image to video is challenging due to issues such as motion blur, video defocus, and rare poses. Existing work attempts to exploit temporal information on the box level, but these methods are not trained end-to-end.", "key_idea": "This paper presents flow-guided feature aggregation, an accurate and end-to-end learning framework for video object detection that improves the video recognition accuracy by leveraging temporal coherence on the feature level and improving per-frame features by aggregation of nearby features along the motion paths.", "method": "The authors tested their method on the ImageNet VID dataset, especially on more challenging fast moving objects.", "outcome": "The flow-guided feature aggregation method significantly improves upon strong single-frame baselines in ImageNet VID, and is on par with the best engineered systems that won the ImageNet VID challenges 2016.", "future_impact": "The authors plan to release their code, which could enable other researchers to build upon this work and further improve video object detection.", "venue": "ICCV", "year": 2017, "title": "Flow-Guided Feature Aggregation for Video Object Detection"} +{"pid": "5e8fb5db-1a00-4bda-9c79-6df889fcfa02", "context": "Traditional Markov random field (MRF) models have limitations in defining field potentials according to local image structure.", "key_idea": "The study proposes a Steerable Random Field (SRF) where the field potentials are defined based on filter responses that are steered to local image structure, making use of the structure tensor to obtain derivative responses aligned with or orthogonal to the predominant local image structure.", "method": "Clique potentials are defined over steered filter responses using a Gaussian scale mixture model and are learned from training data. Image denoising and inpainting performance is evaluated to test the effectiveness of the SRF model.", "outcome": "The study demonstrates that steering the random field to the local image structure improves image denoising and inpainting performance compared with traditional pairwise MRFs.", "future_impact": "The SRF model provides a statistical motivation for anisotropic regularization and may further advance MRF models and methods in image processing.", "venue": "ICCV", "year": 2007, "title": "Steerable Random Fields"} +{"pid": "02f237ba-fa4d-43c8-b639-3d276ce20488", "context": "Subspace clustering of data points from a union of subspaces is a popular research area. Clustering based on the assumption of a nonlinear subspace is more difficult for traditional methods. Autoencoders, due to their capacity to extract nonlinear features, are often used, however, the problem of adding constraints to these features for better effectiveness has not been thoroughly addressed in previous work.", "key_idea": "The paper proposes a method of adding both global and local structure constraints to features in autoencoders using the low-rank property and Laplace operator, respectively. This method improves the effectiveness of feature extraction.", "method": "The authors conduct extensive experiments to validate the effectiveness of the proposed model.", "outcome": "Experiments demonstrate the effectiveness of the proposed method, although specific results are not stated in the abstract.", "future_impact": "The approach of placing constraints as per the proposed way could potentially be employed in other deep neural network architectures.", "venue": "ICCV", "year": 2019, "title": "Structure-Constrained Feature Extraction by Autoencoders for Subspace Clustering"} +{"pid": "3f14be83-6e11-4541-9f9a-f93aedcae92b", "context": "Existing approaches to fusing silhouette information from multiple arbitrary views often involve using 3D constructs such as camera calibration to shape 3D voxels or to project visual cones in 3D space.", "key_idea": "The paper introduces a purely image-based method that uses planar homographies and foreground likelihood information to perform the fusion of multi-view silhouettes, thus avoiding complex 3D constructs.", "method": "The team uses planar homographies and foreground likelihoods to form a 2D grid which represents a cross-sectional slice of the object. They then extend this process to planes parallel to a reference plane to obtain the object's other slices. These grids are then stacked on top of each other to create a 3D data structure which encapsulates the object's shape and location, and an energy functional is minimized over the object's surface using a level sets formulation.", "outcome": "The proposed method successfully generates three-dimensional structures that encapsulate complex object shapes and their locations in cluttered environments containing multiple objects.", "future_impact": "The image-based approach to fusing multi-view silhouette information can simplify the 3D reconstruction process in complicated environments, making it applicable in fields that require efficient and precise object reconstruction.", "venue": "ICCV", "year": 2007, "title": "A Homographic Framework for the Fusion of Multi-view Silhouettes"} +{"pid": "38a5cabe-47fa-483e-88bc-e1aeb3642004", "context": "Traditional neural networks only try to find a model with fixed parameters that optimize the average behavior over all inputs, without considering data-specific properties.", "key_idea": "A meta-level neural network (MLNN) model is proposed. This model uses meta-knowledge learned on data-specific properties of images and dynamically adapts its weights based on the properties of the input images.", "method": "The MLNN model is validated using over ten network architectures under various application scenarios and loss functions. The authors conduct comparison experiments on low-level and high-level vision applications, specifically image super-resolution, demising, and image classification tasks.", "outcome": "The MLNN model has demonstrated 0.1-0.3 dB improvements on PSNR in low-level vision applications, while for high-level image classification, it has an accuracy improvement of 0.4-0.6% for Cifar10 and 1.2-2.1% for ImageNet, compared to convolutional NNs.", "future_impact": "The method is expected to lead to more pronounced improvements as the scale or diversity of data is increased.", "venue": "ICCV", "year": 2017, "title": "Supplementary Meta-Learning: Towards a Dynamic Model for Deep Neural Networks"} +{"pid": "8703cafe-83c8-4106-91b2-8fcdbb07719a", "context": "While generic face detection and facial landmark localization in static imagery are mature areas in machine learning and computer vision, facial landmark tracking in videos has not received the same attention. Tracking-by-detection frameworks that use face detection and landmark localization in every frame to avoid drifting can't outperform static imagery methods.", "key_idea": "The authors propose using the results of generic face detection and landmark localization methods to recursively train person-specific face detectors and landmark localizers for offline deformable face tracking.", "method": "The authors apply the proposed pipeline to track landmarks in challenging long-term sequences captured under arbitrary conditions. The pipeline was tested as a semi-automatic annotation tool on the majority of the videos of the 300-VW Challenge.", "outcome": "The results from the application of the proposed method on the 300-VW Challenge are not explicitly mentioned.", "future_impact": "The study introduces a novel method for facial landmark tracking in videos which could impact future developments in this area, though no specific impacts or potential research directions are identified.", "venue": "ICCV", "year": 2015, "title": "Offline Deformable Face Tracking in Arbitrary Videos"} +{"pid": "4b9dd617-70db-4bc2-9fff-f239579d0a24", "context": "State-of-the-art object detectors heavily rely on pre-trained networks on large-scale classification datasets like Image Net, which incurs learning bias due to the difference on both the loss functions and the category distributions between classification and detection tasks. Transferring pre-trained models from classification to detection between discrepant domains is also challenging.", "key_idea": "The authors propose the Deeply Supervised Object Detector (DSOD), a framework that can train object detectors from scratch, eliminating the need for pre-trained models. This is achieved through deep supervision enabled by dense layer-wise connections.", "method": "The authors develop DSOD following the single-shot detection (SSD) framework and conduct experiments on the PASCAL VOC 2007, 2012 and MS COCO datasets.", "outcome": "The DSOD outperforms SSD on all three benchmarks with real-time detection speed, requiring only half the parameters of SSD and one-tenth the parameters of Faster RCNN.", "future_impact": "The principles and findings in DSOD could potentially influence future work in object detection tasks, especially in training detectors from scratch and creating more compact models.", "venue": "ICCV", "year": 2017, "title": "DSOD: Learning Deeply Supervised Object Detectors from Scratch"} +{"pid": "6196560e6750f87a92a93ccc", "context": "Human parsing and pose estimation are central for understanding human behaviors. Given their close correlation, employing one model for both tasks can yield substantial benefits. However, harmonizing these tasks in joint learning is difficult as they require different processes - pixel-wise classification for human parsing and regression for pose estimation.", "key_idea": "Inspired by the Neural Architecture Search's (NAS) capability to automatically allocate efficient feature connections for specific tasks, the authors propose a network architecture, NPPNet, designed to handle both tasks simultaneously.", "method": "The authors propose to search for encoder-decoder architectures respectively for getting task-specific features and laying the foundation of feature interaction. Further, NAS units are added to multi-scale feature interaction and high-level feature fusion to ensure optimal communication between the tasks. They validate this approach by experimental results on benchmark datasets for parsing and pose estimation.", "outcome": "The experimental results demonstrate that the proposed model achieves state-of-the-art performance on both human parsing and pose estimation tasks.", "future_impact": "While not explicitly stated in the abstract, the proposal of this method combining NAS and task-specific features could potentially impact future efforts in jointly tackling related tasks in the field of human behaviour understanding.", "venue": "ICCV", "year": 2021, "title": "Neural Architecture Search for Joint Human Parsing and Pose Estimation."} +{"pid": "afa34e5a-cc19-42b3-ae99-5263f9c256d4", "context": "Existing methods for reconstructing the 3D surface of an object from depth discontinuities when viewed under orthographic projection don't adequately address gaps in the recovered surface.", "key_idea": "The authors present a new method for 3-D surface reconstruction of an object using depth discontinuity motion under orthographic projection during rigid rotation on a turntable, along with a novel shape completion scheme to fill in the gaps in the recovered surface.", "method": "The authors validate the theory by constructing a large-format orthographic multi-flash camera capable of recovering the depth discontinuities using controlled illumination. They also analyze the limitations of multi-flash depth edge detection using orthographic imaging with both point sources and directional illumination.", "outcome": "This paper identifies and presents a new method for dealing with the problem of reconstructing 3D surfaces and shows through verification the practical functionality of their shape completion scheme and large-format orthographic multi-flash camera.", "future_impact": "The authors anticipate future applications for the shape completion scheme and the specialized hardware introduced in this work.", "venue": "ICCV", "year": 2009, "title": "Shape from depth discontinuities under orthographic projection"} +{"pid": "36cdbdd3-6c33-4200-abb8-36231d830903", "context": "Active Appearance Models (AAMs) are widely used, although they often involve complex optimization problems and can be computationally prohibitive, limiting their utility.", "key_idea": "A new framework is proposed to derive well-known optimization problems in AAMs, resulting in two optimization problems for fast and exact AAM fitting, and a new algorithm which can be applied to 3D.", "method": "After proposing the new optimization problems and algorithm, the authors trained AAMs in-the-wild to investigate its performance relative to other methods, without use of sophisticated shape priors, robust features or robust norms.", "outcome": "The results show that AAMs perform well in comparison to current state-of-the-art methods in some cases. The proposed algorithms are computationally realizable and particularly fast for most current systems.", "future_impact": "The proposed algorithm and optimization problems for AAM could become more broadly applied, especially if further improvements are made that can replicate the successes shown in this study.", "venue": "ICCV", "year": 2013, "title": "Optimization Problems for Fast AAM Fitting in-the-Wild"} +{"pid": "6516338d3fda6d7f065e4fbf", "context": "The existing human mesh recovery problem does not deal with the challenges of generating 3D body meshes from partial observation in 2D clothing images.", "key_idea": "This study defines and tackles the Cloth2Body problem aiming to generate 3D human body meshes from 2D clothing images. The authors propose an end-to-end framework that can estimate 3D body mesh parameterized by pose and shape.", "method": "Three specific solutions are provided within the framework to address the challenges in the Cloth2Body problem: 1) Kinematics-aware Pose Estimation for estimating body pose parameters; 2) An Adaptive Depth Trick to align the re-projected 3D mesh with the 2D clothing image; 3) Physics-informed Shape Estimation for estimating body shape parameters. A skeleton transplanting method, Evolution-based pose generation, is designed for diverse pose generation during inference.", "outcome": "Experimental results on both synthetic and real-world data show that the proposed framework can effectively recover natural and diverse 3D body meshes from 2D images that align well with clothing, achieving state-of-the-art performance.", "future_impact": "The study addresses an emerging challenge in graphics and computer vision, creating a pathway for future research on 3D mesh generation from 2D images.", "venue": "ICCV", "year": 2023, "title": "Cloth2Body: Generating 3D Human Body Mesh from 2D Clothing"} +{"pid": "e0ace8e5-b574-46d0-ab6c-615541a00a7e", "context": "Prior methods for rigid point set registration often rely on solving systems of second-order ordinary differential equations, which can be computationally expensive and sensitive to parameter choices.", "key_idea": "The authors propose a new registration method, called Barnes-Hut Rigid Gravitational Approach (BH-RGA), that interprets the input as two interacting particle swarms and directly minimizes the gravitational potential energy of the system using non-linear least squares.", "method": "The approach accelerates exhaustive particle interactions using a Barnes-Hut tree and handles massive point sets in quasilinear time. The performance of the method is evaluated through systematic experiments on handling incomplete, noisy and perturbed data, and alignment with prior matches.", "outcome": "The experiments demonstrate that BH-RGA surpasses the performance of baseline methods in terms of the convergence basin and accuracy when handling noisy and incomplete data, and that it compares positively to competing methods for alignment with prior matches.", "future_impact": "One advantage of BH-RGA is the possibility to define boundary conditions or additional alignment cues by varying point masses, which could allow for more flexible uses in future applications.", "venue": "ICCV", "year": 2019, "title": "Accelerated Gravitational Point Set Alignment With Altered Physical Laws"} +{"pid": "6528c219939a5f408299cc99", "context": "With the emergence of Vision-Language Pretraining Models (VLMs), prompt learning has become essential for their fast adaptation to downstream tasks with limited resources. However, existing research mostly focuses on single-prompt paradigms and rarely investigates multi-prompt learning.", "key_idea": "This paper extends the recent constant modality gap phenomenon to learnable prompts and justifies the superiority of vision-language transfer with multi-prompt augmentation. The authors propose an Energy-based Multi-prompt Learning (EMPL) to generate multiple prompt embeddings by drawing instances from an energy-based distribution, which is implicitly defined by VLMs.", "method": "The authors conducted comprehensive experiments to justify their claims and evaluate the performance of the proposed EMPL model.", "outcome": "While the specific results are not mentioned in the abstract, it states that comprehensive experiments have justified their claims and shown the excellence of EMPL.", "future_impact": "The proposed EMPL method could redefine the conventional single-prompt paradigm, offering a more efficient vision-language transfer with multi-prompt augmentation, leading to better balance between in-domain and out-of-domain open-vocabulary generalization.", "venue": "ICCV", "year": 2023, "title": "A Retrospect to Multi-prompt Learning across Vision and Language."} diff --git a/data/cot_examples.json b/data/cot_examples.json new file mode 100644 index 0000000..ae0c1a3 --- /dev/null +++ b/data/cot_examples.json @@ -0,0 +1,52 @@ +{ + "idea_generation": [ + { + "user": "Given the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', generate key ideas that could advance this area of study. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "Considering the limitations of traditional methods that depend heavily on static database statistics and execution plans, a shift towards dynamic, predictive models can offer real-time insights and adjustments. Prediction: The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans." + }, + { + "user": "Given the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', generate key ideas that could advance this area of study. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "Traditional kernels struggle with high-dimensional spaces due to their homogeneous nature. Incorporating adaptability and depth could enhance sensitivity to complex data distributions. Prediction: The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data." + } + ], + "method_recommendation": [ + { + "user": "Given the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.' and the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', recommend the most suitable method to validate this idea. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "To validate the proposed approach, leveraging a large dataset of diverse SQL queries would allow for robust training and testing of the machine learning model, ensuring its efficacy across various scenarios. Prediction: The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans." + }, + { + "user": "Given the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.' and the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', recommend the most suitable method to validate this idea. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "Experimental validation using diverse datasets, including those with high-dimensional attributes, would effectively demonstrate the model's adaptability and superiority over traditional methods. Prediction: Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method." + } + ], + "outcome_prediction": [ + { + "user": "Based on the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', and the recommended method: 'The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans.', predict the potential outcome of this research. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "Given the novel approach's reliance on real-time data and machine learning, it is expected to surpass traditional methods in accuracy and adaptability, providing actionable insights more reliably. Prediction: Empirical results demonstrate that the predictive accuracy of several query properties is higher when utilizing the new approach based on machine learning techniques, compared to existing methods." + }, + { + "user": "Based on the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', and the recommended method: 'Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method.', predict the potential outcome of this research. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "The integration of deep learning for kernel adaptation promises significant improvements in test performance, especially in complex scenarios where traditional kernels fall short. Prediction: The authors demonstrate the superior performance of their deep kernels in hypothesis testing on benchmark and real-world data, indicating they improve upon previous methods." + } + ], + "future_work_recommendation": [ + { + "user": "Based on the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', the method: 'The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans.', and the outcome: 'Empirical results demonstrate that the predictive accuracy of several query properties is higher when utilizing the new approach based on machine learning techniques, compared to existing methods.', suggest projected impact for this research. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "Given the success of this predictive model, exploring its applicability to other database systems or incorporating it into database management tools could further broaden its impact, making database operations more efficient and intuitive. Prediction: The introduction of a data-driven machine learning approach to predict SQL query properties can significantly accelerate and improve query formulation and analysis by providing insights prior to their execution, potentially transforming the way SQL queries are formed." + }, + { + "user": "Based on the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', the method: 'Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method.', and the outcome: 'The authors demonstrate the superior performance of their deep kernels in hypothesis testing on benchmark and real-world data, indicating they improve upon previous methods.', suggest projected impact for this research. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "Further advancements could include developing even more adaptable neural network architectures or extending these methods to other statistical tests to harness similar benefits in different contexts. Prediction: The ideas and methods proposed in this paper have the potential to significantly improve the efficacy of two-sample tests in the future, particularly in instances where the data is high dimensional and complex. As a special case, this scheme includes popular classifier-based two-sample tests but improves upon them." + } + ], + "title_prediction": [ + { + "user": "Given the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics.', the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class.', the method: 'The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries.', the outcome: 'Empirical results demonstrate that the predictive accuracy of several query properties is higher when utilizing the new approach based on machine learning techniques.', and the future impact: 'The introduction of a data-driven machine learning approach to predict SQL query properties can significantly accelerate and improve query formulation and analysis by providing insights prior to their execution.', predict the title of this research paper. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "The paper's focus is on enhancing the predictability and efficiency of SQL query processing using advanced machine learning models. Prediction: Facilitating SQL Query Composition and Analysis Through Data-Driven Predictive Modeling." + }, + { + "user": "Given the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces.', the method: 'Deep neural networks are trained to maximize the power of these two-sample tests. The authors perform experiments on benchmark and real-world data to validate their method.', the outcome: 'The authors demonstrate the superior performance of their deep kernels in hypothesis testing on benchmark and real-world data.', and the future impact: 'The ideas and methods proposed in this paper have the potential to significantly improve the efficacy of two-sample tests in the future.', predict the title of this research paper. Let's think step by step. You should first present your reasoning. After that, the final prediction should start after the marker 'Prediction:'.", + "assistant": "This paper introduces a novel, adaptable kernel model enhanced by deep learning techniques for two-sample testing, which could set a new standard for statistical tests in complex data environments. Prediction: Learning Deep Kernels for Non-Parametric Two-Sample Tests." + } + ] +} diff --git a/data/few_shot_examples.json b/data/few_shot_examples.json new file mode 100644 index 0000000..0980cdc --- /dev/null +++ b/data/few_shot_examples.json @@ -0,0 +1,32 @@ +{ + "idea_generation": [ + {"user": "Given the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', generate key ideas that could advance this area of study.", + "assistant": "The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans."}, + {"user": "Given the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', generate key ideas that could advance this area of study.", + "assistant": "The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data."} + ], + "method_recommendation": [ + {"user": "Given the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.' and the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', recommend the most suitable method to validate this idea.", + "assistant": "The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans."}, + {"user": "Given the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.' and the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', recommend the most suitable method to validate this idea.", + "assistant": "Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method."} + ], + "outcome_prediction": [ + {"user": "Based on the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', and the recommended method: 'The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans.', predict the potential outcome of this research.", + "assistant": "Empirical results demonstrate that the predictive accuracy of several query properties is higher when utilizing the new approach based on machine learning techniques, compared to existing methods."}, + {"user": "Based on the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', and the recommended method: 'Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method.', predict the potential outcome of this research.", + "assistant": "The authors demonstrate the superior performance of their deep kernels in hypothesis testing on benchmark and real-world data, indicating they improve upon previous methods."} + ], + "future_work_recommendation": [ + {"user": "Based on the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', the method: 'The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans.', and the outcome: 'Empirical results demonstrate that the predictive accuracy of several query properties is higher when utilizing the new approach based on machine learning techniques, compared to existing methods.', suggest projected impact for this research.", + "assistant": "The introduction of a data-driven machine learning approach to predict SQL query properties can significantly accelerate and improve query formulation and analysis by providing insights prior to their execution, potentially transforming the way SQL queries are formed."}, + {"user": "Based on the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', the method: 'Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method.', and the outcome: 'The authors demonstrate the superior performance of their deep kernels in hypothesis testing on benchmark and real-world data, indicating they improve upon previous methods.', suggest projected impact for this research.", + "assistant": "The ideas and methods proposed in this paper have the potential to significantly improve the efficacy of two-sample tests in the future, particularly in instances where the data is high dimensional and complex. As a special case, this scheme includes popular classifier-based two-sample tests but improves upon them."} + ], + "title_prediction": [ + {"user": "Given the context: 'Formulating efficient SQL queries is a challenging and iterative process requiring tuning and execution cycles. Current methods for providing insights about SQL query properties prior to execution depend on database instance statistics or query execution plans.', the key idea: 'The authors propose a new approach for predicting SQL query properties, including the query answer size, run-time, and error class, relying on data-driven machine learning techniques and large query workloads, instead of database stats or execution plans.', the method: 'The authors employ data-driven machine learning techniques, alongside extensive query workloads to model and predict properties of SQL queries, in contrast to traditional techniques dependent on database-related statistics or execution plans.', the outcome: 'Empirical results demonstrate that the predictive accuracy of several query properties is higher when utilizing the new approach based on machine learning techniques, compared to existing methods.', and the future impact: 'The introduction of a data-driven machine learning approach to predict SQL query properties can significantly accelerate and improve query formulation and analysis by providing insights prior to their execution, potentially transforming the way SQL queries are formed.', predict the title of this research paper. The title should be concise and reflective of the core aspects.", + "assistant": "Facilitating SQL Query Composition and Analysis"}, + {"user": "Given the context: 'Two-sample tests are utilized to determine if two sets of samples come from the same distribution. Current tests typically use kernels which are spatially homogeneous and adapt only in lengthscale. Such kernels are not well-suited for dealing with high dimensions and complex data.', the key idea: 'The authors suggest a new kernel-based two-sample test that is parameterized by deep neural networks. This test is designed specifically to adapt to variations in distribution shape and smoothness across different spaces, making it better suited for high dimensional and complex data.', the method: 'Deep neural networks are trained to maximize the power of these two-sample tests, allowing them to be particularly effective with complex data. The authors perform experiments on benchmark and real-world data to validate their method.', the outcome: 'The authors demonstrate the superior performance of their deep kernels in hypothesis testing on benchmark and real-world data, indicating they improve upon previous methods.', and the future impact: 'The ideas and methods proposed in this paper have the potential to significantly improve the efficacy of two-sample tests in the future, particularly in instances where the data is high dimensional and complex. As a special case, this scheme includes popular classifier-based two-sample tests but improves upon them.', predict the title of this research paper. The title should be concise and reflective of the core aspects.", + "assistant": "Learning Deep Kernels for Non-Parametric Two-Sample Tests"} + ] +} diff --git a/data/gpt35_0531.jsonl b/data/gpt35_0531.jsonl new file mode 100644 index 0000000..4ff3257 --- /dev/null +++ b/data/gpt35_0531.jsonl @@ -0,0 +1,120 @@ +{"id": "01f161fe-dd40-45dd-89bd-fb1562771d73", "Context": "Capturing contextual information within event forecasting is challenging due to several factors: (i) uncertainty of context structure and formulation, (ii) high dimensional features, and (iii) adaptation of features over time.", "Key Idea": "The authors propose a novel graph convolutional network for predicting future social events by extracting and learning graph representations from historical/prior event documents.", "Method": "The authors employ a graph convolutional network to predict future events and identify sequences of dynamic graphs as event context. They extract and learn graph representations from historical/prior event documents.", "Outcome": "Experimental results on multiple real-world data sets show that the proposed method is competitive against various state-of-the-art methods for social event prediction.", "Future Impact": "The proposed method could be beneficial in real-world applications such as automated analysis generation and resource allocation as it captures dynamic properties of event contexts as social indicators."} +{"id": "0b31e456-4944-47e5-80ed-deaf6421c375", "Context": "The paper proposes a tool, Directed Dialogue Protocols (DDPs) for interface design based on Katou0027s (1986) method of verbal data collection by question-asking protocols.", "Key Idea": "The paper proposes three extensions to the question-asking method: atomic tasks, interventions by the experimenter, and sequential disclosure to improve the process of verbal data collection.", "Method": "The paper discusses the design and application of the tool, DDPs. The method was applied to identify design choices that improve learnability and usability into a productu0027s user-interface.", "Outcome": "The paper doesn't present any measurable outcomes of this study but discusses the identified design choices that improve learnability and usability into a productu0027s user-interface. ", "Future Impact": "The method proposed in this paper can impact future studies in human-computer interaction and design that could utilize DDPs as a tool for facilitating the quantization of verbal data and improving the design of learnable and usable interfaces."} +{"id": "07c3daea-a88c-4a67-9aac-20ef0ec62e79", "Context": "Non-pharmacological interventions are the most common and the most effective for people with dementia. Using biographical or personalized materials has been proven to benefit these interventions, but it can be challenging to obtain and maintain such information.", "Key Idea": "The authors propose a web platform to collect and manage biographical materials in a streamlined manner, enabling reminiscence and other biographical cognitive stimulation practices for people with dementia.", "Method": "The authors conducted a case study with one psychologist and three patients across two weeks. The patients' data were collected using the web platform, and the impact of the therapy was measured.", "Outcome": "The web platform improved the collection of meaningful data about the patients, and the therapy's overall awareness was maintained. The study showed improvements in the collection of patient data and maintenance of the therapy's overall awareness.", "Future Impact": "The proposed platform could be extended to open new horizons in biographical cognitive interventions and personalized treatments for people with dementia."} +{"id": "081d6673-3c7c-4aec-b101-cf55d75ac718", "Context": "Previous partial permutation synchronization algorithms are often computationally intensive and memory-demanding.", "Key Idea": "The authors propose an improved algorithm, CEMP-Partial, for estimating the corruption levels of the observed partial permutations, and subsequently propose MatchFAME, a fast, accurate, and memory-efficient matching algorithm based on sparse matrix operations.", "Method": "The authors prove the CEMP-Partial algorithm's ability to classify corrupted and clean partial permutations. The authors also demonstrate the accuracy, speed, and memory efficiency of MatchFAME on synthetic and real datasets.", "Outcome": "MatchFAME shows state-of-the-art accuracy, speed, and memory efficiency compared to previous PPS algorithms and is able to classify corrupted and clean partial permutations under certain assumptions.", "Future Impact": "The improved algorithm and resulting matching algorithm may enable more efficient and accurate multi-object matching for large-scale structure-from-motion datasets."} +{"id": "14b0ebd1-b654-4eed-bdd8-ebeb74250b15", "Context": "In the few-shot relational triple extraction (FS-RTE), one extracts relational triples from plain texts by utilizing only few annotated samples. The entity-then-relation paradigm, which first extracts all entities and then classifies their relations, ignores the entity discrepancy between relations.", "Key Idea": "The authors propose a novel task decomposition strategy, Relation-then- Entity, for FS-RTE and a model, RelATE to instantiate this strategy. RelATE builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities.", "Method": "The authors experimentally show that RelATE outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).", "Outcome": "RelATE outperforms previous models with an absolute improvement of 18.98% and 28.85% F1 scores.", "Future Impact": "N/A"} +{"id": "16c1b4ae-73f8-4c23-8bdb-b931ade1baa5", "Context": "Various tasks in decision making and decision support require selecting a preferred subset of items from a given set of feasible items based on individual attribute values.", "Key Idea": "The paper proposes to solve the problem of computing an optimal subset of items given a specified preference based on the attribute values of individual elements within the set.", "Method": "The paper presents two algorithm classes: direct set construction and implicit enumeration as solutions to appropriate CSPs. New algorithms for each class are presented and compared empirically against previous results.", "Outcome": "The problem is shown to be NP-hard in the general case, and the proposed heuristic search methods modelled as CSPs perform better than previous approaches.", "Future Impact": "The new algorithms proposed in this paper could provide a foundation for solving more complex decision-making problems with conflicting preferences. "} +{"id": "10c15fe5-c315-4b6d-8910-e6bc3279c817", "Context": "Social event detection in an incremental learning setting, where acquiring, preserving, and extending knowledge are major concerns.", "Key Idea": "The authors propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection that explores the expressive power of GNNs for knowledge extraction and continuously adapts to incoming data by adopting contrastive loss terms.", "Method": "The authors use KPGNN for incremental social event detection, adopting contrastive loss terms that cope with a changing number of event classes in the inductive learning ability of GNNs to detect events and expand its knowledge from previously unseen data. They also adopt a mini-batch subgraph sampling strategy for scalable training and periodically remove obsolete data to maintain a dynamic embedding space.", "Outcome": "Extensive experiments demonstrate the superiority of KPGNN over various baselines.", "Future Impact": "The proposed KPGNN has the potential to improve applications in fields such as product recommendation and crisis management by providing valuable insights into group social behaviors and public concerns."} +{"id": "18f27ced-9f37-45d7-9b76-6663c349d408", "Context": "Cross-modal retrieval is a challenge where one type of data is used as a query to retrieve relevant data of another type. Existing approaches learn a common subspace in a joint manner, which requires all modalities to be involved in the training process.", "Key Idea": "The authors propose a novel cross-modal retrieval method called Scalable Deep Multimodal Learning (SDML) that trains modality-specific networks independently to transform multimodal data into a predefined common subspace.", "Method": "The authors train m modality-specific networks independently for m modalities and project the data into the predefined common subspace to achieve multimodal learning. The authors perform comprehensive experimental results on four benchmark datasets to demonstrate that SDML is effective and efficient.", "Outcome": "The proposed method (SDML) is effective and efficient in multimodal learning and outperforms the state-of-the-art methods in cross-modal retrieval on four benchmark datasets.", "Future Impact": "The proposed method could be useful in real-world applications, making cross-modal retrieval more scalable, efficient, and effective especially when handling samples from different modalities."} +{"id": "1b0e4045-d39b-4bea-8dec-e747f5c674f5", "Context": "Probabilistic databases have been developed to manage imprecise information inherent in applications such as sensor monitoring systems, location-based services, and biological databases.", "Key Idea": "The authors propose two efficient algorithms to discover frequent patterns and association rules from probabilistic data under the Possible World Semantics.", "Method": "The authors propose two algorithms that discover frequent patterns from probabilistic data in either top-down or bottom-up manners. The algorithms can also be extended to discover maximal frequent patterns. The authors conduct experiments on real and synthetic datasets to validate the performance of their proposed methods.", "Outcome": "The performance of the proposed methods is validated through extensive experiments conducted on real and synthetic datasets.", "Future Impact": "N/A"} +{"id": "1946f496-f6cd-4736-8c30-a6ae70baa8b2", "Context": "Existing click models treat intrinsic relevance as an atomic query-document-specific parameter, which is solely estimated from historical clicks without using any content information about a document or relationship among the clicked/skipped documents under the same query.", "Key Idea": "The authors propose a novel Bayesian Sequential State model for click modeling, where the document content and dependencies among the sequential click events within a query are characterized by a set of descriptive features via a probabilistic graphical model.", "Method": "The authors apply the posterior regularized Expectation Maximization algorithm for parameter learning and tailor the model to meet specific ranking-oriented properties, e.g., pairwise click preferences. Experiment results show that the proposed model is effective compared with state-of-the-art click models.", "Outcome": "The experimental results demonstrate that the proposed model is effective compared with state-of-the-art click models.", "Future Impact": "The proposed model may lead to the improvement of click models and better understanding of user-click behavior with richer information exploited from the user clicks."} +{"id": "1e396f93-a73e-4d33-9a8e-56097a8c3c28", "Context": "Deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER), but requires large amounts of labeled data.", "Key Idea": "Combining deep learning with active learning drastically reduces the amount of labeled training data required.", "Method": "The authors introduce a lightweight architecture called CNN-CNN-LSTM and carry out incremental active learning. The authors use standard datasets for the task.", "Outcome": "The proposed model achieves nearly state-of-the-art performance on standard datasets for NER while being computationally much more efficient.", "Future Impact": "The proposed approach of combining deep learning with active learning could potentially be applied to other NLP tasks and be used with different architectures for further improvement of the performance."} +{"id": "192f7803-df4d-40c0-b816-ba34339026b3", "Context": "Most current single image super-resolution (SR) methods use empirical risk minimisation, which often produces blurry and over-smoothed images. A more desirable approach would employ maximum a posteriori (MAP) inference, preferring solutions that always have a high probability under the image prior, making them appear more plausible.", "Key Idea": "The authors propose to use amortised MAP inference whereby they calculate the MAP estimate directly using a convolutional neural network to achieve plausible image outputs.", "Method": "The authors introduce a novel neural network architecture intended for the amortised MAP inference. They propose three methods to solve the optimisation problem, Generative Adversarial Networks (GAN), denoiser-guided SR, and a baseline method using a maximum-likelihood-trained image prior. The authors tested the proposed methods on real image data.", "Outcome": "The GAN-based approach outperformed the other two proposed methods when tested on real image data.", "Future Impact": "The amortised MAP inference method proposed in this study has potential use in a wide range of image-processing tasks beyond image super-resolution."} +{"id": "1dea5ec2-d311-4c03-bba5-e38d7a62fbd4", "Context": "The paper tackles the problem of spotting signs occurring in video sequences. The paper proposes to model spatio-temporal signatures of signs using an extension of sequential patterns that contain temporal intervals.", "Key Idea": "The authors proposed Sequential Interval Patterns (SIP) to model spatio-temporal signatures and a novel multi-class classifier called Hierarchical Sequential Pattern Tree (HSP-Tree) that organises different Sequential Interval Patterns in a hierarchical tree structure to enable exploration of any subsequence sharing that exists between SIPs of different classes.", "Method": "The authors evaluated the proposed method on both concatenated isolated sign sequences and continuous sign sequences and shows that the proposed method is superior in robustness and accuracy compared to state-of-the-art sign recognition approaches.", "Outcome": "The proposed method achieves superior robustness and accuracy compared to the state-of-the-art sign recognition approach.", "Future Impact": "N/A"} +{"id": "2a220303-8653-497f-b2b5-c829583c2714", "Context": "Counting and sampling directed acyclic graphs from a Markov equivalence class are fundamental tasks in graphical causal analysis.", "Key Idea": "The authors propose polynomial-time algorithms for counting and sampling directed acyclic graphs from a Markov equivalence class, solving a long-standing open problem in this area.", "Method": "The authors present effective and easily implementable algorithms. They conduct experiments to compare the performance of their method with that of state-of-the-art methods.", "Outcome": "The authors show that their algorithms significantly outperform state-of-the-art methods in terms of computational efficiency for counting and sampling directed acyclic graphs.", "Future Impact": "The proposed algorithms have practical implications for graphical causal analysis and can be applied to real-world datasets."} +{"id": "1ef9b762-e9be-46c5-ad19-090fe16200c4", "Context": "Paper addresses the issue of accurately deriving pointing information from a corresponding gesture in human-robot interaction.", "Key Idea": "The paper proposes a novel approach that fuses information from two different input streams, head pose estimated by visually tracking the face, and hand pointing orientation using Dempster-Shafer theory of evidence, to decide on the pointed object.", "Method": "The proposed approach tracks off-plane face rotations and hand pointing orientation to fuse information and estimate pointed targets.", "Outcome": "The method is validated experimentally and proven effective in realistic application setups.", "Future Impact": "The method could enhance the accuracy of pointed target identification in human-robot interaction setups and similar applications."} +{"id": "1e5571af-71cf-40b8-ba2a-18b42cae5b42", "Context": "The paper addresses knowledge-based sequence mining.", "Key Idea": "The authors introduce a new framework for knowledge-based sequence mining using Answer Set Programming (ASP), and show how preferred patterns of interest can be easily extracted.", "Method": "The authors demonstrate their framework using several modular extensions, and use ASP's preference handling capacities to mine patterns of interest with expert-declared importance. The authors also compare the effectiveness of their approach with a related sequence mining mechanism through empirical study.", "Outcome": "The paper demonstrates that the proposed approach can easily incorporate domain-specific knowledge in the form of preferred patterns, which is effective in real-world sequence mining tasks.", "Future Impact": "The proposed knowledge-based sequence mining framework can contribute to the development of more intelligent pattern mining algorithms and systems. Future research might explore improving the scalability and efficiency of the approach."} +{"id": "24e29617-a320-450a-aaa5-19d8700d74b7", "Context": "Matrix completion has been a crucial problem in machine learning and involves filling in missing values in a given matrix. Existing synchronous algorithms have not been efficient in solving this problem.", "Key Idea": "The authors proposed NOMAD, a non-locking and asynchronous decentralized algorithm with non-blocking communication between processors for matrix completion. One key feature of the algorithm is that it is lock-free and ownership of a variable is asynchronously transferred between processors.", "Method": "The authors' method is an extensive empirical evaluation that shows NOMAD's superior performance to existing synchronous algorithms on commodity hardware and HPC cluster both in multi-core and distributed memory scenarios.", "Outcome": "NOMAD outperforms synchronous algorithms in both commodity hardware and HPC cluster. NOMAD's variable updates are serializable despite being asynchronously updated.", "Future Impact": "N/A"} +{"id": "206d2d53-dbaf-4a2f-810d-856309d8eb83", "Context": "The paper proposes a neural head reenactment system.", "Key Idea": "The system is driven by a latent pose representation learned solely on image reconstruction losses, which can successfully decompose pose from identity and perform cross-person reenactment.", "Method": "The authors show the effectiveness of the proposed system by predicting the foreground segmentation and reproducing mimics of the driving person. The authors also demonstrate that the learned descriptors are useful for keypoint prediction and pose-based retrieval.", "Outcome": "The proposed system successfully predicts the foreground segmentation and can perform cross-person reenactment. The learned descriptors are useful for keypoint prediction and pose-based retrieval.", "Future Impact": "The proposed system has potential applications in the fields of computer vision and graphics, such as entertainment, video communication, and virtual reality."} +{"id": "2065b977-7782-4981-ad70-3121a2315687", "Context": "The authors propose a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition.", "Key Idea": "The authors integrate unigram and bigram LSA into the N-gram LM via marginal adaptation and linear interpolation respectively, achieving improved results on both the Mandarin RT04 test set and the large-scale evaluation on Arabic.", "Method": "The model is trained using variational EM and fractional Kneser-Ney smoothing is proposed to handle fractional counts. Bootstrapping of bigram LSA from unigram LSA is used to address scalability issues.", "Outcome": "The proposed approach yields a 6%-8% relative perplexity reduction and 2.5% relative character error rate reduction compared to applying only unigram LSA on the Mandarin RT04 test set. A 3% relative word error rate reduction is achieved on the large-scale evaluation on Arabic.", "Future Impact": "N/A"} +{"id": "36d7073a-d006-4a4a-850f-ca7a3f0ca19b", "Context": "Many important problems can be framed as learning from graph data. Previous methods either do not scale well or lack expressivity.", "Key Idea": "The authors propose a framework for learning convolutional neural networks for arbitrary graphs, regardless of whether they are directed or undirected and have discrete or continuous node and edge attributes. The authors present a general approach to extracting locally connected regions from graphs.", "Method": "The authors demonstrate the effectiveness and efficiency of the proposed method compared to previous methods, using established benchmark datasets.", "Outcome": "The learned feature representations are competitive with state-of-the-art graph kernels, and their computation is highly efficient.", "Future Impact": "The proposed framework has the potential to advance the state of the art in learning from graph data, enabling the development of deep learning models for graph data that are scalable and expressive. It may also find applications in various fields, such as social network analysis and drug discovery."} +{"id": "2eb66e5a-472c-4db1-b02f-47fe5eb7e71e", "Context": "The proposed model does not need semantic information such as paraphrase pairs or syntactic information such as constituency parses. The proposed model relies solely on the inductive bias found in attention-based architectures such as Transformers.", "Key Idea": "The authors propose QKVAE, a generative model that exhibits disentangled latent representations of syntax and semantics using a decoder that reads latent variables where one variable infers keys and another infers values.", "Method": "The authors run experiments on latent representations and syntax/semantics transfer to validate QKVAE's disentangled syntax and semantics. The authors compare QKVAE's syntax transfer capabilities to supervised models and evaluate the amount of data required for comparable supervised models to outperform QKVAE.", "Outcome": "QKVAE displays clear signs of disentangled syntax and semantics according to experiments on latent representations and syntax/semantics transfer. QKVAE displays competitive syntax transfer capabilities compared to supervised models with less data.", "Future Impact": "The proposed model could potentially be applied to solving related unsupervised NLP tasks, such as multilingual text generation and dialogue generation."} +{"id": "37da719b-8db9-4523-8a4a-3f2eef238978", "Context": "The paper deals with image watermarking and recovery of tampered images.", "Key Idea": "The key idea is to use a fragile watermarking with self-embedding scheme that does not use authentication bits. The authors propose a spread spectrum-based watermarking scheme using block-based embedding and DCT-based compression.", "Method": "The authors propose a fragile watermarking scheme using self-embedding. The watermark is embedded in the blocks of an image using DCT-based compression. The authors use simulation to test the effectiveness of the proposed method.", "Outcome": "The simulation results demonstrate that the proposed method provides good recovery performance.", "Future Impact": "The proposed method could be applied in other multimedia fields to protect intellectual property, such as video and audio watermarking."} +{"id": "29dd9fd3-6c98-4e4b-b70c-0474ff361419", "Context": "Building a deformable shape model for a new species requires 3D data. The accurate shape model is not available for birds despite being a species with almost double the number of species as mammals.", "Key Idea": "The authors propose a method to capture new bird species using an articulated template and images of that species.", "Method": "The authors fit the articulated template to each training sample to disentangle pose and shape and learn a shape space that captures variation among and within bird species from image evidence. The authors learn models of multiple bird species from the CUB dataset and contribute new species-specific and multi-species shape models.", "Outcome": "The authors show that their learned 3D shape space better reflects the phylogenetic relationships among birds than learned perceptual features through the use of a low-dimensional embedding.", "Future Impact": "The proposed method can be extended to capture shape models of other animal species beyond birds using articulated templates and images, facilitating downstream reconstruction tasks. The learned 3D shape space can also be used in downstream computer vision tasks."} +{"id": "370141c7-e1bb-4010-9938-efcad6cf2e62", "Context": "Innovations in neural architectures have fostered significant breakthroughs in language modeling and computer vision. Novel architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized. A number of architecture-specific initialization schemes have been proposed.", "Key Idea": "The authors propose GradInit, an automated and architecture agnostic method for initializing neural networks using a simple scalar multiplier to adjust the layer norms to improve convergence and test performance.", "Method": "The authors apply GradInit to many convolutional architectures, including architectures without normalization layers, and the original Transformer architecture for machine translation. The authors optimize the scalar multiplier using a simple numerical scheme.", "Outcome": "GradInit improves the convergence and test performance of many convolutional architectures and improves the stability of the original Transformer architecture for machine translation.", "Future Impact": "GradInit offers an automated and architecture-agnostic approach for initializing neural networks, which could improve the stability and convergence of various neural architectures in the future. GradInit could become a standard method for initializing neural networks, as it does not depend on architecture-specific initialization schemes."} +{"id": "31ab88b5-e66f-4b69-98b6-7a470dce9875", "Context": "Conditional neural networks play an important role in sequence-to-sequence modeling tasks. In conditional neural networks, the output of a model is often influenced by a conditioning vector, in addition to the input.", "Key Idea": "The authors propose a novel approach of neural network conditioning by learning intermediate layer activations based on the conditioning vector to eliminate the cost of the increased model size.", "Method": "The authors explore the proposed approach and show that learned activation functions can produce models with comparable or better quality, while decreasing model sizes. The authors evaluate the approach on the personalized sound enhancement and personalized ASR tasks.", "Outcome": "The authors find that conditioning via activation function learning is an effective modeling strategy. The proposed approach produces conditional models with comparable or better quality while decreasing the model size, making them ideal candidates for on-device deployment.", "Future Impact": "The proposed approach can reduce computational costs and memory footprints for conditional neural network models, making them more feasible for deployment on devices with limited resources."} +{"id": "3ac464c2-2214-4bf2-a6b6-03da2498cb03", "Context": "Query scheduling is a fundamental problem in database management systems and has received renewed attention due to the rise of the Database as a Service (DaaS) model. Researchers have focused on different scheduling algorithms, but little is known about what information these algorithms should know about the queries to be scheduled.", "Key Idea": "The authors investigate using histograms describing the distribution of likely query execution times as input to the query scheduler and propose a novel distribution-based scheduling algorithm called Shepherd.", "Method": "The authors compare the proposed Shepherd algorithm with state-of-the-art point-based methods through extensive experimentation with synthetic and TPC workloads.", "Outcome": "The authors show that the proposed Shepherd algorithm substantially outperforms state-of-the-art point-based methods.", "Future Impact": "The proposed approach can improve query scheduling algorithms and can be explored in more complex distributed environments like cloud-based deployments."} +{"id": "2b5cc037-4841-4fb1-85ff-673230198be1", "Context": "This paper focuses on the problem of recovering a life-long incomplete m x n matrix of rank r with columns arriving online over time. The matrix completion problem is widely applied to various domains including recommendation systems, computer vision, and system identification. Prior works have considered the noise-free case, but designing provable algorithms under realistic noise models is a challenge.", "Key Idea": "The authors propose algorithms for life-long matrix completion that are tolerant to two different realistic noise models: bounded deterministic noise and sparse random noise. The algorithms have low sample complexity as compared to previous results in the noise-free case.", "Method": "The authors present algorithms for both bounded deterministic noise and sparse random noise and study the scenario where the hidden matrix lies on a mixture of subspaces. The proposed algorithms are evaluated experimentally on both synthetic and real-world datasets.", "Outcome": "The proposed algorithms return a matrix of small error with sample complexity almost as small as the best prior results in the noiseless case for the bounded deterministic noise model. For the sparse random noise model, the proposed algorithm exactly recovers an mu_0-incoherent matrix by probability at least 1 - delta with sample complexity as small as O(mu_0rnlog(r/delta)), advancing the state-of-the-art work and matching the lower bound in the worst case.", "Future Impact": "The proposed algorithms for life-long matrix completion that are tolerant to realistic noise models could significantly improve the performance of recommendation systems, computer vision, and system identification tasks that are vulnerable to data corruption and incompleteness. The study of the scenario where the hidden matrix lies on a mixture of subspaces could lead to more efficient algorithms with even lower sample complexity."} +{"id": "3b076d81-3ed0-4d57-84e8-4145f67052bd", "Context": "Referring expressions and other object descriptions should be maximal under the Local Brevity, No Unnecessary Components, and Lexical Preference preference rules; otherwise, they may lead hearers to infer unwanted conversational implicatures.", "Key Idea": "The authors propose a polynomial time generation algorithm that incorporates preference rules to avoid unwanted conversational implicatures while generating referring expressions.", "Method": "The authors incorporate Local Brevity, No Unnecessary Components, and Lexical Preference preference rules into a polynomial time generation algorithm. The authors also discuss alternative formalizations of conversational implicature that make the generation task NP-Hard.", "Outcome": "N/A", "Future Impact": "The proposed polynomial time algorithm can be used to generate referring expressions and other object descriptions that avoid conversational implicatures, which have implications in natural language processing and communication applications."} +{"id": "3b6e7572-2ef8-4565-a3b7-301a3fd38acd", "Context": "Backward locking and update locking are sources of inefficiency in backpropagation that prevent from concurrently updating layers. Local error signals have been suggested as a solution to train network blocks asynchronously to overcome these limitations", "Key Idea": "The authors propose a differentiable algorithm named SEDONA to automate the process of discovering the optimal configuration for local training.", "Method": "The authors run experiments to show that their algorithm can discover transferable decoupled architectures for VGG and ResNet variants and they evaluate their approach on CIFAR-10, Tiny-ImageNet, and ImageNet.", "Outcome": "The authors report consistently better performance on the evaluated tasks compared to end-to-end backpropagation and other state-of-the-art greedy-leaning methods. They also report up to a 2× speedup over backpropagation in total training time.", "Future Impact": "The proposed search algorithm could be applied to larger and more complex models to further improve performance and speed, and reduce reliance on human design and intuition."} +{"id": "45b76955-9670-4664-939c-f5a61eb597df", "Context": "The paper proposes a novel metric for auditing group fairness in ranked lists by accounting for varying user behaviors and allowing non-binary protected attributes to enable investigating inherently continuous attributes.", "Key Idea": "The authors model user attention as varying rather than assuming a logarithmic loss in importance as a function of the rank and allow non-binary protected attributes to enable measurements across aggregated sets of search results, rather than separately for each result list, to better address the human factors inherent in fair group representation in ranked lists.", "Method": "The authors propose a novel metric and perform three simulated fairness audits using the metric.", "Outcome": "The metric can better address the human factors inherent in fair group representation in ranked lists.", "Future Impact": "N/A"} +{"id": "4164fd9a-7ab8-4013-b416-a205231f10f2", "Context": "Plan recognition has been shown to be solvable using modified planning algorithms without the need for a plan library", "Key Idea": "The authors extend the above approach to probabilistic plan recognition by defining the posterior goal probabilities in terms of cost differences of achieving the goal under two conditions: complying with observations and not complying with them. These posterior probabilities are computed using off-the-shelf classical planners.", "Method": "The authors use two calls to a classical planner to compute the cost difference mentioned above and posterior goal probabilities. The authors provide examples to illustrate the quality, flexibility, and scalability of their approach.", "Outcome": "The authors show that their approach efficiently solves the problem of probabilistic plan recognition. They do not provide measurable results.", "Future Impact": "The proposed approach could simplify the process of probabilistic plan recognition by using off-the-shelf classical planners and could extend the applicability of probabilistic plan recognition to a wider range of scenarios."} +{"id": "438caf1c-5c7e-4283-a3cd-bbab302df185", "Context": "Active learning commonly uses uncertainty or diversity sampling selection, aiming to select data points from the pool of unlabeled data. Of these, uncertainty and diversity sampling are best-known acquisition functions.", "Key Idea": "The authors present Contrastive Active Learning - an acquisition function called CAL - that selects contrastive examples. CAL selects similar data points in the feature space, which have maximally different predictive likelihoods.", "Method": "The authors compare CAL with diverse acquisition functions in four natural language understanding tasks and seven datasets. The authors conduct an extensive ablation study of the proposed method and analyze all actively acquired datasets.", "Outcome": "CAL consistently outperforms or equals the best performing baseline across all tasks, on both in-domain and out-of-domain data.", "Future Impact": "CAL achieves a better trade-off between selecting the right examples by calculating similarity and predictive differences between example pairs. The authors expect that CAL can improve datasets in various domains where acquiring the right examples is key, such as medical and legal domains."} +{"id": "48bacac1-1ca9-4be8-90e6-470596de0e26", "Context": "Previous studies have explored (semi-) unsupervised methods of learning polarity of words and phrases for the purpose of building a lexicon automatically in sentiment analysis.", "Key Idea": "The authors propose to use structural clues to extract polar sentences from Japanese HTML documents and build lexicon from the extracted sentences. The structural clues have extremely high precision at the cost of recall.", "Method": "The authors explore the use of structural clues to extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences. They compensate for the low recall using a massive collection of HTML documents.", "Outcome": "The authors were able to extract enough polar sentence corpus from the massive collection of HTML documents to build a lexicon for sentiment analysis.", "Future Impact": "N/A"} +{"id": "432f5702-7b7d-4995-812e-40925f1a18dd", "Context": "The research effort in machine learning is generally focused on conditional modeling when x is high dimensional. The authors address the opposite case of high dimensional y.", "Key Idea": "The authors propose the Landmark Selection Method for Multiple Output Prediction. Their approach involves selecting a small subset of the dimensions of y and modeling x → yL and yL → y to obtain a conditional model x → y that has convenient statistical properties.", "Method": "The authors provide several multilabel classification and multivariate regression experiments across different datasets to compare the proposed method with one-vs-all approach and several sophisticated multiple output prediction methods.", "Outcome": "The proposed method outperforms one-vs-all approach as well as several sophisticated multiple output prediction methods when tested on several datasets for multilabel classification and multivariate regression problems.", "Future Impact": "N/A"} +{"id": "4e6f8004-9384-4c5c-8d7f-265410a290df", "Context": "Several deep learning methods have been proposed for completing partial data from shape acquisition setups. However, these methods only complete the partial shape with a single output, ignoring the ambiguity when reasoning the missing geometry.", "Key Idea": "The authors pose a multi-modal shape completion problem and propose to use a conditional generative modeling approach to complete the partial shape via a one-to-many mapping.", "Method": "The authors develop a method that learns a multimodal distribution of possible results and conditions the completion on this distribution. They evaluate their approach on several datasets.", "Outcome": "The authors demonstrate the merit of their approach in completing partial shapes with both diversity and quality. They compare their approach against several baseline methods and variants of their method qualitatively and quantitatively.", "Future Impact": "This proposed method may have potential applications in fields such as robotics and autonomous driving, where partial data may come from sensors with varying degrees of accuracy and completeness."} +{"id": "4fdcceeb-f50f-4c4f-8b92-5985498114f8", "Context": "The authors discuss the problem of recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images. Feature correspondence over multiple frames is assumed.", "Key Idea": "The authors propose a method of using algebraic geometry and polynomial equations to recover the 3D motion of a non-rigid object undergoing uniform expansion and shearing from stereo images.", "Method": "The authors reduce the problem of recovering the 3D motion uniquely to the solution of a set of homogeneous polynomial equations using the commutative algebra software package MACAULAY, and the Fortran polynomial continuation program POLSYS. The approach depends on feature correspondence over multiple frames.", "Outcome": "It is shown that the proposed method requires only two stereo snapshots with four points correspondence to determine the motion uniquely.", "Future Impact": "N/A"} +{"id": "532e797a-4b72-488a-80e4-03713d3c8435", "Context": "Existing methods for Non-Rigid Structure-from-Motion (NRSfM) rely on linear low-order or low-rank shape models. These methods are not effective when the observed object is performing non-linear motion.", "Key Idea": "The authors propose a new strategy to NRSfM that exploits the property of shape recurrency, which is a generalization of the concept of rigidity.", "Method": "The authors reduce NRSfM to a rigid-SfM method using camera view clustering and automatic recurrency detection. The proposed method is evaluated with experiments on both simulated sequences and real data.", "Outcome": "The proposed method is demonstrated to be effective in the reconstruction of non-rigid dynamic shapes, as shown by experiments on both simulated sequences and real data.", "Future Impact": "The proposed method offers a novel perspective on rethinking structure-from-motion, opening up possibilities for novel problems in the field."} +{"id": "58279154-e623-46d2-a431-cc409e094e2c", "Context": "The paper addresses a model of analogy-driven theorem proving that is more general and cognitively more adequate than previous approaches.", "Key Idea": "The authors propose an analogy-driven proof-plan construction method that employs a source proof-plan to guide the construction of a proof-plan for the target problem.", "Method": "The model works at the level of proof-plans and includes a reformulation of the source proof-plan. Several well-known theorems were processed using this approach.", "Outcome": "Using the proposed analogy-driven proof-plan construction method, the authors were able to prove several theorems that could not be proven by previous approaches.", "Future Impact": "N/A"} +{"id": "56992082-e04e-4a8b-a985-abfea27fc2e0", "Context": "Dynamic network pruning achieves runtime acceleration by dynamically determining the inference paths based on different inputs.", "Key Idea": "The authors propose a method for explicitly modeling discrete weight channel selections to achieve more sparse runtime inference paths and visualizing the network decision paths for model interpretability.", "Method": "The authors conduct experiments on CIFAR10 and ImageNet datasets to demonstrate the effectiveness of their approach. They also propose an adversarial example detection algorithm by discriminating the runtime decision features.", "Outcome": "The proposed dynamic network achieves higher prediction accuracy under similar computing budgets on CIFAR10 and ImageNet datasets compared to traditional static pruning methods and other dynamic pruning approaches. The proposed adversarial detection algorithm significantly improves the state-of-the-art detection rate across multiple attacks.", "Future Impact": "The proposed method provides an opportunity to build an interpretable and robust model with higher accuracy and reduced computing costs, which can be applied in various fields involving large-scale neural networks and adversarial attacks."} +{"id": "49309d5a-5959-4f8f-ae30-9fd2350f0cbc", "Context": "Deploying CNNs on constrained devices poses a challenge due to the huge memory and compute resources required by CNNs. Existing compression techniques struggle to be computationally friendly.", "Key Idea": "The authors propose focused quantization, a novel quantization strategy that exploits the weight distributions after fine-grained pruning and dynamically discovers the most efficient numerical representation for weights in layers with varying sparsities. The proposed method significantly reduces model sizes and replaces expensive multiplications with cheaper bit-shift operations.", "Method": "The proposed method is evaluated on ResNet-50 and ResNet-18. The authors achieve a compression ratio of 18.08x with only 0.24% loss in top-5 accuracy, outperforming existing compression methods on ResNet-50. They also show that ResNet-18 is higher in compression ratios and more hardware efficient than other state-of-the-art quantization methods given the same throughput.", "Outcome": "Focused quantization achieves a significantly reduced model size with a minimal loss in accuracy. The proposed method outperforms existing compression methods on ResNet-50 in terms of compression ratios and accuracy and is more hardware efficient than other state-of-the-art quantization methods in ResNet-18.", "Future Impact": "The proposed focused quantization method may enable CNN deployment on constrained devices where traditional compression techniques struggle to provide computationally friendly solutions."} +{"id": "5b9f94f9-d93f-455d-a110-007ad67ada6d", "Context": "Existing specification languages for tree based grammars do not support identifier management adequately.", "Key Idea": "The paper argues that XMG provides an effective treatment of identifiers for a linguist-friendly grammar design.", "Method": "The authors demonstrate that Coreference Handling in XMG provides a sophisticated treatment of identifiers irrespective of the quantity of cross-referencing. They provide examples of the mechanism in XMG using multiple examples.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "5b47ca6b-ff9b-429c-adb1-ad9a171eea0e", "Context": "The paper discusses the building of style-adapted maximum entropy language models for speech recognition, given a large corpus of written language data and a small corpus of speech transcripts. The linear interpolation method is typically used in such cases.", "Key Idea": "The authors investigate a Bayesian adaptation method for building style-adapted maximum entropy language models that consistently outperforms linear interpolation.", "Method": "The authors perform experiments to show the superiority of the Bayesian adaptation method over linear interpolation.", "Outcome": "Experiments show that the Bayesian adaptation method consistently outperforms linear interpolation in building style-adapted maximum entropy language models for speech recognition.", "Future Impact": "N/A"} +{"id": "5a3da6ef-67b0-41bc-a994-fc5ff455a27b", "Context": "Online reviews play a crucial role in today's electronic commerce, but due to pervasive spam reviews, customers may be misled, and decent stores may be defamed by malicious reviews. The problem of detecting spam reviews in singleton reviews is challenging.", "Key Idea": "The authors propose to detect spam attacks in singleton review based on the unusually correlated temporal patterns via identification and construction of a multidimensional time series based on aggregate statistics.", "Method": "The authors propose a hierarchical algorithm to robustly detect the time windows where such spam attacks are likely to have happened. The algorithm also pinpoints such windows in different time resolutions to facilitate faster human inspection.", "Outcome": "Experimental results show that the proposed method is effective in detecting singleton review attacks which was found to be a significant source of spam reviews, largely affecting the ratings of online stores.", "Future Impact": "The proposed approach provides a practical solution in detecting review spam in online e-commerce which is widely applicable to other applications dealing with bursty events."} +{"id": "5d87de73-77d4-4efd-b8e1-d7561b13f69f", "Context": "The paper addresses the problem of preposition sense disambiguation and evaluates its results against the SemEval 2007 Preposition Sense Disambiguation datasets.", "Key Idea": "The authors present a supervised classification approach for disambiguating preposition senses by using linguistically motivated features derived from both sides of the preposition utilized through phrase structure.", "Method": "The authors employed five different classifiers to test the effectiveness of proposed approach.", "Outcome": "The approach shows an increased accuracy that outperforms the best system in the SemEval task.", "Future Impact": "N/A"} +{"id": "5c268324-e160-489d-9722-6b59d5e3471b", "Context": "The Web accessibility is limited as the majority of designers do not know how to design more usable and accessible websites for disabled people.", "Key Idea": "The authors propose a new approach to achieve a more accessible Web through better knowledge-acquisition mechanisms and designing mechanisms that elicit knowledge from minorities of Web users in order to include marginal Web users.", "Method": "The approach includes detecting the Web designer's needs for knowledge, collecting this knowledge from minorities of Web users, and designing mechanisms that elicit such knowledge from Web users.", "Outcome": "The proposed approach can build a more accessible Web for the visually-impaired knowledge contributors and enable self-interested designers to be more accessible through better knowledge-acquisition mechanisms.", "Future Impact": "The approach presents a model for designing for marginalized users and promotes inclusion in the design process, which can be replicated to different areas of design to promote inclusion in their fields."} +{"id": "5dccca98-2b58-47e3-9b8f-3b1888aa3976", "Context": "Despite the vision of real-time data warehousing, the heterogeneity of today’s IT environment and increasing demands from mobile users pose significant challenges to be overcome by the traditional middleware solutions.", "Key Idea": "The authors propose a new middleware paradigm called space based computing to meet the challenges of creating real-time data warehousing in heterogeneous platforms and systems. ", "Method": "The paper presents space-based computing as an innovative middleware component that offers a level of abstraction superior to conventional middleware solutions, including distributed transactions and the seamless integration of mobile devices using open standards. The authors present a real-time build-up of data warehouses as an example of how this approach could work in practice.", "Outcome": "N/A", "Future Impact": "The proposed approach may have a significant impact on the development of real-time data warehousing and other applications that require seamless integration of heterogeneous platforms and systems while improving the performance of middleware solutions."} +{"id": "5e1f387c-d883-4d1f-8397-e4a533a3387b", "Context": "Hyperspectral imaging suffers from various sources of degradations, and lack of accurate ground-truth 'clean' hyperspectral signals makes restoration tasks challenging. Traditional neural networks struggle to train correctly due to limited available data.", "Key Idea": "The authors propose a new approach to hyperspectral image restoration based on sparse coding principles, merging classical techniques with deep learning to allow training of parameters without significant data requirements.", "Method": "The proposed spectral-spatial sparse coding model is applied to various denoising benchmarks and is shown to be computationally efficient while significantly outperforming the current state of the art.", "Outcome": "The proposed spectral-spatial sparse coding model is shown to be highly effective in improving denoising benchmarks.", "Future Impact": "The hybrid approach based on sparse coding principles proposed in this paper is a promising avenue for future research in hyperspectral imaging restoration and may be useful for other applications with limited training data."} +{"id": "69aacc53-6730-4db0-b420-9a45b96a642e", "Context": "Finding the configuration of a collection of geometric bodies that satisfies a set of given constraints is a significant geometric reasoning problem.", "Key Idea": "Automatically synthesizing a set of specialized routines called plan fragments from first principles about geometric bodies, actions, and topology and using them to solve the geometric constraint satisfaction problem.", "Method": "The paper proposes an approach that synthesizes the plan fragments automatically using first principles about geometric bodies, actions, and topology.", "Outcome": "The proposed approach is shown to solve the geometric constraint satisfaction problem efficiently by employing a set of specialized routines called plan fragments.", "Future Impact": "The proposed approach may be used in applications relying on geometric reasoning, such as computer-aided design (CAD), and may pave the way for more efficient and automated solutions for geometric constraint satisfaction problems."} +{"id": "68f8d058-1403-4066-b3d3-a8a2836b35e1", "Context": "The paper focuses on the problem of imperfect vector training labels with registration uncertainty, which is important in applications such as streamline classification on Earth imagery or tissue segmentation on medical imagery.", "Key Idea": "The authors propose a novel deep learning framework that explicitly quantifies vector labels' registration uncertainty and a registration-uncertainty-aware loss function to train neural networks with uncertain label locations.", "Method": "The authors design an iterative uncertainty reduction algorithm by estimating the posterior of true vector label locations distribution based on a Gaussian process and evaluate their approach on real-world datasets in National Hydrography Dataset refinement.", "Outcome": "The proposed approach significantly outperforms several baselines in both the registration uncertainty estimations performance and classification performance.", "Future Impact": "The proposed framework has the potential to improve the efficiency of annotating precise vector labels, potentially leading to significant cost savings in image analysis applications that rely on such labels. The uncertainty quantification method may be useful in other image analysis applications as well."} +{"id": "6dc39f88-d613-4ec0-b70d-d5daa6f3643c", "Context": "BitGourmet is a novel data analysis system that supports deterministic approximate query processing (DAQ).", "Key Idea": "BitGourmet operates on a carefully selected data subset to satisfy the user-defined precision constraint by dividing each column vertically, bit-by-bit, and uses a specialized query processing engine that evaluates queries on subsets of these bit vectors.", "Method": "The proposed system executes aggregation queries and produces deterministic bounds that contain the true value. The specialized query processing engine uses scenario-specific query optimizer, quality, and cost models to decide the optimal bit selection and execution plan. In the demonstration, the authors show that DAQ realizes an interesting trade-off between result quality and execution time making data analysis more interactive.", "Outcome": "The proposed system provides deterministic bounds that are guaranteed to contain the true value and operates in case of user-defined precision constraints. The authors also demonstrate that DAQ realizes an interesting trade-off between result quality and execution time that results in more interactive data analysis.", "Future Impact": "The authors do not explicitly state the Impact of their work, but the proposed approach could potentially reduce the computational cost of data analysis and provide deterministic bounds with user-defined precision constraints. Further research in this direction may lead to more interactive data analysis and reduced computational costs."} +{"id": "6de74297-fb80-448f-b7ae-41f8d9701044", "Context": "Previous literatures have used different imaging models to describe central catadioptric and fisheye cameras separately, and they have not been considered within a unified imaging model.", "Key Idea": "The authors present a unified imaging model that can be used to describe both central catadioptric and fisheye cameras, and show that existing calibration methods for central catadioptric cameras can be directly applied to fisheye cameras.", "Method": "The authors propose a unified imaging model for both central catadioptric and fisheye cameras. They demonstrate how fisheye images can be transformed into central catadioptric images under this new model, and how this model enables the use of existing calibration methods for central catadioptric cameras on fisheye cameras as well.", "Outcome": "The experimental results of calibration from some central catadioptric and fisheye images confirm the validity and usefulness of the proposed unified imaging model.", "Future Impact": "The unified imaging model proposed in this study can be used to develop new calibration methods for fisheye cameras that are metric, and can also be used to develop new imaging techniques that require a unified model for both central catadioptric and fisheye cameras."} +{"id": "73e353a8-e0d6-466f-af93-6fccf38fcb18", "Context": "Video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting. Dialogue context is mostly used as a simple text input without modeling the inherent information flows at the turn level in previous approaches.", "Key Idea": "The authors propose to discover information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer, and introducing a new approach that learns to predict reasoning paths over this semantic graph.", "Method": "The authors construct semantic graphs based on dialogue context, and propose a model that predicts reasoning paths through the graph. Additionally, their reasoning model processes both visual and textual information through this reasoning path and the propagated features are used to generate the answer. ", "Outcome": "Experimental results demonstrate the effectiveness of the proposed method.", "Future Impact": "N/A"} +{"id": "7b0671d7-c8f6-4e81-828b-c73958a6a63a", "Context": "The paper describes the design experience of an interactive exhibit about medieval music with three major goals: educational value, entertainment aspects, and historic authenticity.", "Key Idea": "The authors focused on the design challenges and solutions of an interactive exhibit about medieval music that balances educational value, entertainment aspects, and historic authenticity.", "Method": "The authors describe their design experience in developing the interactive exhibit that relies on audio as its only feedback channel.", "Outcome": "The paper presents insight into the challenges in designing an interactive exhibit about medieval music with three major goals: educational value, entertainment aspects, and historic authenticity, and how they could be solved.", "Future Impact": "N/A"} +{"id": "754f792b-fcf3-42f4-bb2c-5c1dcce21d7a", "Context": "The authors argue in favor of using formal meaning representations for natural language. An important use case is the problem of identifying the possible antecedents of anaphoric expressions.", "Key Idea": "The authors propose that a formal meaning representation should have structural properties that facilitate the identification of possible antecedents for anaphoric expressions.", "Method": "The authors do not describe a specific experiment. The research is theoretical and involves proposing a set of structural properties that a formal meaning representation should have.", "Outcome": "N/A", "Future Impact": "The adoption of formal meaning representations with structural properties to facilitate the identification of possible antecedents may improve the ability of computational models to deal with anaphoric expressions."} +{"id": "7c065e41-7c2e-430e-a918-6a776037bf50", "Context": "The paper summarizes the SIGMOD 2005 panel on Databases and Information Retrieval: Rethinking the Great. The goal of the panel was to discuss whether we should rethink data management systems architectures to truly merge Database (DB) and Information Retrieval (IR) technologies.", "Key Idea": "The panel discusses whether we should rethink data management systems architectures to truly merge Database (DB) and Information Retrieval (IR) technologies.", "Method": "The panel was a discussion session held at the SIGMOD 2005 conference with high attendance and lively discussions.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "7d850699-212d-466b-976f-0afed0653fef", "Context": "The authors build on their previous work from HOO 2011 Shared Task to extend it for determiner and preposition error detection in non-native English essays from the Cambridge Learner Corpus FCE Dataset.", "Key Idea": "The authors propose a missing determiner detector and incorporate word clustering to their n-gram prediction approach to improve the determiner and preposition error detection.", "Method": "The authors extend their n-gram-based data-driven prediction approach from HOO 2011 Shared Task to improve determiner and preposition error detection. They used word clustering to group semantically similar words and incorporated it into the prediction approach.", "Outcome": "N/A", "Future Impact": "The proposed approach can be applied to other languages to identify determiner and preposition errors in non-native speaker essays during automated essay scoring and may improve the accuracy of grammatical error correction systems."} +{"id": "7b21425c-a2b7-4d19-b030-a8350b2a7a80", "Context": "Conditional set generation learns a mapping from an input sequence of tokens to a set. Several NLP tasks are instances of set generation.", "Key Idea": "The authors propose a novel algorithm called SETAUG to jointly model the set cardinality and output, endowing any Seq2Seq model with the signals of order-invariance and cardinality.", "Method": "The SETAUG algorithm is designed to augment any Seq2Seq model and enable it with the capabilities of order-invariance and cardinality. The authors show that training a Seq2Seq model with this augmented data resulted in an average relative improvement of 20% on four benchmark datasets.", "Outcome": "The proposed SETAUG algorithm achieved an average relative improvement of 20% on four benchmark datasets. The Seq2Seq models achieved superior performance on entity typing and dialogue emotion tagging.", "Future Impact": "SETAUG is a model-independent data augmentation approach that may improve the performance of various Seq2Seq models across different domains. This technique could be applied to other set generation tasks, enabling improvements in various NLP applications."} +{"id": "6ecf725b-661e-4897-8169-22d71826d0e8", "Context": "Different Open Information Extraction (OIE) tasks require different types of information, so the OIE field requires strong adaptability of OIE algorithms to meet different task requirements.", "Key Idea": "The authors propose a new adaptable and efficient OIE system called OIE@OIA which follows the methodology of Open Information eXpression (OIX) to parse a sentence to an Open Information Annotation (OIA) graph and then adapts the graph to different OIE tasks with simple rules.", "Method": "The authors implement an end-to-end OIA generator by annotating a dataset and designing an efficient learning algorithm for the OIA graph. The authors adapt the proposed system to accomplish three popular OIE tasks and show that it achieves new SOTA performances on these tasks with fewer training samples.", "Outcome": "The proposed OIE@OIA system achieves new state-of-the-art (SOTA) performances on the three popular OIE tasks and requires significantly fewer training samples than other end-to-end OIE baselines.", "Future Impact": "The OIE@OIA system can adapt to different OIE tasks with simple rules, which could lead to the development of more efficient and adaptable OIE algorithms in the future."} +{"id": "816707fd-9214-4435-ac40-b2655e55c9d0", "Context": "Pretrained language models (PLMs) have achieved superhuman performance on many benchmarks", "Key Idea": "The authors introduce CoDA21, a challenging benchmark that measures natural language understanding (NLU) capabilities of PLMs by aligning definitions with appropriate context.", "Method": "CoDA21 requires understanding of contexts and definitions, including complex inference and world knowledge. The authors compare performance of humans and PLMs on CoDA21.", "Outcome": "The authors find a large gap between human and PLM performance on CoDA21, suggesting that it measures an aspect of NLU that is not sufficiently covered in existing benchmarks.", "Future Impact": "CoDA21 could be used as a new benchmark to evaluate the performance of PLMs and other NLP models in terms of natural language understanding. It could also motivate researchers to develop NLP models with better NLU capabilities that exceed human performance."} +{"id": "802a5b78-a022-4d38-bfb3-f28eee4ef89a", "Context": "Neural network-based dependency parsing has attracted much interest due to its advantages in overcoming the data sparsity and the feature engineering problem. However, modeling the complicated syntactic and semantic compositions of the dense neural features is still a challenge.", "Key Idea": "The authors propose two heterogeneous gated recursive neural networks: Tree-GRNN and DAG-GRNN and integrate them to automatically learn the compositions of the dense features for transition-based dependency parsing.", "Method": "The authors use Tree-GRNN to model the feature combinations for the trees in the stack with partial dependency structures. They use DAG-GRNN to model the feature combinations for the nodes whose dependency relations have not been built yet. The authors evaluate the proposed model on two benchmark datasets: PTB3 and CTB5.", "Outcome": "The experiment results show that the proposed model is effective for dependency parsing, achieving state-of-the-art results on the PTB3 dataset and competitive results on the CTB5 dataset.", "Future Impact": "The proposed method can potentially benefit various natural language processing tasks that rely on dependency parsing, such as machine translation, text summarization, and sentiment analysis."} +{"id": "80698baf-89cb-4a50-9f5c-0c74151b798b", "Context": "The paper aims to examine how students in a MOOC can be motivated to do a better job during peer grading.", "Key Idea": "The paper proposes a controlled study to examine how 'grading the graders' (examining the peer-grading work of students) can motivate students to do a better job when grading assignments.", "Method": "The authors conducted a controlled study on over a thousand students on a popular MOOC, asking two questions to differentiate how to motivate students and validate their hypothesis. The quality of the peer grading is compared and examined to test their hypothesis.", "Outcome": "The study found strong statistical evidence that 'grading the graders' can increase the quality of peer grading.", "Future Impact": "The paper provides a potential motivation strategy for instructors to integrate in their peer-grading evaluation process in MOOCs and could encourage further research into how to motivate students in online education."} +{"id": "8737b031-f77c-4f32-8a68-4be9b0c9ecf8", "Context": "Training a deep network to perform semantic segmentation requires large amounts of labeled data. Researchers have investigated the use of synthetic data to reduce the manual effort of annotating real images.", "Key Idea": "The authors propose a new method for using synthetic data that builds on foreground-background segmentation. They treat foreground objects in a detection-based manner and background pixels in a more traditional classification-based manner.", "Method": "The authors perform experiments on Cityscapes and CamVid datasets using models trained only on synthetic data.", "Outcome": "The proposed method achieves state-of-the-art performance on Cityscapes, outperforming other methods that use a mix of synthetic and real data.", "Future Impact": "Our approach provides a drastically different and more efficient way of handling synthetic images, which can lead to more rapid development and deployment of semantic segmentation models by reducing the need for real images. Further research could investigate the use of this method on other types of computer vision tasks beyond semantic segmentation."} +{"id": "8ae36735-e4ac-48da-bd2b-5538a6a00a73", "Context": "Social networking sites can use user-provided traits such as interests and demographics to provide better content to match their usersu0027 interests. Accurate probability estimates are needed to determine the correct content to return.", "Key Idea": "The authors propose a maximum entropy constraint to address the bias introduced by collective inference methods, and a massively scalable variational inference algorithm for large scale relational networks.", "Method": "The authors analyze the effect of full semi-supervised RML and conduct experiments on seven real-world datasets to demonstrate their method's improvement over baselines.", "Outcome": "The proposed method outperforms baselines on seven real-world datasets, including large scale networks with over five million edges.", "Future Impact": "The proposed method provides a framework for accurately predicting user preferences in massive relational network domains, which has numerous applications in recommendation and matching systems in social networking sites and e-commerce platforms."} +{"id": "9292bc08-a70f-4adb-a6d7-920728893a39", "Context": "The paper discusses propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics.", "Key Idea": "The authors show the equivalence between a stronger argumentation system in a full classical language and a system of causal reasoning.", "Method": "The authors discuss the implications of this correspondence between the two systems.", "Outcome": "The paper establishes the equivalence between stronger argumentation system and system of causal reasoning.", "Future Impact": "N/A"} +{"id": "93bf12a1-2174-43e3-9cd4-c2b8aeed2f93", "Context": "Unsupervised domain mapping aims to learn a mapping between two unmatched datasets such that a sample in one can be translated to its analog in another", "Key Idea": "The authors propose a method to learn a one-sided mapping between two unmatched datasets, without having to learn the inverse mapping.", "Method": "The authors present a method for learning the mapping that maintains the distance between a pair of samples using a proposed distance loss function. The proposed method is compared to the existing circularity-based constraint through experiments.", "Outcome": "The proposed method can learn one-sided mapping between unmatched datasets and leads to preferable numerical results over the existing circularity-based constraint.", "Future Impact": "N/A"} +{"id": "8aedb046-2f51-4229-bc19-ea6db98355cb", "Context": "Ranking webpages has been a significant organizational structure of the web, affecting various web applications such as web search and crawling. Researchers previously used a HostGraph with the random walk model to rank websites.", "Key Idea": "The authors proposed a novel method named AggregateRank rooted in the theory of stochastic complement that accurately approximates the sum of PageRank and has lower computational complexity than PageRank.", "Method": "The authors mathematically proved that the probability of visiting a website by the random web surfer should be equal to the sum of the PageRank values of the pages inside that website. The authors proposed AggregateRank, which has lower computation complexity than PageRank and performed an experimental evaluation of AggregateRank in comparison to previous methods.", "Outcome": "AggregateRank is a better method for ranking websites than previous methods. Both theoretical analysis and experimental evaluation support the effectiveness of AggregateRank.", "Future Impact": "N/A"} +{"id": "99bef0ca-923c-480d-a89e-9a2f27e02157", "Context": "Abstraction has emerged as a key enabler for solving large incomplete-information games over the last dozen years.", "Key Idea": "The paper reviews key developments in abstraction, including reasons for abstracting games, practical algorithms for information and action abstraction, recent theoretical breakthroughs on strategy quality, and reverse mapping opponent's actions into the abstraction.", "Method": "The author reviews existing literature and presents various algorithms and theoretical results.", "Outcome": "N/A", "Future Impact": "The paper discusses current and future research in the field of abstraction for solving large incomplete-information games."} +{"id": "9999e1df-f439-4f62-bd03-337e494e9da3", "Context": "The paper discusses a data type encapsulation scheme that enables natural expression of space operations using base language operators.", "Key Idea": "The authors propose a data type encapsulation scheme using a conceptual separation of operators and procedure calls in the base language, which results in a powerful and expressive language.", "Method": "The authors implemented the proposed scheme and provide several examples to demonstrate its effectiveness.", "Outcome": "The paper does not mention any quantitative or qualitative outcomes of the proposed scheme.", "Future Impact": "The proposed scheme provides a way for programmers to achieve data encapsulation and can potentially improve code readability and maintainability. It may inspire other researchers to explore similar approaches."} +{"id": "977698a6-56ef-4e57-94c5-b3a68a451a80", "Context": "Emotion cause extraction is a challenging task compared to emotion classification.", "Key Idea": "The authors propose a new QA-based approach that models context information using a new mechanism to store relevant context in different memory slots.", "Method": "The authors use deep memory networks to frame emotion cause identification as a reading comprehension task in QA and extract both word level sequence and lexical features. They evaluate their proposed method on a recently released dataset and compare it with competitive baselines.", "Outcome": "The proposed method achieves the state-of-the-art performance on a recently released emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.", "Future Impact": "N/A"} +{"id": "9748b061-590a-4602-9015-a68e879ffced", "Context": "HCI researchers are looking for customized liquid-based material printing platforms with desirable features. Conventional printing systems like inkjet or pneumatic syringe-based have limitations in printing a wide range of materials with different processing requirements.", "Key Idea": "The authors propose a design strategy for building and customizing modularized parts of a liquid-based smart material printing platform. The printing platform supports open-source, customizable software", "Method": "The authors followed design principles, building and customizing modularized parts of a liquid-based smart material platform. They demonstrated the system design using three use cases to show the material variability and customizability for different users with different demands.", "Outcome": "The liquid-based smart material printing platform xPrint provides the ability to print a large range of materials, including synthesized polymers and natural micro-organism-living cells, with a printing resolution from 10μm up to 5mm (droplet size).", "Future Impact": "The proposed strategy offers a flexible way for researchers and designers to build a liquid-based material printer optimized to meet their specific material requirements. It can potentially be used to print a wide range of customized materials for various applications."} +{"id": "927df1bd-273a-4088-8c56-2e79cac37072", "Context": "K-means is one of the most popular data processing algorithms, and a proper initialization is crucial for obtaining a good final solution. K-means++ initialization algorithm achieves this secret by obtaining an initial set of centers that is provably close to the optimum solution.", "Key Idea": "The authors propose k-means||, a parallelizable initialization algorithm that drastically reduces the number of passes needed to obtain a good initialization for k-means. The proposed method obtains a nearly optimal solution after a logarithmic number of passes.", "Method": "The authors perform experimental evaluations on real-world large-scale data to demonstrate the superiority of k-means|| over k-means++ in both sequential and parallel settings.", "Outcome": "The proposed initialization algorithm k-means|| outperforms k-means++ in both sequential and parallel settings when evaluated on real-world large-scale data.", "Future Impact": "The proposed initialization algorithm can improve the performance of k-means on massive data by reducing the time and computational resources needed to obtain a good initialization. The methodology of k-means|| can be generalized to other clustering algorithms that could lead to the improvement of various machine learning applications."} +{"id": "9b95b592-1562-4ef9-b0ed-e0655fadc73b", "Context": "The paper presents a general formulation for geodesic distance propagation of surfaces on a 3-manifold. The paper sets up a partial differential equation governing the propagation of surfaces at equal geodesic distance.", "Key Idea": "The paper proposes a geodesic distance evolution scheme for surface matching on a 3-manifold, which generalizes a result of Kimmel et al. [11] and provides a way to compute distance maps on manifolds. Moreover, the propagation equation is generalized to any number of dimensions.", "Method": "The authors use an eulerian formulation with level-sets and develops a stable numerical algorithm based on the extended partial differential equation to compute the geodesic distance propagation.", "Outcome": "The paper proposes a new method for surface matching, and the matching paths are obtained as the orbits of the vector field defined as the sum of two distance map's gradient values. The proposed method works for large deformation and arbitrary topology.", "Future Impact": "The proposed method can find applications in computational geometry, computer graphics, and medical imaging for matching and registration of the surfaces in different domains. It can also pave the way for further research in using geodesic distance propagation for distance maps on manifolds."} +{"id": "9f641e23-3886-4ac6-b65e-28db86ee48be", "Context": "Google unveiled the generalized second price (GSP) auction, against the usual recommendation of using Vickrey-Clarke-Groves (VCG) auction.", "Key Idea": "Advertisers' preferences map to a model called value maximization instead of profit maximization. For value maximizers, GSP is the truthful auction. GSP is an auction whose prices are truthful for value maximizers, which can be applied to arbitrary single-parameter domains.", "Method": "The authors provide a deep justification for the success of GSP. They explain their model on advertisers' preferences and demonstrate its implications.", "Outcome": "GSP is a successful auction model for advertisers' preferences consisting of value maximizers. GSP is a powerful auction with elegant simplicity.", "Future Impact": "N/A"} +{"id": "9cd7e7e1-8893-4db6-8327-48f098187699", "Context": "Twitter provides an important alternative information channel to traditional media during natural disasters, but the amount and diversity of messages poses the challenge of information overload to end users.", "Key Idea": "The authors develop an automatic classifier of tweets to feed a mobile application that reduces the difficulties that citizens face to get relevant information during natural disasters.", "Method": "The authors build and validate a ground truth dataset from the Chilean earthquake of 2010 and present the process to build a classifier that filters tweets relevant and non-relevant to an earthquake. The authors show in detail the effect of class imbalance and dimensionality reduction over 5 classifiers.", "Outcome": "The authors present important considerations at the moment of building these systems and show how the performance of these models is affected by class imbalance and dimensionality reduction.", "Future Impact": "The proposed system could help citizens more easily access relevant information during natural disasters, which could improve preparedness and response. Additionally, the study's findings on dealing with class imbalance and dimensionality reduction could be applied to other classification tasks in the field of natural language processing."} +{"id": "a5bc5ab8-70fb-4dff-824b-7606228e44a9", "Context": "There is a need to embed mathematical expressions in Web pages, but the notations available are difficult to learn.", "Key Idea": "EzMath provides a new notation for embedding mathematical expressions that is easy to learn.", "Method": "N/A", "Outcome": "The outcome of the paper is the proposed notation for embedding mathematical expressions in Web pages.", "Future Impact": "N/A"} +{"id": "a722b600-3725-4738-a47b-435aebd63e13", "Context": "Smart home devices have evolved to include products that monitor, automate, and present themselves as human, but this has raised privacy concerns among users who expect privacy in their homes.", "Key Idea": "The authors propose a design philosophy for intelligent agents in the smart home that prioritize privacy and respect for the user's private space, and offer an alternative to the current way of building smart home devices.", "Method": "The authors apply their design philosophy to develop privacy-empowering technologies for the smart home.", "Outcome": "N/A", "Future Impact": "The proposed design philosophy could influence the future development of smart home devices by prioritizing user privacy and respect for private spaces."} +{"id": "a10dafe9-6093-47f0-8429-7b62c46566ea", "Context": "Enterprise mashup scenarios often involve feeds derived from data created primarily for eye consumption, such as email, news, calendars, blogs, and web feeds. These data sources can test the capabilities of current data mashup products, as the attributes needed to perform join, aggregation, and other operations are often buried within unstructured feed text.", "Key Idea": "The authors propose the integration of SystemT, an information extraction system from IBM Research, with IBM's InfoSphere MashupHub to build domain-specific annotators with SystemT's declarative rule language, AQL, to convert unstructured text into structured information.", "Method": "The authors demonstrate the integration of the SystemT information extraction system with IBM's InfoSphere MashupHub by using domain-specific annotators to combine structured and unstructured information.", "Outcome": "The demo shows how to build domain-specific annotators with SystemT's declarative rule language, AQL, and how to use these annotators to combine structured and unstructured information in an enterprise mashup.", "Future Impact": "N/A"} +{"id": "a1739057-ef00-4b01-9c26-4ab2b5d5708e", "Context": "The increasing use of data mining tools requires a Knowledge Discovery and Data Mining System (KDDMS) that optimizes for sequence of queries and multiple simultaneous queries scenario. This paper focuses on frequent pattern mining queries involving one or multiple datasets.", "Key Idea": "The authors propose a systematic mechanism featuring a knowledgeable cache to optimize the mining queries and algorithms to design an architecture for implementing the mechanism.", "Method": "The authors present a new algorithm for optimizing the system architecture. Experimental evaluation is done with both real and synthetic datasets to show the effectiveness of the proposed mechanism.", "Outcome": "The proposed mechanism achieves a speedup up to a factor of 9 compared with non-caching or non-optimization systems.", "Future Impact": "N/A"} +{"id": "ad69e31c-2c7f-4db0-916c-3deccaab37fd", "Context": "The paper focuses on developing a machine learning approach to Go and related board games, particularly on learning a good evaluation function in a scalable way.", "Key Idea": "The authors propose a system that automatically learns the propensity of local patterns from a library of games, feeds it into a recursive neural network, and produces local outputs that represent local territory ownership probabilities for developing a strategic evaluation function for the game of Go.", "Method": "The authors use a library of local tactical patterns, integrate patterns across the board, and develop a Bayesian network-based recursive neural network for developing the evaluation function. They use datasets of amateur and professional games for training and testing the system.", "Outcome": "The system trained using only 9x9 amateur game data performs surprisingly well on a test set derived from 19x19 professional game data.", "Future Impact": "The system has potential applications in developing better evaluation functions for other board games. Possible directions for further improvement include exploring better patterns of intel and using reinforcement learning techniques in addition to supervised learning."} +{"id": "b0ad60d7-3a3a-42c2-acbc-fba55e708ba0", "Context": "The meaning of a lexeme often varies due to the specific surrounding context in natural language.", "Key Idea": "The authors propose a technique to produce a context-dependent 'meaning' representation for a lexeme in a specific surrounding context by representing the 'meaning' of a lexeme in a specific context through a list of semantically replaceable elements composed of other lexemes from the experimental lexicon.", "Method": "The authors perform experiments with two lexicons, English words, and individual words and phrases. The resulting lists can be used to compare the 'meaning' of conceptual units and can also serve as features for machine learning approaches to classify semantic roles and relationships.", "Outcome": "The authors successfully developed a general new technique to produce a context-dependent 'meaning' representation for a lexeme in a specific surrounding context. The resulting lists can be used to compare the 'meaning' of conceptual units in different contexts and can also be used as features for machine learning approaches.", "Future Impact": "The technique proposed by the authors may lead to more effective natural language processing tools that can accurately capture the contextual meaning of words, phrases, and sentences in various domains."} +{"id": "adfd8058-64b3-4062-953c-034b732e2fa0", "Context": "The paper presents a video coding scheme using texture synthesis, where P and B-frames are parametrically coded with the help of DEMD algorithm and I-frames are coded with the help of H.264.", "Key Idea": "The paper proposes the use of Directional Empirical Mode Decomposition (DEMD) algorithm for texture synthesis of P and B-frames, and wavelet decomposition of residual image to perform multi-resolution parametric modeling of image data for video compression.", "Method": "The authors perform experiments to demonstrate the effectiveness of the proposed algorithm, where subsequent IMF images are synthesized with the help of correlation search and wavelet decomposition over residual image.", "Outcome": "The proposed algorithm achieves significant video compression with acceptable quality and enables scalable coding of IMF parameters.", "Future Impact": "The proposed algorithm can be used to improve video compression and enable high scalability with improved perceptual quality in real-time video streaming applications like live video streaming, video conferencing, and IPTV."} +{"id": "bd84a1fd-ee85-4ac9-a6ae-e534b6013506", "Context": "The paper considers convolutional networks from a reproducing kernel Hilbert space viewpoint.", "Key Idea": "The authors establish harmonic decompositions of convolutional networks, which are expansions into sums of elementary functions of increasing order related to the spherical harmonics.", "Method": "The authors use harmonic analysis techniques to obtain characterizations of the integral operators associated with convolutional networks and statistical bounds for them.", "Outcome": "N/A", "Future Impact": "The harmonic decomposition framework could provide a new perspective on the representation power and expressiveness of convolutional networks, and potentially lead to new insights and improvements in their design and optimization."} +{"id": "b391a193-83e3-4f11-801f-1842647d626e", "Context": "Previous studies have employed graph-based deep learning methods for action unit detection. However, dependencies among action units in real-world data are often noisy.", "Key Idea": "The authors propose an uncertain graph neural network (UGN) in which the probabilistic mask captures individual dependencies among action units and uncertainties.", "Method": "The authors propose an adaptive weighted loss function and perform experiments on two benchmark datasets, BP4D and DISFA.", "Outcome": "The proposed UGN model achieves the state-of-the-art performance for AU detection on BP4D and DISFA dataset.", "Future Impact": "Using an uncertain graph neural network with an adaptive weighted loss function could lead to better results for other tasks with noisy dependencies and data imbalances."} +{"id": "b14f2bc1-607f-4d8b-a731-b4afdf30a633", "Context": "Current explanation datasets often employ synthetic data with simple reasoning structures, which cannot express more complex reasoning processes, such as the rebuttal to a reasoning step and the degree of certainty of the evidence.", "Key Idea": "The authors propose a comprehensive logical reasoning explanation form that includes three main components: the condition of rebuttal, logical formulae, and reasoning strength indicated by degrees of certainty.", "Method": "The authors propose a comprehensive benchmark, and evaluate the performance of the current state-of-the-art models on this new explanation form.", "Outcome": "The experimental results show that generating reasoning graphs remains a challenging task for current models even with the help of giant pre-trained language models.", "Future Impact": "The proposed benchmark can be used as a standardized tool to evaluate logical reasoning capabilities. The benchmark can also drive future research in developing models that can handle more complex real-life scenarios."} +{"id": "b6b29c8a-7c8c-444c-b434-2ff9e166d9aa", "Context": "Offline Reinforcement Learning (RL) aims to learn policies from previously collected datasets without exploring the environment. Directly applying off-policy algorithms to offline RL usually fails due to the extrapolation error caused by the out-of-distribution (OOD) actions.", "Key Idea": "The authors propose a purely uncertainty-driven offline algorithm called Pessimistic Bootstrapping for offline RL (PBRL) that conducts uncertainty quantification via the disagreement of bootstrapped Q-values and performs pessimistic updates by penalizing the value function based on the estimated uncertainty to tackle extrapolation error.", "Method": "The authors propose a novel OOD sampling method. They show that such OOD sampling and pessimistic bootstrapping yields provable uncertainty quantifier in linear MDPs, thus providing the theoretical underpinning for PBRL. The authors conduct experiments on D4RL benchmark and compare the performance of PBRL with the state-of-the-art algorithms.", "Outcome": "The experiments on D4RL benchmark show that PBRL has better performance compared to the state-of-the-art algorithms.", "Future Impact": "The proposed PBRL algorithm could be applied to various problems such as robotics, self-driving cars, and others, where collecting new data is challenging. Further exploration could be done in using PBRL algorithm on continuous control and real-world problems."} +{"id": "c268a190-6974-4190-8f48-db5dcbda8bc8", "Context": "The challenge of data variety in multiple formats like relational and semi-structured data has led to the constraint of traditional databases in handling data in different formats.", "Key Idea": "The authors propose a multi-model processing framework for joining relational and semi-structured data and a worst-case optimal join algorithm that guarantees intermediate results no larger than the worst-case join result.", "Method": "The authors design a worst-case optimal join algorithm and compare it with baseline join methods using preliminary results.", "Outcome": "The multi-model algorithm proposed by the authors outperforms the baseline join methods in terms of running time and intermediate result size.", "Future Impact": "The proposed multi-model processing framework and worst-case optimal join algorithm could potentially lead to faster join processing and improved handling of different formats of data, solving the challenges posed by the data variety in the modern data management ecosystem."} +{"id": "c5c22c6f-e9e9-4b0a-ac57-d3baa100033b", "Context": "The paper deals with structured stochastic convex optimization problems with a large number of linear constraints, which is a common problem in SDP-relaxations of combinatorial problems.", "Key Idea": "The authors propose two novel conditional gradient-based methods that process only a subset of the constraints at each iteration to gain a computational advantage over prior works that require full passes.", "Method": "The proposed algorithms rely on variance reduction and smoothing used in conjunction with conditional gradient steps. Preliminary numerical experiments are provided for illustrating the practical performance of the methods.", "Outcome": "The authors provide rigorous convergence guarantees for their proposed algorithms and demonstrate their practical performance through preliminary numerical experiments.", "Future Impact": "Further research can investigate the suitability of these methods for other types of optimization problems and improve their efficiency and scalability."} +{"id": "ca53b2c4-2912-4515-aae6-938c3f268a60", "Context": "The authors address the problem of detecting epidemic tendency by mining search logs.", "Key Idea": "The authors propose an algorithm to select epidemic related queries/terms by using click-through information. They adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms in search logs.", "Method": "The authors use click-through information to select epidemic-related terms, followed by linear regression to model epidemic occurrences and the frequencies of epidemic-related terms in search logs. They compare the performance of combining different epidemic-related terms against using a single epidemic-related term.", "Outcome": "The proposed algorithm is effective in identifying ERTs that show a high correlation with epidemic occurrences. The authors show that combining different ERTs performs better than using a single ERT.", "Future Impact": "N/A"} +{"id": "c676aecf-7468-4258-bb41-22bc1811bc3a", "Context": "Community search is a fundamental problem aiming to find a connected subgraph containing the given query nodes in a social network. Most of the existing community search models only focus on the internal cohesiveness of a community. However, a high-quality community often has high modularity, making dense connections inside communities and sparse connections to the nodes outside the community.", "Key Idea": "The authors present Density Modularity based Community Search (DMCS) for searching communities which have high modularity and contain all the query nodes.", "Method": "The authors design a new graph modularity function named Density Modularity. To efficiently address DMCS, they present new algorithms that run in log-linear time to the graph size. The authors conduct extensive experimental studies in real-world and synthetic networks.", "Outcome": "The proposed algorithm achieves up to 8.5 times higher accuracy in terms of NMI than the baseline algorithms in the conducted experiments.", "Future Impact": "N/A"} +{"id": "c37e0f9f-1654-4f1f-b812-bf6f67c0c840", "Context": "Recent research has shown the susceptibility of deep networks to adversarial attacks. Developing defenses against such attacks is an active research area, with some approaches proposing robust models that are immune to such adversaries, while other techniques attempt to detect such adversarial inputs.", "Key Idea": "The authors present a novel statistical approach for adversarial detection in image classification. The approach is based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class.", "Method": "The authors make use of various statistical distances such as ED (Energy Distance) and MMD (Maximum Mean Discrepancy) for adversarial detection, and analyze the performance of each metric. They experimentally show that their approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size, and degree of adversarial perturbation.", "Outcome": "The authors show that their proposed approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size, and degree of adversarial perturbation.", "Future Impact": "The proposed method may lead to better defense mechanisms to combat adversarial attacks on deep learning-based AI systems, especially in cases where developing robust models immune to adversarial inputs is difficult."} +{"id": "c3aa9543-0695-42b4-99af-d033e3912801", "Context": "The real estate industry is a largely untapped area by the KDD community despite its enormous size and prominence. The lack of synergy is due to the industry's lack of appreciation for data science methods and the Data Science community's unawareness of challenging real estate problems suited to its methods.", "Key Idea": "The paper provides an introduction to real estate for data scientists and outlines a spectrum of data science problems that are suited to data science methods, which can be utilized in real estate to provide solutions to these problems.", "Method": "The paper presents concrete examples from three companies: Airbnb, a short-term rental marketplace, Cherre, a real estate data integration platform, and Compass, the largest independent real estate brokerage in the U.S.", "Outcome": "N/A", "Future Impact": "The paper opens up the possibility of utilizing Data Science methods in the real estate industry, leading to more sophisticated and efficient methods in the industry, improved market analysis resulting in lower risk and more reliable forecasts, among others."} +{"id": "ce74316d-c5dc-47f1-b0c4-0591bc3fb4b6", "Context": "Several methods have been proposed to evaluate queries over a native XML DBMS, where the queries specify both path and keyword constraints. These methods include graph traversal approaches optimized with auxiliary structures known as structure indexes and approaches based on information-retrieval style inverted lists.", "Key Idea": "The authors propose a strategy that integrates both structure indexes and inverted lists to evaluate branching path expression queries in a wide range of methods.", "Method": "The authors propose an algorithm for branching path expressions based on the integration of structure indexes and inverted lists. The algorithm is evaluated over the Niagara XML DBMS.", "Outcome": "The experiments demonstrate the benefits of integrating structure indexes and inverted lists. Additionally, the authors integrate these techniques with the Threshold Algorithm proposed by Fagin et al. to obtain instance optimal algorithms to push down top k computation.", "Future Impact": "N/A"} +{"id": "dca09f6f-b63a-42a1-9eb5-fbfa45bc6389", "Context": "Activity logs collected from wearable devices are challenging to model due to their heterogeneous nature, complex interdependencies, and diverse scale and resolution.", "Key Idea": "The authors propose FitRec, an LSTM-based model that captures two levels of context information, context within a specific activity, and context across a user's activity history for personalized fitness recommendations.", "Method": "The authors develop context-aware sequential models to capture the personalized and temporal patterns of the fitness data. The authors propose, and evaluate the model on, a novel dataset containing over 250 thousand workout records coupled with hundreds of millions of parallel sensor measurements (e.g., heart rate, GPS), and metadata.", "Outcome": "FitRec is able to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise and outperforms baselines on several personalized recommendation tasks.", "Future Impact": "The proposed context-aware approach paves the way for utilizing wearables data for personalized exercise scheduling, workout recommendation, and heart rate anomaly detection."} +{"id": "db900c02-9a35-4a83-aa19-15b763259100", "Context": "Deep learning and Convolutional Neural Networks (CNN) have become essential in image recognition and signal processing applications. Quantum computing is a powerful computing paradigm that has applications in several fields, including machine learning.", "Key Idea": "The authors propose a quantum algorithm for evaluating and training deep convolutional neural networks called the Quantum CNN (QCNN), which can reproduce the outputs of the classical CNN and allows non-linearities and pooling operations.", "Method": "The authors present numerical simulations for the classification of the MNIST dataset to provide practical evidence for the efficiency of the QCNN.", "Outcome": "The proposed QCNN algorithm shows potential speedups over classical CNNs for evaluation and training, and can be used for larger kernels, high-dimensional inputs, and high-depth input channels. The numerical simulation presented shows evidence for the efficiency of the proposed algorithm.", "Future Impact": "The proposed QCNN algorithm could allow new frontiers in the image recognition domain, and may have practical implications for using quantum computing to speed up deep learning tasks."} +{"id": "ddf8f49c-342a-4cd0-8b3b-b588af08ed0d", "Context": "Soboroff, Nicholas and Cahan proposed an approach for evaluating retrieval systems even when relevance judgments are absent. They demonstrated that their methodology is correlated with actual evaluations.", "Key Idea": "The authors propose assessing the similarity of retrieved results to quantify similarity among retrieval systems in the absence of relevance judgments.", "Method": "The authors evaluate retrieval systems based on the similarity of their retrieved results. They compare the effectiveness of their method to the methodology proposed by Soboroff et~al. and demonstrate highly correlated results.", "Outcome": "The proposed methodology allows for evaluation of retrieval systems in absence of relevance judgments with reasonable accuracy. The methodology effectively evaluates and ranks retrieval systems by popularity rather than performance.", "Future Impact": "N/A"} +{"id": "ccae9338-7379-4af1-8fea-8945ba429c5c", "Context": "Object tracking is a challenging problem in computer vision. Deep learning is increasingly being used to improve tracking accuracy by extracting powerful features. Pairwise loss is the commonly used loss function in Siamese network for object tracking.", "Key Idea": "The authors propose a novel triplet loss that is added to the Siamese network framework to extract expressive deep features for object tracking. Triplet loss utilizes more elements for training to achieve a more powerful feature via the combination of original samples.", "Method": "The authors perform experiments on three real-time trackers based on the Siamese network and evaluate the proposed triplet loss on several popular tracking benchmarks.", "Outcome": "The proposed triplet loss outperforms the baseline trackers in terms of tracking accuracy while maintaining comparable accuracy with recent state-of-the-art real-time trackers. The proposed variants operate at almost the same frame-rate as the baseline trackers.", "Future Impact": "The proposed triplet loss could be applied to other tasks that require feature extraction from Siamese networks. Future work could explore other ways to combine different forms of losses to generate better features while maintaining good performance."} +{"id": "dd282632-ee41-45da-add8-d68d89c57e2d", "Context": "Classification activation map (CAM), utilizing the classification structure to generate pixel-wise localization maps, is a crucial mechanism for weakly supervised object localization (WSOL). However, CAM prefers to discern global discriminative factors rather than regional object cues, making it challenging to locate objects based on regional features.", "Key Idea": "The authors propose a plug-and-play mechanism called BagCAMs to improve the regional localization in WSOL without refining or re-training the baseline structure. BagCAMs adopts a proposed regional localizer generation (RLG) strategy to define a set of regional localizers to discern region-wise object factors for localization tasks. The regional localizers are effectively weighted to form the final localization map.", "Method": "The authors adopt BagCAMs to improve the performance of baseline WSOL methods and compare it with other state-of-the-art models on three WSOL benchmarks.", "Outcome": "The proposed BagCAMs mechanism improves the performance of baseline WSOL methods to achieve state-of-the-art results on three WSOL benchmarks.", "Future Impact": "N/A"} +{"id": "dfba01c5-a632-4394-8607-9a32f20e526c", "Context": "The paper presents a system that automatically builds a 3D scene model containing geometric information and photometric information under various illumination conditions, from real images.", "Key Idea": "To extract photometric information, the paper computes a low-dimensional linear space of the spatio-illumination volume, represented by a set of basis images.", "Method": "The authors use structure-from-motion and correlation-based stereo techniques to match pixels between images of different viewpoints and to reconstruct the scene in 3D space. The photometric property is extracted from images taken under different illumination conditions.", "Outcome": "The model can be used to create realistic renderings from different viewpoints and illumination conditions. Applications include object recognition, virtual reality, and product advertisement.", "Future Impact": "The approach presented in this paper can potentially be applied to other domains, and can improve object recognition and product advertisement industries."} +{"id": "ded9a095-d94d-4ccd-8825-283ec4bb7093", "Context": "There is increasing interest in adopting UX within corporate environments, and what competencies translate to effective UX design. This paper addresses the gap between pedagogy and UX practice through the lens of competence.", "Key Idea": "The paper explores how students perceive their own competence in UX design and how it shifts over time as they begin internships and full-time positions in UX. A co-construction of identity between designer and environment is proposed as a factor affecting competence.", "Method": "The authors conducted a 12-week longitudinal study of surveys and interviews with students and early professionals in UX and collected data on their competence level and factors influencing it.", "Outcome": "The study proposes a co-construction of identity between designer and environment as a factor affecting competence in UX over time. The paper emphasizes the importance of considering this relationship and various factors, including tool and representational knowledge, complexity, and corporate culture, when assessing competency in UX design.", "Future Impact": "The paper suggests opportunities for future research in building an understanding of competency in UX design based on the proposed co-construction of identity and considering various factors that influence it."} +{"id": "e1a191db-6463-4be9-883d-dfc7fb05c5f4", "Context": "The paper presents a novel dependency parsing method that enforces two structural properties on dependency trees to better represent the set of admissible dependency structures in treebanks.", "Key Idea": "The authors cast the problem of dependency parsing with bounded block degree and well-nestedness as an integer linear program that is solved using Lagrangian Relaxation. They derive a heuristic and an exact method based on a Branch-and-Bound search.", "Method": "The authors solve the Integer Linear Program using Lagrangian Relaxation from which they derive a heuristic and an exact method based on a Branch-and-Bound search. They compare their methods to a baseline unconstrained parser.", "Outcome": "Experimentally, the authors show that their methods are efficient and competitive compared to the baseline unconstrained parser while enforcing structural properties in all cases.", "Future Impact": "The proposed approach can better represent the set of admissible dependency structures in treebanks and connect dependency parsing to context-sensitive grammatical formalisms."} +{"id": "e58b9947-7a3a-414d-a0e8-d6cf02ed7127", "Context": "With recent advances in radio-frequency identification (RFID), wireless sensor networks, and Web-based services, physical things are becoming an integral part of the emerging ubiquitous Web.", "Key Idea": "The paper proposes a unified probabilistic based framework that fuses information across relationships between users and things to make more accurate recommendations, addressing the things recommendation problem in IoT. The framework exploits the merits of social relationships and thing-thing correlations, inheriting the matrix factorization's advantages.", "Method": "The proposed approach is validated based on an Internet of Things platform, and its effectiveness is demonstrated through experimental results.", "Outcome": "The experimental results demonstrate the feasibility and effectiveness of the proposed approach.", "Future Impact": "N/A"} +{"id": "e2466d98-98f9-41a8-8a38-e862adc2ed47", "Context": "This paper explores the importance of IT Governance models for public organizations, as well as serves as a basis for further research in IT Governance adoption models.", "Key Idea": "The authors present an IT Governance model that can be easily adopted by both practitioners and researchers. ", "Method": "The authors perform a literature review in IT Governance, as well as an empirical survey using a questionnaire based on COBIT 4.1 maturity model used to investigate IT Governance practice in multiple case studies from the Kingdom of Bahrain.", "Outcome": "The research will enable public sector organizations to adopt an IT Governance model in a simple and dynamic manner, which provides a clear focus for decision-making attention. The model provides a basis for further research in IT Governance adoption models.", "Future Impact": "The proposed IT Governance model can be adopted by public sector organizations in the future, which may improve IT Governance processes."} +{"id": "e827ee51-aafd-4f3b-99ca-60a5e289a555", "Context": "The paper addresses the PASCAL 2005 Recognizing Textual Entailment Challenge.", "Key Idea": "The authors aim to test the practicability of a purely logical approach for recognizing textual entailment. The authors extract atomic propositions from the text and the entailment hypothesis and express it in a custom logical notation.", "Method": "The authors use Link Parser to extract and encode the propositions. They detect independent entailment relations using Otter and WordNet.", "Outcome": "The proposed logical approach is successful in recognizing the entailment relation between the input text and hypothesis.", "Future Impact": "N/A"} +{"id": "e10b883a-ded6-4b4e-9934-6daba2d2f2b2", "Context": "Existing semantic hashing methods generate binary codes for documents based on similarity in a keyword feature space, which does not fully reflect semantic relationships that go beyond keyword matching. Tag information is often associated with documents in many real-world applications, but it has not been fully exploited yet.", "Key Idea": "The authors propose a new hashing approach called Semantic Hashing using Tags and Topic Modeling (SHTTM) to address the limitations in existing methods by incorporating tag information and similarity information from probabilistic topic modeling.", "Method": "The authors propose a unified framework that employs a formal latent factor model to ensure that hashing codes are consistent with tag information and preserve document topic/semantic similarity. An iterative coordinate descent algorithm is used for learning the optimal hashing codes. The authors conduct extensive empirical studies on four different datasets to demonstrate the advantages of the proposed SHTTM approach against several other state-of-the-art semantic hashing techniques.", "Outcome": "Experimental results indicate that the proposed SHTTM approach outperforms several other state-of-the-art semantic hashing methods on four different datasets. Furthermore, the combination of utilizing tag information and topic modeling in the unified framework obtains even better results.", "Future Impact": "The proposed SHTTM approach has potential applications in large-scale similarity search in various fields such as information retrieval and computer vision."} +{"id": "f23bdd28-b2d6-4a42-a56c-c9774f6451b5", "Context": "With the development of hypersonic vehicles in near space such as X-51A, HTV-2 and so on, tracking for them is becoming a new task and hotspot.", "Key Idea": "The paper introduces a learning tracking algorithm for hypersonic targets using the Sine model and IMM algorithm that learns the target tracking error characteristics to adjust the sampling rate adaptively.", "Method": "The proposed algorithm is compared with single accurate model algorithm and general IMM algorithms with fixed sampling rate through simulation experiments.", "Outcome": "Through the simulation experiments, it is proved that the proposed algorithm can effectively improve the tracking accuracy.", "Future Impact": "N/A"} +{"id": "ea414544-c89c-4039-8227-23b11e9a1239", "Context": "The paper presents a method to identify single-snippet answers to definition questions in question answering systems that supplement web search engines.", "Key Idea": "The authors propose a practically unsupervised learning method that leverages on-line encyclopedias and dictionaries to generate positive and negative definition examples, which are then used to train an SVM.", "Method": "The authors generate positive and negative definition examples using on-line encyclopedias and dictionaries. These examples are used to train an SVM to separate the two classes. The trained SVM is used to identify single-snippet answers to definition questions.", "Outcome": "The proposed method is experimentally shown to be viable, outperforming the alternative of training the system on questions and news articles from TREC. The proposed method helps the search engine handle definition questions significantly better.", "Future Impact": "N/A"} +{"id": "eb15ebe7-aa58-4a98-8f9e-939967c6359f", "Context": "The paper addresses the synchronization problem on multi-graphs, that are graphs with more than one edge connecting the same pair of nodes, where edges are labelled with the ratio of the incident vertices, and labels belong to a group.", "Key Idea": "The authors propose MultiSynch, a principled constrained eigenvalue optimization algorithm that can deal with any linear group, to address the multi-graph synchronization problem.", "Method": "The proposed MultiSynch algorithm is presented based on a constrained eigenvalue optimization technique. The authors evaluate the algorithm on both synthetic and real datasets.", "Outcome": "The authors prove empirically that the averaging approach is less precise and accurate than the proposed algorithm. MultiSynch shows better performance both on synthetic and real datasets.", "Future Impact": "N/A"} +{"id": "f41bfd75-a6b0-44d7-a842-b117797ebd7e", "Context": "The paper focuses on how to generate features from various data representations for efficient answer extraction. The authors mainly discuss feature generation in parse trees.", "Key Idea": "The authors propose and compare three methods to represent syntactic features in Support Vector Machines: feature vector, string kernel, and tree kernel.", "Method": "The authors conduct experiments on the TREC question answering task to compare the proposed feature generation methods. The contribution of individual features is also presented.", "Outcome": "The features generated from the more structured data representations significantly improve the performance based on the surface text features. However, the specific outcomes of the experiments are not provided.", "Future Impact": "N/A"} +{"id": "f2c9f8c3-f9d1-4cae-b7ae-a919ada1daaf", "Context": "Access to online visual search engines implies sharing of private user content -- the query images. Existing retrieval systems may not sufficiently protect the privacy of user images.", "Key Idea": "The authors introduce the concept of targeted mismatch attack for deep learning based retrieval systems to generate an adversarial image to conceal the query image.", "Method": "The authors design various loss functions for the adversarial image construction, including loss functions for unknown global pooling operation or input resolution by the retrieval system. They evaluate the attacks on standard retrieval benchmarks and compare the retrieved results with the original and adversarial image.", "Outcome": "The authors successfully generate adversarial images with very different content but lead to similar retrieval results on standard benchmarks. They also demonstrate successful attacks on partially unknown systems.", "Future Impact": "This paper highlights the privacy implications of online visual search engines and could lead to the development of more robust retrieval systems to better protect user privacy."} +{"id": "f3cef657-1887-42a7-9e11-8fdf54f8fa90", "Context": "This paper introduces kernel continual learning, a variant of continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting.", "Key Idea": "Kernel continual learning is introduced combining kernel methods with continual learning by episodically learning task-specific classifiers based on kernel ridge regression and by formulating kernel continual learning as a variational inference problem.", "Method": "Kernel continual learning is formulated as a variational inference problem to learn a data-driven kernel for each task by incorporating a random Fourier basis as the latent variable. The variational posterior distribution over the random Fourier basis is inferred from the coreset of each task. The authors perform extensive evaluation on four benchmarks.", "Outcome": "Experiments show the effectiveness and promise of kernels for continual learning as kernel continual learning systematically avoids task interference in the classifiers and achieves more efficient continual learning based on episodic memory.", "Future Impact": "Kernel continual learning is a promising direction for addressing catastrophic forgetting in continual learning and improving model efficiency. Further research may focus on evaluating this approach against more complex problems and exploring its limitations."} +{"id": "f32e53d5-c7f4-407e-a5cb-26fef230b5fd", "Context": "Matrix factorization is widely used for recommendation by learning latent embeddings of users and items from user-item interaction data. However, previous methods assume the learned embeddings are static or homogeneously evolving with the same diffusion rate.", "Key Idea": "The authors propose a dynamic matrix factorization model, Dynamic Bayesian Logistic Matrix Factorization (DBLMF), to learn heterogeneous user and item embeddings that are drifting with inconsistent diffusion rates.", "Method": "The authors extend logistic matrix factorization to model the probability a user would interact with an item at a given timestamp and a diffusion process to connect latent embeddings over time. An efficient Bayesian inference algorithm has also been proposed to make DBLMF scalable on large datasets.", "Outcome": "The proposed method has been extensively experimented on real datasets and is compared to the state-of-the-art methods, demonstrating its effectiveness.", "Future Impact": "DBLMF proposes a novel way to model drifting embeddings in recommendation systems and has been demonstrated to be effective in real datasets, signaling its potential to improve the quality of recommendations in various scenarios."} +{"id": "f32d6bc3-d75e-4e84-8bfa-3c83578281dc", "Context": "Transaction log analysis allows examination of both user commands and system responses when conducting an online information search. The objective is to monitor the extent to which systems are used and determine the actual user patterns when conducting an information search.", "Key Idea": "The authors propose transaction log analysis as a powerful methodology to discover how systems are used and assist in the improvement of existing and future systems.", "Method": "Machine-readable transaction log tapes from online catalogs are obtained and analyzed mathematically. This involves parsing user sessions, utilizing Markov chain analysis, and developing state transition probability matrices to illustrate the probability of proceeding from one state to another.", "Outcome": "The methodology allows for the monitoring and evaluation of information systems and can assist in system design. Patron use and system response patterns from several online public catalogs have been obtained by transaction log tapes.", "Future Impact": "The predictive power of the methodology may allow real-time aids to be developed for monitoring and evaluation of information systems."} +{"id": "f82f2e26-5437-4397-b781-50668ce5851b", "Context": "Alignment is trained in a previous stage as the translation model in Statistical Machine Translation (SMT) systems. Alignment model parameters are not tuned in function of the translation task but only indirectly.", "Key Idea": "The authors propose a novel framework for discriminative training of alignment models with automated translation metrics as maximization criterion, optimizing alignments for the translation task, and not needing link labels at the word level.", "Method": "The authors evaluate their framework in terms of automatic translation evaluation metrics to observe an improvement in translation quality.", "Outcome": "Evaluation of the proposed framework shows an observed improvement in translation quality.", "Future Impact": "The proposed framework may be applied to other machine translation models, leading to a more direct and optimized approach to achieve high-quality translations without the need for additional labeled data."} +{"id": "fb538ce2-abf4-4bd8-b35c-1bfe3ab9f48e", "Context": "The complex word identification task refers to the process of identifying difficult words in a sentence from the perspective of readers belonging to a specific target audience. This task has immense importance in the field of lexical simplification.", "Key Idea": "The authors propose two systems for identifying complex words using various lexical and semantic features.", "Method": "The authors developed one system using Naive Bayes and another based on Random Forest Classifiers. They incorporated rule-based processing techniques to improve the results.", "Outcome": "The Naive Bayes classifier based system achieves the maximum G-score of 76.7% for identifying complex words in SemEval-2016 shared task 11 after incorporating rule based post-processing techniques.", "Future Impact": "N/A"} +{"id": "f4f1bbaf-c1a2-44d5-8305-27235fa69d62", "Context": "There has been a lot of interest in the analysis of (social) networks, and with the growing complexity of heterogeneous data, feature-rich networks have emerged as a powerful modeling approach.", "Key Idea": "The tutorial provides a unified perspective on feature-rich networks, focusing on different modeling approaches, in particular multiplex and attributed networks, to devise novel algorithms and tools for the analysis of such networks.", "Method": "The tutorial outlines important principles, methods, and tools for the analysis of feature-rich networks, and provides future research directions in this emerging field.", "Outcome": "N/A", "Future Impact": "The tutorial could lead to the development of more effective methods and tools that enable better analysis of feature-rich networks, which could have implications for data science, data mining, web mining, and web science."} +{"id": "ffd14676-a525-479f-a74e-2c5d3a85c510", "Context": "There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units. This paper considers the difficulties involved in representing shapes in parallel systems.", "Key Idea": "The authors suggest ways of representing shapes in parallel systems which provides a mechanism for shape perception and visual attention, and allows a novel interpretation of the Gestalt slogan.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The proposed mechanism for shape perception and visual attention can be extended to other domain-specific problems involving the representation of complex shapes such as in medical imaging, robotics, and autonomous driving."} +{"id": "f64fdfde-7e93-411b-865a-1e29d71c95b2", "Context": "Large-scale topic models are basic tools for feature extraction and dimensionality reduction in many applications. Hierarchical topic models (HTMs) can learn topics of various levels of abstraction leading to better generalization compared to their flat counterparts. Scalable systems for flat topic models cannot handle HTM due to their complicated data structures and susceptibility to local optima.", "Key Idea": "The authors study the hierarchical latent Dirichlet allocation (hLDA) model, propose an efficient partially collapsed Gibbs sampling algorithm for hLDA and an initialization strategy to deal with local optima introduced by tree-structured models. The authors also propose efficient data layout for vectorizing HTM as well as distributed data structures to address system challenges in building scalable systems for HTMs.", "Method": "The authors perform empirical studies to show the efficiency of their proposed system which is 87 times more efficient than previous open-source implementation for hLDA. The authors demonstrate their scalability on a 131-million-document corpus with 28 billion tokens. They extract 1,722 topics from the corpus with 50 machines in just 7 hours. ", "Outcome": "The proposed system is significantly more efficient than the previous open-source implementation for hLDA. The system can scale to thousands of CPU cores and can process a corpus 4-5 orders of magnitude larger than previously used. ", "Future Impact": "The proposed system can be used in practical applications such as feature extraction and dimensionality reduction for large datasets. The system can be further developed to address system challenges for other HTMs and for other statistical models beyond topic models."} +{"id": "feb75e1f-7838-48ca-9a78-cc31b717e5bf", "Context": "The paper discusses the problem of finding fraudulent content on a crowd-sourced review site due to businesses adding themselves to multiple websites to more easily be discovered. The existing work uses supervised machine learning and focuses only on textual and stylometry features.", "Key Idea": "The authors propose OneReview, a system for finding fraudulent content on a crowd-sourced review site, using correlations with other independent review sites and focusing on anomalous changes in a business’s reputation across multiple review sites.", "Method": "The authors utilize Change Point Analysis method on the reviews of every business independently on every website, and then use the proposed Change Point Analyzer to evaluate change-points, detect suspicious reviews, and identify them as fraudulent. They used crowd-labeling and k-cross validation to obtain 97% accuracy on their ground truth.", "Outcome": "The proposed model classified 61,983 Yelp reviews, about 8% of all reviews, as fraudulent. It identified 3,980 businesses with fraudulent reviews, as well as, 14,910 suspected spam, where at least 40% of their reviews are classified as fraudulent.", "Future Impact": "OneReview could be used to tackle fraudulent reviews on other platforms or to analyze the change in a business's reputation across different time periods or geographical regions."} +{"id": "fed7302a-43a7-412e-8ace-d07905e38c3c", "Context": "One-class collaborative filtering is difficult due to the challenge of interpreting and modeling the latent signal from the missing class. ", "Key Idea": "The authors propose a Bayesian generative model approach for implicit collaborative filtering that treats the latent signal as an unobserved random graph connecting users with items they might have encountered.", "Method": "The authors use stochastic gradient descent and mean field variational inference over random graph samples for large-scale distributed learning, and compare the proposed model to a state-of-the-art baseline on real-world data.", "Outcome": "The proposed model achieves better performance than the state-of-the-art baseline on real-world data.", "Future Impact": "The proposed model can be used to improve one-class collaborative filtering in real-world systems such as the Xbox Live architecture. Future work could explore applying this model to other recommendation systems."} +{"id": "9821d5f7-72b0-4841-a54f-d2af4a04ea3a", "Context": "The paper addresses the problem of inverse rendering where estimating the spherical harmonic illumination coefficients and texture parameters in a specular invariant colour subspace is challenging.", "Key Idea": "The authors propose a novel approach for inverse rendering based on a linear basis approximation of surface texture, which can account for non-Lambertian specular reflectance and complex illumination of the same light source colour.", "Method": "The proposed approach involves fitting a 3D morphable model to a single colour image of faces through the solution of bilinear equations in a specular invariant colour subspace. This approach recovers statistical texture model parameters without relying on computationally expensive analysis-by-synthesis techniques.", "Outcome": "The proposed approach recovers texture model parameters with an accuracy comparable to that of more computationally expensive methods, while requiring only the solution of convex optimization problems.", "Future Impact": "The proposed approach could be extended to other objects besides faces, potentially offering a more efficient and accurate solution to the problem of inverse rendering."} diff --git a/data/gpt4_0531.jsonl b/data/gpt4_0531.jsonl new file mode 100644 index 0000000..f884a00 --- /dev/null +++ b/data/gpt4_0531.jsonl @@ -0,0 +1,120 @@ +{"id": "01f161fe-dd40-45dd-89bd-fb1562771d73", "Context": "Capturing contextual information within event forecasting is challenging due to uncertainty of context structure, high dimensional features, and adaptation of features over time. Graph representations have been successful in various applications, but their use in modeling the dynamic context of social events remains unexplored.", "Key Idea": "The authors propose a novel graph convolutional network for predicting future events, such as civil unrest movements, by identifying dynamic properties of event contexts as social indicators.", "Method": "The proposed model involves extracting and learning graph representations from historical or prior event documents to predict future events and identifying sequences of dynamic graphs as event context.", "Outcome": "Experimental results on multiple real-world data sets demonstrate that the proposed method is competitive against various state-of-the-art methods for social event prediction.", "Future Impact": "N/A"} +{"id": "10c15fe5-c315-4b6d-8910-e6bc3279c817", "Context": "Social events provide insights into group behaviors and public concerns, with applications in product recommendation and crisis management. Traditional methods for social event detection, including incremental clustering and community detection, overlook the rich semantics and structural information in social data and fail to preserve acquired knowledge.", "Key Idea": "The authors introduce a Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for social event detection that models social messages into unified social graphs, leveraging GNNs for knowledge extraction and preservation.", "Method": "KPGNN utilizes contrastive loss terms for adapting to changing data, employs inductive learning for event detection, adopts a mini-batch subgraph sampling strategy for scalability, and periodically removes outdated data to maintain a dynamic embedding space.", "Outcome": "Experimental results demonstrate that KPGNN outperforms various baselines in social event detection, highlighting its efficiency and effectiveness in acquiring, preserving, and extending knowledge.", "Future Impact": "N/A"} +{"id": "16c1b4ae-73f8-4c23-8bdb-b931ade1baa5", "Context": "Selecting a preferred subset of items from a set of feasible items is a critical task in decision making and support, with recent methods focusing on specifying preferences based on attribute values of the elements in the set.", "Key Idea": "This paper tackles the problem of computing an optimal subset given specific preferences, building on the approach of Brafman et al. (2006), and demonstrates that the problem is NP-hard, necessitating heuristic search methods.", "Method": "Two algorithm classes are considered for tackling the problem: direct set construction and implicit enumeration as solutions to appropriate constraint satisfaction problems (CSPs), with new algorithms presented in each class.", "Outcome": "Empirical comparison of the newly presented algorithms against previous results is conducted, showing the effectiveness of the proposed methods.", "Future Impact": "N/A"} +{"id": "081d6673-3c7c-4aec-b101-cf55d75ac718", "Context": "Previous partial permutation synchronization (PPS) algorithms, used for multi-object matching, often involve computation-intensive and memory-demanding matrix operations, becoming intractable for large-scale structure-from-motion datasets.", "Key Idea": "The authors propose a new algorithm named MatchFAME (Fast, Accurate and Memory-Efficient Matching), which is based on the CEMP-Partial method for estimating corruption levels of observed partial permutations and implements a nonconvex weighted projected power method without spectral initialization.", "Method": "The authors prove the effectiveness of CEMP-Partial under adversarial corruption and demonstrate MatchFAME's accuracy, speed, and memory efficiency through experiments on both synthetic and real datasets.", "Outcome": "MatchFAME is shown to provide state-of-the-art accuracy, speed, and memory efficiency for partial permutation synchronization, outperforming previous algorithms especially in large-scale cases.", "Future Impact": "N/A"} +{"id": "18f27ced-9f37-45d7-9b76-6663c349d408", "Context": "Most existing cross-modal retrieval approaches learn a common subspace in a joint manner, necessitating the involvement of data from all modalities during the entire training process, making the model retrain for new modalities.", "Key Idea": "The paper introduces Scalable Deep Multimodal Learning (SDML), a novel method for cross-modal retrieval that predefines a common subspace to maximize between-class variation and minimize within-class variation, and allows for training modality-specific networks independently.", "Method": "Comprehensive experiments were conducted on four widely-used benchmark datasets to validate the effectiveness and efficiency of SDML in multimodal learning and its performance in cross-modal retrieval.", "Outcome": "SDML is effective and efficient in multimodal learning, demonstrating superior performance in cross-modal retrieval compared to the state-of-the-art methods on four benchmark datasets.", "Future Impact": "N/A"} +{"id": "07c3daea-a88c-4a67-9aac-20ef0ec62e79", "Context": "Non-pharmacological interventions, particularly those utilizing biographical or personalized materials, are effective for people with dementia, but there is a challenge in obtaining these materials and maintaining awareness of what is meaningful for a person.", "Key Idea": "The authors created a web platform to support psychologists by streamlining the collection of relevant biographical information about people with dementia, facilitating reminiscence and other biographical cognitive stimulation practices.", "Method": "The design of the web platform is introduced, and its effectiveness is evaluated through a case study involving one psychologist and three patients over a period of two weeks.", "Outcome": "The platform showed improvements in the collection of meaningful data about people with dementia and in maintaining awareness of the therapy as a whole.", "Future Impact": "N/A"} +{"id": "192f7803-df4d-40c0-b816-ba34339026b3", "Context": "Most current single image super-resolution (SR) methods employ empirical risk minimisation with a pixel-wise mean squared error (MSE) loss, resulting in outputs that are often blurry, over-smoothed and appear implausible.", "Key Idea": "The paper introduces a novel approach for \textit{amortised MAP (Maximum a Posteriori) inference} for image SR using a convolutional neural network (CNN) that ensures the high-resolution output is always consistent with the low-resolution input.", "Method": "The paper proposes three methods to solve the amortised MAP inference problem: Generative Adversarial Networks (GANs), denoiser-guided SR, and a baseline method using a maximum-likelihood-trained image prior, focusing on minimising cross-entropy between two distributions.", "Outcome": "Experiments indicate the GAN-based approach outperforms others on real image data, and a connection between GANs and amortised variational inference, as in variational autoencoders, is established.", "Future Impact": "N/A"} +{"id": "0b31e456-4944-47e5-80ed-deaf6421c375", "Context": "The development of user interface design tools is discussed, focusing on the method of verbal data collection known as \"question-asking protocols\".", "Key Idea": "The key contribution is the introduction of \"directed dialogue protocols,\" a novel interface design tool that enhances the question-asking method with three significant extensions, aiming at improving the collection and quantification of verbal data for design insights.", "Method": "The tool incorporates three extensions: an experimental procedure based on atomic tasks for verbal data quantization, experimenter interventions for probing expectations and prompting verbalizations, and a sequential disclosure technique for answering subject queries.", "Outcome": "Applications of the directed dialogue protocols have identified design choices that improve the learnability and usability of a product's user interface.", "Future Impact": "N/A"} +{"id": "1dea5ec2-d311-4c03-bba5-e38d7a62fbd4", "Context": "The problem addressed is the spotting of a set of signs occurring in videos with sequences of signs, indicating a need for more effective methods.", "Key Idea": "The paper introduces a new method for modeling the spatio-temporal signatures of a sign using Sequential Interval Patterns (SIP) and a novel multi-class classifier organized in a Hierarchical SIP Tree (HSP-Tree).", "Method": "The proposed Hierarchical SIP Tree (HSP-Tree) is evaluated by combining multiple trees into a forest (HSP-Forest) and using it to spot sequences of signs in both concatenated sequences of isolated signs and continuous sign sequences.", "Outcome": "The proposed method shows superior robustness and accuracy for spotting a sequence of signs compared to a state-of-the-art sign recognizer.", "Future Impact": "N/A"} +{"id": "14b0ebd1-b654-4eed-bdd8-ebeb74250b15", "Context": "In few-shot relational triple extraction (FS-RTE), traditional approaches follow an entity-then-relation paradigm which first extracts all entities and then classifies their relations, overlooking the entity discrepancy between different relations.", "Key Idea": "The authors propose a novel task decomposition strategy for FS-RTE called Relation-then-Entity, which first detects the relations in a sentence and then extracts the corresponding head/tail entities of these relations. A model named RelATE is further introduced to implement this strategy.", "Method": "RelATE uses a dual-level attention mechanism to aggregate relation-relevant information for detecting relation occurrences and utilizes the annotated samples of the detected relations for extracting the corresponding head/tail entities.", "Outcome": "The RelATE model significantly outperforms previous works in FS-RTE, showcasing an absolute gain of 18.98% and 28.85% in F1 in two different few-shot settings.", "Future Impact": "N/A"} +{"id": "1946f496-f6cd-4736-8c30-a6ae70baa8b2", "Context": "Existing click models treat intrinsic relevance of documents to queries as a query-document-specific parameter estimated solely from historical clicks, without considering the document content or the relationship among clicked/skipped documents under the same query.", "Key Idea": "The authors propose a novel Bayesian Sequential State model that takes into account the document content and dependencies among sequential click events within a query through a probabilistic graphical model.", "Method": "The posterior regularized Expectation Maximization algorithm is applied for parameter learning to exploit richer information from user clicks and tailor the model to meet specific ranking-oriented properties like pairwise click preferences.", "Outcome": "Experiments on a large set of real click logs show the proposed model to be more effective than several state-of-the-art click models.", "Future Impact": "N/A"} +{"id": "1e5571af-71cf-40b8-ba2a-18b42cae5b42", "Context": "N/A", "Key Idea": "The paper introduces a framework for knowledge-based sequence mining using Answer Set Programming (ASP), demonstrating ease of incorporating knowledge into the mining process and efficiency in extracting condensed patterns.", "Method": "The authors model the basic task of sequence mining and refine it by showing how condensed patterns can be extracted through modular extensions and how ASP's preference handling capacities can aid in mining patterns of interest. An empirical study comparing this approach with a related sequence mining mechanism is provided.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "1ef9b762-e9be-46c5-ad19-090fe16200c4", "Context": "An important issue in human-robot interaction is accurately deriving pointing information from gestures, where it is often the pointed object, not the exact pointing direction, that is important.", "Key Idea": "The authors formulate a novel approach that takes into account prior information about the location of possible pointing targets and uses the Dempster-Shafer theory of evidence to fuse information from head pose and hand pointing orientation.", "Method": "The proposed approach is validated through experiments in realistic application setups, involving tracking the off-plane rotations of the face for head pose estimation and assessing hand pointing orientation.", "Outcome": "Detailed experimental results validate the effectiveness of the method, showing its utility in accurately determining pointed targets in human-robot interaction scenarios.", "Future Impact": "N/A"} +{"id": "2065b977-7782-4981-ad70-3121a2315687", "Context": "N/A", "Key Idea": "The authors introduce a correlated bigram LSA approach for unsupervised Language Model (LM) adaptation in automatic speech recognition, utilizing efficient variational EM for training and fractional Kneser-Ney smoothing for handling fractional counts.", "Method": "The scalability issue is addressed via bootstrapping bigram LSA from unigram LSA, and for LM adaptation, unigram and bigram LSA are integrated into the background N-gram LM through marginal adaptation and linear interpolation. Experiments are conducted on the Mandarin RT04 test set and a large-scale evaluation on Arabic.", "Outcome": "The application of both unigram and bigram LSA yields a 6%-8% relative perplexity reduction and a 2.5% relative character error rate reduction on the Mandarin RT04 test set, and a 3% relative word error rate reduction on a large-scale evaluation on Arabic, all statistically significant improvements.", "Future Impact": "N/A"} +{"id": "1b0e4045-d39b-4bea-8dec-e747f5c674f5", "Context": "Data uncertainty is inherent in applications such as sensor monitoring systems, location-based services, and biological databases, leading to the development of probabilistic databases.", "Key Idea": "The paper addresses the challenge of mining frequent patterns and association rules from probabilistic data under the Possible World Semantics, which is difficult due to the exponential number of possible worlds in a probabilistic database.", "Method": "The authors propose two efficient algorithms for discovering frequent patterns in a bottom-up and top-down manner, which can be extended to discover maximal frequent patterns. They also detail how these patterns can be used to generate association rules. The performance of these methods is validated through extensive experiments using real and synthetic datasets.", "Outcome": "Extensive experiments on real and synthetic datasets validate the efficiency and effectiveness of the proposed algorithms for discovering frequent patterns and association rules in probabilistic databases.", "Future Impact": "N/A"} +{"id": "1e396f93-a73e-4d33-9a8e-56097a8c3c28", "Context": "Deep learning has yielded state-of-the-art performance on many natural language processing tasks including named entity recognition (NER), but it requires large amounts of labeled data.", "Key Idea": "The authors propose combining deep learning with active learning to reduce the needed amount of labeled training data for NER tasks, and introduce a lightweight CNN-CNN-LSTM model that is computationally efficient.", "Method": "Incremental active learning during the training process is utilized to assess the effectiveness of the proposed model and approach.", "Outcome": "The proposed CNN-CNN-LSTM model achieves nearly state-of-the-art performance on standard NER datasets while being more computationally efficient, and nearly matches state-of-the-art performance with just 25% of the original training data.", "Future Impact": "N/A"} +{"id": "206d2d53-dbaf-4a2f-810d-856309d8eb83", "Context": "N/A", "Key Idea": "The authors propose a neural head reenactment system driven by a latent pose representation that predicts the foreground segmentation alongside the RGB image and decomposes pose from identity using a training process based on image reconstruction losses.", "Method": "N/A", "Outcome": "The system successfully reproduces mimics of the driving person and performs cross-person reenactment. The learned descriptors are also useful for keypoint prediction and pose-based retrieval.", "Future Impact": "N/A"} +{"id": "2a220303-8653-497f-b2b5-c829583c2714", "Context": "Counting and sampling directed acyclic graphs from a Markov equivalence class are fundamental tasks in graphical causal analysis.", "Key Idea": "The authors present polynomial-time algorithms for counting and sampling directed acyclic graphs (DAGs) from a Markov equivalence class, addressing a long-standing open problem.", "Method": "The authors benchmark their algorithms against state-of-the-art methods through experimental analysis.", "Outcome": "Experimental results indicate that the proposed algorithms significantly outperform existing state-of-the-art methods.", "Future Impact": "N/A"} +{"id": "29dd9fd3-6c98-4e4b-b70c-0474ff361419", "Context": "Animals exhibit diverse shapes, yet building a deformable shape model for a new species is often hindered by the absence of 3D data. Although birds represent a significant portion of species diversity, no accurate shape model exists for them.", "Key Idea": "The research introduces a method for capturing new species' shape models, focusing on birds, using an articulated template and images. This method allows for the creation of species-specific and multi-species shape models by disentangling pose and shape to learn a shape space from image evidence.", "Method": "The method involves fitting an articulated template to each training sample from the CUB dataset to capture the variation among and within species. This process contributes to developing new species-specific and multi-species shape models.", "Outcome": "The generated shape models are useful for downstream reconstruction tasks and the learned 3D shape space accurately reflects the phylogenetic relationships among birds, outperforming learned perceptual features, using a low-dimensional embedding.", "Future Impact": "N/A"} +{"id": "24e29617-a320-450a-aaa5-19d8700d74b7", "Context": "Existing matrix completion algorithms often suffer from efficiency issues in parallel distributed settings, typically requiring lock-based synchronization or facing challenges with decentralized data.", "Key Idea": "NOMAD is introduced as an efficient, parallel distributed algorithm specifically designed for matrix completion that features non-blocking communication, decentralized variable ownership with asynchronous transfer, and lock-free parallelism.", "Method": "The algorithm's efficiency and performance were validated through extensive empirical evaluation in distributed settings on commodity hardware and compared against state-of-the-art algorithms in both multi-core and distributed memory settings on a High Performance Computing (HPC) cluster.", "Outcome": "NOMAD not only showcased superior performance in decentralized and asynchronous settings but also outperformed synchronous state-of-the-art algorithms in multi-core and distributed memory settings on HPC clusters.", "Future Impact": "N/A"} +{"id": "2b5cc037-4841-4fb1-85ff-673230198be1", "Context": "The problem studied is life-long matrix completion, which is crucial to areas such as recommendation systems, computer vision, and system identification. The main challenge is designing provable algorithms that are noise-tolerant and have small sample complexity.", "Key Idea": "The authors offer algorithms with strong guarantees under two realistic noise models (bounded deterministic noise and sparse random noise), which achieve smaller error or exact recovery with nearly minimal sample complexity.", "Method": "The paper presents the development of algorithms for two noise models, and their performance is validated experimentally on both synthetic and real-world datasets.", "Outcome": "The algorithms achieve a small error in the presence of bounded deterministic noise and exact recovery with high probability in the presence of sparse random noise, showing effectiveness in practical scenarios.", "Future Impact": "N/A"} +{"id": "36d7073a-d006-4a4a-850f-ca7a3f0ca19b", "Context": "Numerous important problems can be framed as learning from graph data.", "Key Idea": "The authors propose a framework for learning convolutional neural networks (CNNs) for arbitrary graphs, which can handle undirected, directed, and graphs with both discrete and continuous node and edge attributes.", "Method": "The authors demonstrate the effectiveness and efficiency of their proposed method using established benchmark datasets.", "Outcome": "The learned feature representations from the proposed CNNs are competitive with state-of-the-art graph kernels and their computation is highly efficient.", "Future Impact": "N/A"} +{"id": "2eb66e5a-472c-4db1-b02f-47fe5eb7e71e", "Context": "Previous models for text generation required syntactic information such as constituency parses or semantic information such as paraphrase pairs to disentangle latent representations of syntax and semantics.", "Key Idea": "The authors propose QKVAE, a generative model for text generation that exhibits disentangled latent representations of syntax and semantics without needing syntactic or semantic information, relying solely on the inductive bias in Transformers.", "Method": "Experiments on latent representations and on syntax/semantics transfer are conducted to evaluate the QKVAE model.", "Outcome": "QKVAE shows clear signs of disentangled syntax and semantics, competitive syntax transfer capabilities to supervised models, and requires less data (fewer than 50K samples) to outperform comparable supervised models on syntactic and semantic transfer.", "Future Impact": "N/A"} +{"id": "31ab88b5-e66f-4b69-98b6-7a470dce9875", "Context": "Conditional neural networks are crucial in sequence-to-sequence modeling tasks but are often burdened by increased model size due to common conditioning approaches like input concatenation or modulation with a conditioning vector.", "Key Idea": "The authors introduce a novel approach for conditioning neural networks by learning intermediate layer activations based on the conditioning vector, aiming at reducing model size while maintaining or improving model quality.", "Method": "The study explores learned activation functions' effectiveness through experiments in personalized sound enhancement (PSE) as a pre-processing technique for telephony or ASR performance improvement under noise, and in personalized ASR for single speaker scenarios.", "Outcome": "The proposed approach of conditioning via learned activation functions resulted in producing conditional models with comparable or better quality than traditional methods while decreasing model sizes, indicating its effectiveness and efficiency.", "Future Impact": "The broad applicability and effectiveness of the proposed conditioning technique suggest its potential across various application domains, especially for on-device deployment where resource efficiency is critical."} +{"id": "37da719b-8db9-4523-8a4a-3f2eef238978", "Context": "N/A", "Key Idea": "The paper proposes a fragile watermarking method with self-embedding for recovery of tampered images without using authentication bits. The method utilizes a robust spread spectrum-based watermarking scheme involving block-based embedding and DCT based compression.", "Method": "Simulation results showing recovery performance of the proposed method are presented.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "3b076d81-3ed0-4d57-84e8-4145f67052bd", "Context": "Referring expressions and other object descriptions should adhere to specific preference rules to avoid leading hearers to infer unwanted conversational implicatures.", "Key Idea": "The authors incorporate preference rules such as Local Brevity, No Unnecessary Components, and Lexical Preference into a polynomial time generation algorithm for avoiding unwanted conversational implicatures.", "Method": "N/A", "Outcome": "The incorporation of preference rules can lead to a polynomial time generation algorithm, whereas alternative formalizations of conversational implicature could make the generation task NP-Hard.", "Future Impact": "N/A"} +{"id": "3b6e7572-2ef8-4565-a3b7-301a3fd38acd", "Context": "Backward locking and update locking are inefficiencies in backpropagation that prevent concurrent updates of layers, and while using local error signals for asynchronous block training shows promise, it requires trial-and-error to optimize.", "Key Idea": "SEDONA, a differentiable search algorithm, is proposed to automate the process of decoupling network blocks and selecting auxiliary networks for each block, aiming at improving efficiency and performance.", "Method": "Experimental validation was conducted comparing SEDONA to end-to-end backpropagation and other greedy learning methods using VGG and ResNet variants on CIFAR-10, Tiny-ImageNet, and ImageNet datasets.", "Outcome": "SEDONA consistently discovered transferable decoupled architectures that significantly outperform those trained with end-to-end backpropagation and other methods in terms of both efficiency and performance, reporting up to 2× speedup in total training time.", "Future Impact": "N/A"} +{"id": "4164fd9a-7ab8-4013-b416-a205231f10f2", "Context": "Plan recognition is the understanding of an agent's goals and plans based on observed behavior. Recent advancements have shown that plan recognition can be tackled efficiently without a plan library using modified planning algorithms.", "Key Idea": "This study extends recent advancements to probabilistic plan recognition, aiming to derive a probability distribution over possible goals given that actions have deterministic effects and both the agent and observer are fully informed about the initial state.", "Method": "The authors propose calculating posterior goal probabilities by defining the probability of a partially observed execution given a goal in terms of the cost difference of meeting the goal with or without complying with observations. This is accomplished via two calls to an unmodified classical planner.", "Outcome": "The feasibility of solving probabilistic plan recognition using standard classical planners is demonstrated through examples, showcasing the method's quality, flexibility, and scalability.", "Future Impact": "N/A"} +{"id": "438caf1c-5c7e-4283-a3cd-bbab302df185", "Context": "Common acquisition functions for active learning focus on either uncertainty or diversity sampling to select difficult and diverse data points from the pool of unlabeled data.", "Key Idea": "The authors introduce CAL (Contrastive Active Learning), an acquisition function that selects contrastive examples, which are similar in the model feature space but have maximally different predictive likelihoods, combining the benefits of uncertainty and diversity sampling.", "Method": "CAL is compared with a diverse set of acquisition functions across four natural language understanding tasks and seven datasets.", "Outcome": "CAL performs consistently better or equal to the best performing baseline across all tasks, on both in-domain and out-of-domain data.", "Future Impact": "The authors conduct an extensive ablation study and analyze all actively acquired datasets, showing that CAL achieves a better trade-off between uncertainty and diversity compared to other strategies."} +{"id": "3ac464c2-2214-4bf2-a6b6-03da2498cb03", "Context": "Query scheduling is a fundamental problem in database management systems, and it has seen renewed attention with the rise of the 'database as a service' (DaaS) model. Existing work has primarily focused on different scheduling algorithms without much consideration of the knowledge about the queries to be scheduled.", "Key Idea": "The authors propose using histograms describing the distribution of likely query execution times as input to the query scheduler and introduce a novel distribution-based scheduling algorithm named Shepherd.", "Method": "The authors validate Shepherd by conducting extensive experimentation using both synthetic and TPC workloads to compare it against state-of-the-art point-based methods.", "Outcome": "Shepherd substantially outperforms state-of-the-art point-based methods in scheduling efficiency as proven by the experiments.", "Future Impact": "N/A"} +{"id": "370141c7-e1bb-4010-9938-efcad6cf2e62", "Context": "Neural architectures have led to significant advances in language modeling and computer vision, but novel architectures often bring about challenging hyper-parameter choices and training instability due to improper network parameter initialization.", "Key Idea": "The paper introduces GradInit, an automated and architecture agnostic method for initializing neural networks that adjusts the norm of each network layer to minimize the loss value in a single step of SGD or Adam.", "Method": "GradInit introduces a scalar multiplier variable in front of each parameter block, optimizing these variables using a simple numerical scheme to demonstrate its effectiveness across various convolutional architectures and the original Transformer architecture.", "Outcome": "GradInit accelerates convergence and test performance of many convolutional architectures, with or without skip connections, even in the absence of normalization layers. It also improves the stability of the Transformer architecture for machine translation, allowing it to be trained without learning rate warmup across a range of learning rates and momentum coefficients.", "Future Impact": "N/A"} +{"id": "432f5702-7b7d-4995-812e-40925f1a18dd", "Context": "Conditional modeling x → y is a central problem in machine learning, especially when x is high dimensional. The novelty here focuses on scenarios where y is high dimensional.", "Key Idea": "The proposed approach involves selecting a small subset yL of the dimensions of y and modeling x → yL and yL → y, to obtain a conditional model x → y with convenient statistical properties.", "Method": "The validation comes from multilabel classification and multivariate regression experiments on several datasets, comparing the performance against the one vs. all approach and multiple sophisticated output prediction methods.", "Outcome": "The method outperforms the one vs. all approach as well as several other sophisticated multiple output prediction methods across different datasets.", "Future Impact": "N/A"} +{"id": "48bacac1-1ca9-4be8-90e6-470596de0e26", "Context": "Previous studies have focused on (semi-)unsupervised methods for automatically learning the polarity of words and phrases for sentiment analysis.", "Key Idea": "The paper presents a method that exploits structural clues to extract polar sentences from Japanese HTML documents for the purpose of building a sentiment lexicon with high precision, albeit at the cost of recall.", "Method": "N/A", "Outcome": "By using a massive collection of HTML documents, the authors are able to compile a sufficiently large corpus of polar sentences to compensate for the method's low recall.", "Future Impact": "N/A"} +{"id": "45b76955-9670-4664-939c-f5a61eb597df", "Context": "Existing metrics for auditing group fairness in ranked lists generally do not account for varying user behaviors and often assume a simplified view, such as a logarithmic loss in importance as a function of rank.", "Key Idea": "The proposed novel metric introduces a blueprint for modeling user attention with parametrization to account for varying user behaviors and allows for non-binary protected attributes to investigate inherently continuous attributes.", "Method": "Three simulated fairness audits were used to apply and test the novel metric, demonstrating how it accounts for user attention in evaluating the fairness of ranked lists.", "Outcome": "The study shows that the fairness of a ranked output is dependent on the knowledge or model of the end-users' attention distribution function, which can lead to perceptions of bias both in favor and against protected groups.", "Future Impact": "N/A"} +{"id": "58279154-e623-46d2-a431-cc409e094e2c", "Context": "Previous models of analogy-driven theorem proving were not as general or cognitively adequate.", "Key Idea": "The paper proposes a model that considers analogy as a control strategy in proof planning, using a source proof-plan to guide the construction of a proof-plan for the target problem and includes a reformulation of the source proof-plan.", "Method": "Several well-known theorems were processed using the proposed analogy-driven proof-plan construction model.", "Outcome": "The proposed model was able to prove several well-known theorems analogically, which could not be proven by previous approaches.", "Future Impact": "N/A"} +{"id": "5b9f94f9-d93f-455d-a110-007ad67ada6d", "Context": "Existing specification languages for tree-based grammars do not adequately support identifier management.", "Key Idea": "XMG (eXtensible Meta-Grammar) offers a sophisticated treatment of identifiers, which effectively supports a linguist-friendly grammar design.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "49309d5a-5959-4f8f-ae30-9fd2350f0cbc", "Context": "Deep convolutional neural networks (CNNs) are essential for various vision tasks, yet they require substantial memory and compute resources, posing a challenge for deployment on constrained devices. While existing compression techniques reduce model sizes, they often do not optimize computational efficiency.", "Key Idea": "The paper introduces 'focused quantization,' a novel quantization strategy that uses power-of-two values and dynamically identifies the most efficient numerical weight representations after fine-grained pruning, particularly for sparse CNNs.", "Method": "The authors apply focused quantization to ResNet-50 and ResNet-18, using lossless encoding to build a compression pipeline, which aims for high compression ratios, low computation costs, and minimal loss in accuracy.", "Outcome": "The application of focused quantization to ResNet-50 resulted in an 18.08x compression ratio with only a 0.24% loss in top-5 accuracy, outperforming existing methods. For ResNet-18, the method not only achieved higher compression ratios and top-5 accuracy but also improved hardware efficiency by requiring fewer logic gates.", "Future Impact": "N/A"} +{"id": "5c268324-e160-489d-9722-6b59d5e3471b", "Context": "N/A", "Key Idea": "The authors introduce a novel approach to creating a more accessible Web through accessible knowledge acquisition mechanisms that target the needs of Web designers for knowledge from minority Web users.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The approach could place marginal Web users in a privileged position, appealing for their inclusion and potentially building a more accessible Web for the benefit of visually-impaired knowledge contributors."} +{"id": "5dccca98-2b58-47e3-9b8f-3b1888aa3976", "Context": "The heterogeneity of today's IT environments and the increasing demands from mobile users are major obstacles for the creation of real-time data processing systems, despite the assertion from Gartner Group about the timely relevance of such systems.", "Key Idea": "The paper introduces the concept of space-based computing, a new middleware paradigm that offers a level of abstraction superior to conventional middleware solutions. It facilitates distributed transactions and seamless integration of mobile devices using open standards.", "Method": "The authors exemplify the implementation of this paradigm through the real-time build-up of data warehouses.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "5b47ca6b-ff9b-429c-adb1-ad9a171eea0e", "Context": "The authors investigate adapting maximum entropy language models for speech recognition by utilizing a large corpus of written language data and a small corpus of speech transcripts.", "Key Idea": "The study employs a recently proposed Bayesian adaptation method for building style-adapted maximum entropy language models for speech recognition.", "Method": "Experiments were conducted to compare the performance of the Bayesian adaptation method against the commonly used linear interpolation method in speech recognition.", "Outcome": "The experiments demonstrated that the Bayesian adaptation method consistently outperforms linear interpolation.", "Future Impact": "N/A"} +{"id": "4e6f8004-9384-4c5c-8d7f-265410a290df", "Context": "Deep learning methods have been used for completing partial data from shape acquisition setups but tend to produce a single output, failing to address the ambiguity in reasoning the missing geometry.", "Key Idea": "The paper proposes a multimodal shape completion problem aiming to complete the partial shape with multiple outputs by learning a one-to-many mapping and develops the first such method using conditional generative modeling without needing paired training data.", "Method": "The approach is validated by extensively evaluating the method on several datasets that contain varying forms of shape incompleteness and comparing it with several baseline methods and variants qualitatively and quantitatively.", "Outcome": "The method demonstrates its merit in completing partial shapes with both diversity and quality when compared among several baseline methods and variants.", "Future Impact": "N/A"} +{"id": "4fdcceeb-f50f-4c4f-8b92-5985498114f8", "Context": "The problem being addressed is the recovery of the three-dimensional motion of a non-rigid object from a sequence of stereo images, which undergoes uniform expansion, three-dimensional shearing about an unknown point in space, in addition to rigid motion.", "Key Idea": "The key idea is reducing the problem of recovering the three-dimensional motion to the unique solution of a set of homogeneous polynomial equations using algebraic geometry, the software package MACAULAY, and the POLSYS program.", "Method": "The method involves the use of algebraic geometry, MACAULAY software, and the POLSYS program to solve a set of homogeneous polynomial equations derived from the assumption of feature correspondence over multiple frames.", "Outcome": "It is demonstrated that, with four-point correspondence, only two stereo snapshots are needed to uniquely determine the motion.", "Future Impact": "N/A"} +{"id": "5d87de73-77d4-4efd-b8e1-d7561b13f69f", "Context": "The disambiguation of preposition senses is a challenging task in natural language processing, and various systems have been proposed to address it using different methods.", "Key Idea": "The authors present a supervised classification approach that uses linguistically motivated features derived from the phrase structure around the preposition, as opposed to using a fixed window size.", "Method": "The system's performance is evaluated using the SemEval 2007 Preposition Sense Disambiguation datasets and is compared to the results of systems that participated in the SemEval workshop. The system is tested with five different classifiers.", "Outcome": "The proposed method reports an increased accuracy, outperforming the best system in the SemEval task.", "Future Impact": "N/A"} +{"id": "532e797a-4b72-488a-80e4-03713d3c8435", "Context": "Traditional Non-Rigid Structure-from-Motion (NRSfM) approaches use linear low-order or low-rank shape models to observe a non-rigid object in a monocular video sequence.", "Key Idea": "This paper proposes exploiting the property of shape recurrency, where many deforming shapes tend to repeat themselves, as a generalized rigidity, and reduces NRSfM problems to rigid ones under certain conditions.", "Method": "The method includes developing efficient algorithms for automatic recurrency detection and for camera view clustering via a rigidity-check, applied on both simulated sequences and real data.", "Outcome": "Experiments demonstrate the effectiveness of the proposed method in reconstructing non-rigid dynamic shapes using standard rigid-Structure from Motion (SfM) techniques without any changes.", "Future Impact": "This novel perspective on rethinking structure-from-motion may inspire other new problems and approaches in the field."} +{"id": "5a3da6ef-67b0-41bc-a994-fc5ff455a27b", "Context": "Online reviews significantly impact customer purchasing decisions, but the presence of pervasive spam reviews can mislead customers into buying low-quality products or defaming decent stores. Most existing methods do not examine the larger portion of reviews written by individuals who only write one review (singleton reviews), which can crucially influence a store's rating.", "Key Idea": "The authors address the problem of detecting spam in singleton reviews by observing that normal reviewer arrival patterns are stable and uncorrelated with their rating patterns over time, in contrast to the bursty nature of spam attacks. They propose detecting spam through unusually correlated temporal patterns and map the singleton review spam detection problem to an anomaly detection problem in temporal patterns.", "Method": "The authors propose a hierarchical algorithm that identifies and constructs multidimensional time series based on aggregate statistics to depict and mine correlations between reviewer arrival and rating patterns. This algorithm detects time windows where attacks are likely and facilitates human inspection by pinpointing these windows in different time resolutions.", "Outcome": "Experimental results demonstrate that the proposed method effectively detects singleton review spam attacks and reveals that singleton reviews are a significant source of spam reviews, largely affecting online store ratings.", "Future Impact": "N/A"} +{"id": "56992082-e04e-4a8b-a985-abfea27fc2e0", "Context": "Dynamic network pruning aims for runtime acceleration by adapting inference paths to different inputs, but existing methods use continuous decision values for weight channel selection, lacking clear and interpretable pruning processes.", "Key Idea": "This paper introduces a method to explicitly model discrete weight channel selections, promoting diverse weight usage and sparser runtime inference paths, alongside an interpretable layerwise channel selection for enhanced model interpretability.", "Method": "The authors conducted experiments on the CIFAR10 and ImageNet datasets to validate the effectiveness of their proposed dynamic network pruning and adversarial example detection algorithm. They compare the prediction accuracy under similar computing budgets against traditional static pruning methods and other dynamic approaches.", "Outcome": "The dynamic network demonstrated higher prediction accuracy on CIFAR10 and ImageNet datasets compared to static and other dynamic pruning methods. The novel adversarial example detection algorithm significantly improved detection rates across various attacks.", "Future Impact": "The work provides an opportunity to develop interpretable and robust models by leveraging the proposed methods for dynamic pruning and adversarial example detection."} +{"id": "68f8d058-1403-4066-b3d3-a8a2836b35e1", "Context": "Annotating precise vector labels on raster imagery, such as Earth or medical imagery, is expensive and time-consuming due to registration uncertainty. Previous research on uncertain training labels often focuses on uncertainty in label class semantics or at the pixel level, not addressing contiguous vector labels.", "Key Idea": "This paper introduces a deep learning framework designed to quantify and reduce the registration uncertainty of vector training labels while simultaneously training neural network parameters.", "Method": "The authors propose a registration-uncertainty-aware loss function and an iterative uncertainty reduction algorithm that re-estimates the posterior of true vector label locations based on a Gaussian process.", "Outcome": "The proposed approach significantly outperforms several baselines in estimating registration uncertainty and classification performance, as validated on real-world datasets in National Hydrography Dataset refinement.", "Future Impact": "N/A"} +{"id": "6de74297-fb80-448f-b7ae-41f8d9701044", "Context": "Central catadioptric cameras and fisheye cameras are two kinds of omnidirectional cameras used in computer vision, which were previously described by different imaging models.", "Key Idea": "The paper presents a unified imaging model that can describe both central catadioptric cameras and fisheye cameras, extending the model proposed by Geyer and Daniilidis for central catadioptric cameras.", "Method": "The unified model is validated through experimental calibration of central catadioptric and fisheye images, demonstrating that existing models for fisheye cameras fit well within this unified model.", "Outcome": "The unified model allows for classification of cameras by a characteristic parameter and the transformation of fisheye images into central catadioptric ones and vice versa. It also enables metric calibration from a single fisheye image using projections of lines, which was not possible with existing methods.", "Future Impact": "N/A"} +{"id": "5e1f387c-d883-4d1f-8397-e4a533a3387b", "Context": "Hyperspectral imaging is crucial for various applications but is hindered by degradation issues and the scarcity of accurate ground-truth signals, making the training of deep neural networks for restoration tasks challenging.", "Key Idea": "The paper introduces a hybrid approach based on sparse coding principles that combines the interpretability of classical techniques with the ability to train model parameters end-to-end without large datasets.", "Method": "The efficacy of the proposed method is demonstrated through various denoising benchmarks.", "Outcome": "The proposed method is computationally efficient and significantly outperforms state-of-the-art approaches in hyperspectral image restoration.", "Future Impact": "N/A"} +{"id": "754f792b-fcf3-42f4-bb2c-5c1dcce21d7a", "Context": "The abstract suggests a general favor towards formal meaning representations for natural language but does not provide detailed context, such as specific issues with current representations or anachora resolution.", "Key Idea": "The paper advocates for formal meaning representations in natural language and uses the problem of identifying antecedents of anachoric expressions as a test (Forcing function) for the adequacy of such representations.", "Method": "N/A", "Outcome": "The paper posits that with an appropriate formal representation of language, it becomes possible to address a rich class of anachora.", "Future Impact": "N/A"} +{"id": "7b0671d7-c8f6-4e81-828b-c73958a6a63a", "Context": "N/A", "Key Idea": "The paper details the design experience of creating an interactive exhibit about medieval music, focusing on educational value, entertainment aspects, and historic authenticity.", "Method": "The paper presents insights into the challenges faced in designing a system that incorporates educational value, entertainment, and historic authenticity, and how these challenges were addressed.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "69aacc53-6730-4db0-b420-9a45b96a642e", "Context": "An important problem in geometric reasoning is finding the configuration of a collection of geometric bodies to satisfy a given set of constraints.", "Key Idea": "The paper suggests solving this problem by symbolically reasoning about geometry using a degrees of freedom analysis and introduces the automatic synthesis of plan fragments from first principles about geometric bodies, actions, and topology.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "7c065e41-7c2e-430e-a918-6a776037bf50", "Context": "The goal of the panel was to discuss whether data management systems architectures should be rethought to truly merge Database (DB) and Information Retrieval (IR) technologies.", "Key Idea": "N/A", "Method": "The paper summarizes the discussions and viewpoints presented at the SIGMOD 2005 panel on merging Database (DB) and Information Retrieval (IR) technologies.", "Outcome": "The panel had very high attendance and generated lively discussions.", "Future Impact": "N/A"} +{"id": "6ecf725b-661e-4897-8169-22d71826d0e8", "Context": "Different Open Information Extraction (OIE) tasks require various types of information, highlighting a strong need for the adaptability of OIE algorithms to meet diverse task requirements.", "Key Idea": "The paper introduces a new adaptable and efficient OIE system called OIE@OIA, which utilizes the Open Information eXpression (OIX) methodology by parsing a sentence to an Open Information Annotation (OIA) Graph and adapting the OIA graph to different OIE tasks with simple rules.", "Method": "An end-to-end OIA generator is implemented as the core of the OIE@OIA system, including annotating a dataset (made openly available) and designing an efficient learning algorithm for the complex OIA graph. The OIE@OIA system is then adapted to three popular OIE tasks to validate its performance.", "Outcome": "The experimental results show that OIE@OIA achieves new state-of-the-art (SOTA) performances on three OIE tasks, demonstrating its high adaptability. Additionally, compared to other end-to-end OIE systems that require millions of training samples, OIE@OIA only needs 12K training samples, showcasing its efficiency.", "Future Impact": "N/A"} +{"id": "80698baf-89cb-4a50-9f5c-0c74151b798b", "Context": "The motivation behind studying how students in a MOOC might be motivated to do a better job during peer grading.", "Key Idea": "The study examines whether students' knowledge that their peer grading efforts are being evaluated by peers, and the exposure to other peer graders' evaluations, motivates them to improve their grading quality.", "Method": "A controlled study was conducted involving more than one thousand students of a popular MOOC, focusing on the effects of students knowing their grading efforts are being watched and evaluated by their peers.", "Outcome": "The study found strong statistical evidence that the process of 'grading the graders' increases the quality of peer grading.", "Future Impact": "N/A"} +{"id": "73e353a8-e0d6-466f-af93-6fccf38fcb18", "Context": "Video-grounded dialogues require reasoning over dialogue context in a multi-turn setting, which is more complex than traditional visual question answering. Previous approaches mostly use dialogue context as simple text input without modelling the turn-level information flows.", "Key Idea": "The proposed approach discovers information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer, introducing a novel method to learn reasoning paths over this semantic graph.", "Method": "The method involves predicting a reasoning path over the semantic graph from the current turn through past dialogue turns containing additional visual cues needed to answer the current question. The reasoning model processes both visual and textual information through this path, and the propagated features are used to generate the answer.", "Outcome": "The experimental results demonstrate the effectiveness of the proposed method in video-grounded dialogues by showing how it efficiently uses semantic dependencies in dialogue to retrieve visual cues.", "Future Impact": "The method provides insights on how models can leverage semantic dependencies in dialogues to retrieve visual information, suggesting potential applications in enhancing video-grounded dialogue systems."} +{"id": "7b21425c-a2b7-4d19-b030-a8350b2a7a80", "Context": "Seq2Seq models, widely utilized for set generation in NLP tasks such as entity typing and dialogue emotion tagging, treat a set as a sequence without capturing its key characteristics like order-invariance and cardinality.", "Key Idea": "The paper introduces a novel algorithm for effectively sampling informative orders across the combinatorial space of label orders and a data augmentation approach called SETAUG that models set cardinality and output, enhancing any Seq2Seq model with order-invariance and cardinality understanding.", "Method": "The authors evaluate the performance of Seq2Seq models enhanced by their proposed data augmentation method on four benchmark datasets across various models including BART, T5, and GPT-3, without requiring any additional annotations.", "Outcome": "Training Seq2Seq models with the augmented data provided by SETAUG achieved an average relative improvement of 20% across different models and datasets.", "Future Impact": "N/A"} +{"id": "8737b031-f77c-4f32-8a68-4be9b0c9ecf8", "Context": "Training a deep network for semantic segmentation requires significant amounts of labeled data, and while synthetic data can ease the labeling effort, networks trained on synthetic data perform poorly on real images due to domain shift.", "Key Idea": "The authors introduce an approach that handles synthetic images without requiring real images for training, based on the observation that foreground and background classes are affected differently by domain shift, suggesting a detection-based treatment for the foreground.", "Method": "The effectiveness of the proposed approach is validated through experiments on the Cityscapes and CamVid datasets using models trained solely on synthetic data.", "Outcome": "The experiments demonstrate the effectiveness of the approach on the Cityscapes and CamVid datasets with models trained only on synthetic data.", "Future Impact": "N/A"} +{"id": "6dc39f88-d613-4ec0-b70d-d5daa6f3643c", "Context": "N/A", "Key Idea": "BitGourmet is introduced as a novel data analysis system that supports deterministic approximate query processing (DAQ), emphasizing deterministic bounds for aggregation queries and allowing users to set precision constraints.", "Method": "The system divides each column vertically, bit-by-bit, and evaluates queries on subsets of these bit vectors using a scenario-specific optimizer for optimal bit selection and execution plan based on quality and cost models.", "Outcome": "BitGourmet demonstrates an interesting trade-off between result quality and execution time, thereby enhancing the interactivity of data analysis. The system also provides manual control over the query plan for deeper insights.", "Future Impact": "N/A"} +{"id": "802a5b78-a022-4d38-bfb3-f28eee4ef89a", "Context": "Neural network-based dependency parsing has gained interest for its ability to alleviate data sparsity and reduce the need for feature engineering by using dense features, yet modeling the complex syntactic and semantic compositions of dense features remains a challenge.", "Key Idea": "The authors propose two heterogeneous gated recursive neural networks, Tree-GRNN for trees in stack with partial dependency structures and DAG-GRNN for nodes without established dependency relations, to automatically learn compositions of dense features for dependency parsing.", "Method": "The effectiveness of the proposed model is validated by conducting experiments on two prevalent benchmark datasets: PTB3 and CTB5.", "Outcome": "The proposed Tree-GRNN and DAG-GRNN models demonstrate their effectiveness in transition-based dependency parsing, as shown by experiment results on PTB3 and CTB5 datasets.", "Future Impact": "N/A"} +{"id": "816707fd-9214-4435-ac40-b2655e55c9d0", "Context": "Pretrained language models have achieved superhuman performance on many benchmarks, highlighting the need for more challenging tasks to better evaluate their natural language understanding capabilities.", "Key Idea": "The paper introduces a new benchmark called CoDA21 (Context Definition Alignment) that measures natural language understanding of PLMs by requiring them to align definitions with contexts without knowing the specific words, demanding deep understanding and complex inference.", "Method": "N/A", "Outcome": "The findings reveal a significant gap between human performance and that of pretrained language models on the CoDA21 task, indicating that it measures an aspect of natural language understanding not adequately covered by existing benchmarks.", "Future Impact": "N/A"} +{"id": "7d850699-212d-466b-976f-0afed0653fef", "Context": "The authors previously developed an n-gram-based data-driven prediction approach for identifying determiner and preposition errors in non-native English essays as part of the Helping Our Own (HOO) 2011 Shared Task.", "Key Idea": "The paper extends the previous approach by incorporating a missing determiner detector and utilizing word clustering into the n-gram prediction approach to better identify determiner and preposition errors.", "Method": "The extended approach is applied to non-native English essays from the Cambridge Learner Corpus FCE Dataset as part of the HOO 2012 Shared Task, focusing on errors such as missing determiner, incorrect determiner, and incorrect preposition.", "Outcome": "Approximately two-thirds of the errors annotated in the HOO 2012 training and test data are identified, indicating the effectiveness of the extended approach.", "Future Impact": "N/A"} +{"id": "8aedb046-2f51-4229-bc19-ea6db98355cb", "Context": "Research on website ranking has predominantly utilized a HostGraph based on the inter-connectivity among websites and applied the random walk model to generate ranks. This approach, however, does not align with actual web surfer behavior, making derivative ranks inaccurate.", "Key Idea": "This paper introduces AggregateRank, a novel method for ranking websites grounded on stochastic complement theory, offering a more accurate approximation of PageRank sums and lower computational complexity.", "Method": "AggregateRank's performance and computational efficiency were validated through theoretical analysis and experimental evaluation, comparing it against previous website ranking methods.", "Outcome": "Both theoretical analysis and experimental results demonstrated that AggregateRank surpasses previous methods in ranking websites, proving to be more efficient and accurate.", "Future Impact": "N/A"} +{"id": "8ae36735-e4ac-48da-bd2b-5538a6a00a73", "Context": "Relational machine learning (RML) is used for modeling user attributes and relations in social networking sites for tasks like event recommendation but faces challenges in large scale domains due to the underutilization of unlabeled data and expensive collective inference procedures.", "Key Idea": "The proposed work introduces a maximum entropy constraint to correct biases introduced by collective inference in RML and develops a massively scalable variational inference algorithm that incorporates this constraint for large scale network domains.", "Method": "The authors applied their method to seven real-world datasets, including large-scale networks with over five million edges, to test the efficacy of their proposed solutions.", "Outcome": "The proposed method demonstrated an improvement over various baselines in accurately predicting preferences in large-scale networks by addressing the limitations of previous RML approaches.", "Future Impact": "N/A"} +{"id": "927df1bd-273a-4088-8c56-2e79cac37072", "Context": "The k-means++ initialization algorithm is known for obtaining a set of initial centers close to the optimum for k-means clustering, but its sequential nature limits its applicability to very large datasets.", "Key Idea": "The paper proposes k-means||, an initialization algorithm for k-means that drastically reduces the number of passes needed to obtain a good initialization in parallel, addressing the limitation of k-means++'s sequential nature.", "Method": "The paper includes a proof that k-means|| obtains a nearly optimal solution after a logarithmic number of passes and experimental evaluation on real-world large-scale data to compare k-means|| to k-means++ in both sequential and parallel settings.", "Outcome": "Experimental evaluation demonstrates that k-means|| outperforms k-means++ in both sequential and parallel settings, specifically showing nearly optimal solutions can be achieved with a constant number of passes in practice.", "Future Impact": "N/A"} +{"id": "93bf12a1-2174-43e3-9cd4-c2b8aeed2f93", "Context": "In unsupervised domain mapping, learning a mapping between two unmatched datasets (A and B) involves simultaneously learning the mappings GAB and GBA, which has been shown to yield convincing results.", "Key Idea": "The authors present a new method for learning the mapping GAB without the need for learning its inverse GBA, by maintaining the distance between a pair of samples during the translation process.", "Method": "The experimental validation involves demonstrating that the proposed method achieves one-sided mapping learning and provides superior numerical results compared to existing methods involving circularity-based constraints.", "Outcome": "The experimental results indicate that not only the proposed method enables one-sided mapping learning, but it also leads to preferable numerical results over the existing circularity-based constraint.", "Future Impact": "N/A"} +{"id": "9999e1df-f439-4f62-bd03-337e494e9da3", "Context": "N/A", "Key Idea": "The paper introduces a data type encapsulation scheme that allows 'space' operations to be expressed naturally using the base language's operators, resulting from a conceptual separation of operators and procedure calls in the base language.", "Method": "The implementation of the encapsulation scheme is discussed, and several examples illustrating its application and benefits are provided.", "Outcome": "The scheme has been implemented, demonstrating the language's considerable expressive power through various examples.", "Future Impact": "N/A"} +{"id": "9748b061-590a-4602-9015-a68e879ffced", "Context": "HCI researchers are interested in using liquid-based materials, like hydrogels, to create novel interfaces but face challenges with the existing printing systems.", "Key Idea": "The paper presents xPrint, a design strategy for HCI researchers to build and customize a liquid-based smart material printing platform using a modular design with off-the-shelf or easy-to-machine parts, supporting a wide range of printing materials and resolutions.", "Method": "The paper details the system design of xPrint and demonstrates its applicability through three use cases, showcasing the material variability and customizability for various user demands, including designers, scientific researchers, and artists.", "Outcome": "xPrint supports a broad range of printable materials, from synthesized polymers to living cells, with printing resolutions from 10μm to 5mm. It demonstrates the system's versatility and customizability through detailed system design and three distinct use cases.", "Future Impact": "N/A"} +{"id": "9292bc08-a70f-4adb-a6d7-920728893a39", "Context": "N/A", "Key Idea": "The paper introduces new propositional argumentation systems by extending the underlying language and associated monotonic logics, establishing a special case within an assumption-based argumentation framework and showing the equivalence of a stronger argumentation system to a system of causal reasoning.", "Method": "N/A", "Outcome": "A stronger argumentation system in a full classical language is shown to be equivalent to a system of causal reasoning.", "Future Impact": "The implications of the correspondence between argumentation systems and causal reasoning for nonmonotonic theories are discussed."} +{"id": "9821d5f7-72b0-4841-a54f-d2af4a04ea3a", "Context": "Inverse rendering traditionally involves complex computations, especially when accounting for non-Lambertian specular reflectance and complex illumination. Existing methods often entail computationally expensive analysis-by-synthesis approaches.", "Key Idea": "This study introduces an approach for inverse rendering where surface texture is approximated by a linear basis, using a dichromatic reflectance model to estimate spherical harmonic illumination coefficients and texture parameters in a specular invariant colour subspace through solving a system of bilinear equations.", "Method": "The authors focus on faces, describing both shape and texture with a linear statistical model, fitting a 3D morphable model to a single color image to account for non-Lambertian specular reflectance and complex illumination. Recovery of statistical texture model parameters is achieved through convex optimization problems.", "Outcome": "The method recovers statistical texture model parameters with accuracy comparable to more computationally demanding analysis-by-synthesis approaches, while only requiring the solution of convex optimisation problems.", "Future Impact": "N/A"} +{"id": "977698a6-56ef-4e57-94c5-b3a68a451a80", "Context": "Emotion cause extraction aims to identify the reasons behind emotions expressed in text, which is more difficult than emotion classification.", "Key Idea": "The authors propose a new approach that views emotion cause identification as a reading comprehension task similar to question answering, leveraging deep memory networks and a mechanism inspired by convolutional neural networks to store context information.", "Method": "Performance evaluation using a recently released emotion cause dataset, comparing the proposed approach to competitive baselines.", "Outcome": "The proposed method achieves state-of-the-art performance on the emotion cause dataset, outperforming competitive baselines by at least 3.01% in F-measure.", "Future Impact": "N/A"} +{"id": "9cd7e7e1-8893-4db6-8327-48f098187699", "Context": "During natural disasters, online social networks, particularly Twitter, have become a crucial alternative information channel. However, the volume and diversity of messages result in information overload for end users.", "Key Idea": "The goal of the research is to develop an automatic classifier to feed a mobile application that helps citizens access relevant information during natural disasters, focusing on filtering tweets related to earthquakes.", "Method": "A dataset from the Chilean earthquake of 2010 is used to build and validate a ground truth. The research details the impact of class imbalance and dimensionality reduction on five classifiers assessing their performance.", "Outcome": "The study contributes by detailing how class imbalance and dimensionality reduction affect the performance of models designed to filter relevant and non-relevant tweets during an earthquake.", "Future Impact": "N/A"} +{"id": "9b95b592-1562-4ef9-b0ed-e0655fadc73b", "Context": "N/A", "Key Idea": "The paper presents a general formulation for geodesic distance propagation of surfaces on a 3-manifold in IR4, generalizing the work of Kimmel et al. and proposing a new method for surface matching that can handle large deformation and topological changes.", "Method": "The approach uses a partial differential equation for surface propagation at equal geodesic distances from the original surface, along with an eulerian formulation with level-sets for stable numerical computation of distance maps.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "99bef0ca-923c-480d-a89e-9a2f27e02157", "Context": "Most real-world and recreational games are of incomplete information, which poses significant challenges. Over the past dozen years, abstraction has become a crucial technique for solving large incomplete-information games.", "Key Idea": "The paper reviews the use of abstraction to create a smaller, strategically similar game to the original, solve for an approximate equilibrium, and then map the strategy back to the original game, while addressing issues such as abstraction pathology.", "Method": "The approach involves reviewing practical algorithms for information and action abstraction, theoretical breakthroughs providing bounds on strategy quality, and methods for reverse mapping the opponent's actions.", "Outcome": "The paper presents key developments in the field including theoretical breakthroughs and practical algorithms for abstraction in incomplete-information games, and strategies for dealing with actions outside of the abstraction.", "Future Impact": "The document discusses topics of current and future research in the area of game abstraction and solving large incomplete-information games."} +{"id": "a5bc5ab8-70fb-4dff-824b-7606228e44a9", "Context": "N/A", "Key Idea": "EzMath introduces an easy-to-learn notation for embedding mathematical expressions in Web pages.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "a10dafe9-6093-47f0-8429-7b62c46566ea", "Context": "Enterprise mashup scenarios involve using data sources like email, news, calendars, blogs, and web feeds for data mashup products, but the process is challenged due to the unstructured nature of these data sources.", "Key Idea": "The integration of SystemT, an information extraction system, with IBM's InfoSphere MashupHub is proposed, using annotators to convert unstructured text into structured information to facilitate mashup operations.", "Method": "The presentation of a demonstration showcasing how to build domain-specific annotators using SystemT's declarative rule language, AQL, and the application of these annotators to combine structured and unstructured information in an enterprise mashup.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "a1739057-ef00-4b01-9c26-4ab2b5d5708e", "Context": "With the increasing use of data mining tools and techniques, there is a need for systems that can efficiently support and optimize a sequence of related complex mining queries and multiple simultaneous queries by several users.", "Key Idea": "The paper introduces a systematic mechanism aimed at optimizing complex mining tasks, particularly those involving frequent pattern mining on one or multiple datasets, through a new system architecture and algorithms.", "Method": "The authors have implemented the proposed system and evaluated it using both real and synthetic datasets.", "Outcome": "Experimental results demonstrate that the proposed techniques can achieve a speedup of up to a factor of 9 compared to systems that do not support caching or optimize for multiple queries.", "Future Impact": "N/A"} +{"id": "9f641e23-3886-4ac6-b65e-28db86ee48be", "Context": "Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction, which was considered suboptimal compared to the Vickrey-Clarke-Groves (VCG) auction according to theoretical accounts.", "Key Idea": "The paper provides a justification for the success of GSP by demonstrating that advertisers' preferences align with a model termed value maximization, rather than profit maximization as traditionally believed, making GSP truthful for value maximizers.", "Method": "The authors explain the theoretical foundation of GSP's success through the concept of value maximization and provide an axiomatization of GSP, showing its applicability beyond its original scope by applying it to arbitrary single-parameter domains.", "Outcome": "This analysis leads to a reinterpretation of GSP as a truthful auction for value maximizers and extends the understanding of GSP's applicability, aligning it with the folklore definition of GSP when applied to broader contexts.", "Future Impact": "N/A"} +{"id": "a722b600-3725-4738-a47b-435aebd63e13", "Context": "Smart home devices today monitor and automate private spaces, often leading to emergent privacy problems, despite the home being a highly private space where people do not expect surveillance.", "Key Idea": "The authors aim to develop a design philosophy for intelligent agents in the smart home that focuses on privacy empowerment, presenting an alternative to current smart home technologies.", "Method": "N/A", "Outcome": "The authors propose the first steps towards the design of privacy empowering technologies for smart home devices, moving away from invasive practices of current devices.", "Future Impact": "This work represents a move towards a future where smart home devices respect user privacy more thoroughly, suggesting a shift in how these devices are designed and built."} +{"id": "adfd8058-64b3-4062-953c-034b732e2fa0", "Context": "N/A", "Key Idea": "The paper presents a video coding scheme that employs texture synthesis through Directional Empirical Mode Decomposition (DEMD) for P and B-frames and uses H.264 for coding I-frames.", "Method": "P and B-frames are decomposed into Intrinsic Mode Function (IMF) image and its residue. Only the first level IMF image for P and B frames are coded. At the decoder stage, subsequent IMF images are synthesized using correlation search. Additionally, wavelet decomposition is performed over the residual image, and the energy level at the HH band guides the number of decompositions for optimum synthesis.", "Outcome": "The experimental results demonstrate the effectiveness of the proposed scheme in multi-resolution parametric modeling of image data, achieving significant compression with acceptable quality through efficient coding.", "Future Impact": "The scheme enables scalable coding of IMF parameters for higher compression with perceptual quality."} +{"id": "b6b29c8a-7c8c-444c-b434-2ff9e166d9aa", "Context": "Offline Reinforcement Learning (RL) learns from previously collected datasets without exploring the environment, but applying off-policy algorithms directly to offline RL often fails due to extrapolation error from out-of-distribution (OOD) actions.", "Key Idea": "The paper proposes Pessimistic Bootstrapping for offline RL (PBRL), an uncertainty-driven offline algorithm without explicit policy constraints, which quantifies uncertainty via disagreement of bootstrapped Q-functions and penalizes the value function based on estimated uncertainty.", "Method": "PBRL conducts uncertainty quantification, performs pessimistic updates, and introduces a novel OOD sampling method. It’s evaluated on the D4RL benchmark to showcase its effectiveness.", "Outcome": "PBRL demonstrates superior performance compared to state-of-the-art algorithms in offline RL, backed by extensive experiments on the D4RL benchmark.", "Future Impact": "N/A"} +{"id": "b0ad60d7-3a3a-42c2-acbc-fba55e708ba0", "Context": "In natural language, the meaning of a lexeme often varies due to the specific surrounding context, posing a challenge for computational natural language processing.", "Key Idea": "The authors developed a new technique for producing a context-dependent 'meaning' representation for a lexeme in a specific context, represented by a list of semantically replaceable elements consisting of other lexemes.", "Method": "Experiments were conducted using a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases to evaluate the effectiveness of the proposed technique.", "Outcome": "The technique enables comparison of the 'meaning' of conceptual units (individual words or frequently-occurring phrases) in different contexts and can serve as features for machine learning approaches to classify semantic roles and relationships.", "Future Impact": "N/A"} +{"id": "b14f2bc1-607f-4d8b-a731-b4afdf30a633", "Context": "Current explanation datasets often employ synthetic data with simple reasoning structures, lacking the capacity to express more complex reasoning processes such as rebuttals to a reasoning step and degrees of certainty of evidence.", "Key Idea": "The authors propose a comprehensive benchmark with a new logical reasoning explanation form that includes conditions of rebuttal, logical formulae uncovering the internal texture of reasoning nodes, and reasoning strength by degrees of certainty to better fit human cognitive processes and address complex real-life scenarios.", "Method": "The authors evaluate the performance of current best models on this new explanation form to assess their abilities in generating reasoning graphs.", "Outcome": "The experimental results indicate that generating reasoning graphs with the proposed fine-grained logical reasoning explanation form remains a challenging task for current models, including those assisted by giant pre-trained language models.", "Future Impact": "N/A"} +{"id": "b391a193-83e3-4f11-801f-1842647d626e", "Context": "Many studies utilize graph-based deep learning methods to exploit dependencies among facial action units (AUs) for AU detection tasks, but these dependencies are often noisy in real-world data.", "Key Idea": "The paper introduces an uncertain graph neural network (UGN) that learns a probabilistic mask to capture dependencies among AUs and their uncertainties, and an adaptive weighted loss function to manage unbalanced data distributions among AUs.", "Method": "Extensive experiments were conducted on two benchmark datasets, BP4D and DISFA, to validate the effectiveness of the proposed UGN method and the adaptive weighted loss function.", "Outcome": "The proposed method achieves state-of-the-art performance on the BP4D and DISFA datasets for facial action unit detection.", "Future Impact": "N/A"} +{"id": "ad69e31c-2c7f-4db0-916c-3deccaab37fd", "Context": "Go is an ancient board game that poses unique opportunities and challenges for AI and machine learning, particularly regarding learning a good evaluation function in a scalable way.", "Key Idea": "The proposed machine learning approach focuses on scalability in learning an effective evaluation function for Go, using a system that learns the propensity of local patterns and integrates this information across the board.", "Method": "The study utilizes a recursive neural network, derived from a Bayesian network architecture, to integrate local tactical information across the board and produce local territory ownership probabilities.", "Outcome": "The system, trained using 9x9 amateur game data, performs well on a test set derived from 19x19 professional game data, suggesting the effectiveness of the approach for scaling across different board sizes.", "Future Impact": "Possible directions for further improvements in the machine learning approach to Go are briefly discussed, indicating areas for potential future research."} +{"id": "bd84a1fd-ee85-4ac9-a6ae-e534b6013506", "Context": "The authors consider convolutional networks from a reproducing kernel Hilbert space viewpoint.", "Key Idea": "The paper establishes harmonic decompositions of convolutional networks, which are expansions into sums of elementary functions of increasing order, related to spherical harmonics.", "Method": "N/A", "Outcome": "The harmonic decompositions enable the characterization of the integral operators associated with convolutional networks and lead to obtaining statistical bounds for convolutional networks.", "Future Impact": "N/A"} +{"id": "c5c22c6f-e9e9-4b0a-ac57-d3baa100033b", "Context": "Structured stochastic convex optimization problems with a large number of linear constraints are challenging, especially because prior methods require full passes over all constraints.", "Key Idea": "The authors propose two novel conditional gradient-based methods that only process a subset of constraints at each iteration, using variance reduction and smoothing in conjunction with conditional gradient steps.", "Method": "The methods are evaluated through preliminary numerical experiments to illustrate practical performance.", "Outcome": "Rigorous convergence guarantees for the proposed methods are provided, indicating their effectiveness.", "Future Impact": "N/A"} +{"id": "c3aa9543-0695-42b4-99af-d033e3912801", "Context": "Despite the size and prominence of the real estate industry, the KDD community has barely addressed it, largely because the industry did not previously appreciate the value of data science methods, and the data science community was unaware of the real estate problems suited to its methods.", "Key Idea": "The tutorial aims to bridge the gap between the real estate industry and the data science community by introducing real estate concepts to data scientists and outlining a range of data science problems, some of which are tackled by new prop-tech companies.", "Method": "The tutorial presents concrete examples from three companies (Airbnb, Cherre, and Compass) where the authors work, demonstrating applications of data science in the real estate industry.", "Outcome": "N/A", "Future Impact": "The tutorial is expected to spark interest in the real estate industry among the data science community and encourage further exploration and problem-solving in this area."} +{"id": "c37e0f9f-1654-4f1f-b812-bf6f67c0c840", "Context": "Recent research has highlighted the vulnerability of deep learning systems in various domains to adversarial attacks, which involve adding small perturbations to inputs to cause misclassification. An active area of research is developing defenses against such attacks, focusing on creating robust models or detecting adversarial inputs.", "Key Idea": "The authors introduce a novel statistical approach for the detection of adversarial inputs in image classification, based on constructing a per-class feature distribution and using statistical distances for comparison.", "Method": "The authors utilize statistical distances such as Energy Distance (ED) and Maximum Mean Discrepancy (MMD) to detect adversarial inputs. Experiments are conducted on the MNIST and CIFAR-10 datasets to evaluate the performance of each metric.", "Outcome": "The proposed approach demonstrates effective adversarial detection performance on the MNIST and CIFAR-10 datasets, showing robustness across various attack methods, sample sizes, and degrees of adversarial perturbation.", "Future Impact": "N/A"} +{"id": "c268a190-6974-4190-8f48-db5dcbda8bc8", "Context": "The data management ecosystem is challenged by data variety, with data in multiple formats such as relational and semi-structured (e.g., XML). Traditional databases are limited by their ability to handle only a single data format.", "Key Idea": "The authors propose a multi-model processing framework that can handle both relational and semi-structured data, introducing a worst-case optimal join algorithm that ensures intermediate results are no larger than the worst-case join results.", "Method": "N/A", "Outcome": "Preliminary results indicate that the proposed multi-model algorithm significantly outperforms baseline join methods in terms of running time and intermediate result size.", "Future Impact": "N/A"} +{"id": "ce74316d-c5dc-47f1-b0c4-0591bc3fb4b6", "Context": "Several methods to evaluate queries over a native XML DBMS have been proposed, which specify both path and keyword constraints. These approaches broadly consist of graph traversal optimized with structure indexes, and information-retrieval style inverted lists.", "Key Idea": "The study proposes a strategy that integrates structure indexes and inverted lists, along with a query evaluation algorithm for branching path expressions based on this integrated approach.", "Method": "Experiments were conducted over the Niagara XML DBMS to demonstrate the benefit of integrating the two forms of indexes.", "Outcome": "The experiments showed benefits from integrating the two forms of indexes in the evaluation of queries over the Niagara XML DBMS.", "Future Impact": "N/A"} +{"id": "ca53b2c4-2912-4515-aae6-938c3f268a60", "Context": "The problem of detecting epidemic tendency by mining search logs.", "Key Idea": "The key idea is an algorithm that selects epidemic related queries/terms based on click-through information and adopts linear regression to model epidemic occurrences and frequencies of epidemic related terms in search logs.", "Method": "The effectiveness of the algorithm is tested by evaluating the correlation between the frequencies of epidemic related terms found in search logs and actual epidemic occurrences.", "Outcome": "The results demonstrate that the algorithm effectively identifies epidemic related terms that show a high correlation with epidemic occurrences. It also shows that combining different epidemic related terms improves performance compared to using a single term.", "Future Impact": "N/A"} +{"id": "c676aecf-7468-4258-bb41-22bc1811bc3a", "Context": "Most existing community search models focus only on the internal cohesiveness of a community, overlooking the importance of modularity, which signifies dense connections inside communities and sparse connections outside.", "Key Idea": "The paper introduces a novel approach for community search, emphasizing the importance of high modularity. It introduces a new graph modularity function named Density Modularity for searching communities with high modularity.", "Method": "The authors design new algorithms that can efficiently solve the DMCS problem, proving the problem's NP-hardness and running the algorithms in log-linear time relative to the graph size.", "Outcome": "The algorithm based on density modularity, DMCS, demonstrates up to 8.5 times higher accuracy in terms of Normalized Mutual Information (NMI) over baseline algorithms in extensive experiments on real-world and synthetic networks.", "Future Impact": "N/A"} +{"id": "ccae9338-7379-4af1-8fea-8945ba429c5c", "Context": "Object tracking is a critical and challenging problem in computer vision, and there is increasing interest in applying deep learning to improve tracking accuracy.", "Key Idea": "The authors propose a novel triplet loss to be used within a Siamese network framework for object tracking, which differs from the traditional pairwise loss, aiming to extract more expressive deep features for object tracking.", "Method": "The effectiveness of the proposed triplet loss is validated through theoretical analysis involving comparison of gradients and back-propagation, and by applying it to three real-time trackers based on the Siamese network, tested across several popular tracking benchmarks.", "Outcome": "The application of the proposed triplet loss to real-time trackers resulted in those trackers operating at almost the same frame-rate as baseline trackers while achieving superior tracking performance and comparable accuracy to recent state-of-the-art real-time trackers.", "Future Impact": "N/A"} +{"id": "db900c02-9a35-4a83-aa19-15b763259100", "Context": "Quantum computing has been recognized for its potential in various fields, including machine learning. However, implementing deep learning, especially Convolutional Neural Networks (CNNs), within the quantum computing paradigm has been challenging due to difficulties in implementing non-linearities with quantum unitaries.", "Key Idea": "The paper introduces a quantum algorithm, Quantum Convolutional Neural Networks (QCNN), which can evaluate and train deep CNNs, potentially offering speedups for both the forward and backward passes. The QCNN can reproduce the outputs of classical CNNs while enabling non-linearities and pooling operations.", "Method": "Numerical simulations are presented for the classification of the MNIST dataset to demonstrate the efficiency of the QCNN.", "Outcome": "N/A", "Future Impact": "The QCNN could open new frontiers in image recognition by supporting more convolution kernels, larger kernels, high-dimension inputs, and high-depth input channels, especially for deep networks."} +{"id": "dca09f6f-b63a-42a1-9eb5-fbfa45bc6389", "Context": "Activity logs from wearable devices are heterogeneous, noisy, diverse in scale and resolution, and have complex interdependencies, making them challenging to model for applications like personalized exercise scheduling and workout recommendation.", "Key Idea": "The paper presents FitRec, an LSTM-based model that captures two levels of context information: context within a specific activity, and context across a user's activity history, aiming to estimate users' heart rates for activities and make personalized recommendations.", "Method": "FitRec is evaluated on a novel dataset containing over 250 thousand workout records, along with hundreds of millions of sensor measurements and metadata. The model's performance is assessed against baselines in several personalized recommendation tasks.", "Outcome": "FitRec demonstrates the capability to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise, showing promise for using wearable data in activity modeling and recommendation.", "Future Impact": "N/A"} +{"id": "dd282632-ee41-45da-add8-d68d89c57e2d", "Context": "Classification activation map (CAM) is used for weakly supervised object localization (WSOL) but struggles to localize objects effectively because it prefers global discriminative factors over regional object cues.", "Key Idea": "The paper introduces a plug-and-play mechanism called BagCAMs, which employs a regional localizer generation (RLG) strategy to produce regional localizers from a well-trained classifier for improved object localization.", "Method": "Experiments were conducted to test the effectiveness of BagCAMs in improving weakly supervised object localization, comparing its performance to baseline methods on three WSOL benchmarks.", "Outcome": "Adopting BagCAMs significantly improved the performance of baseline WSOL methods and achieved state-of-the-art performance on three WSOL benchmarks.", "Future Impact": "N/A"} +{"id": "dfba01c5-a632-4394-8607-9a32f20e526c", "Context": "N/A", "Key Idea": "The paper introduces a system that automatically builds a scene model from real images, including both 3D geometric information and photometric information under various illumination conditions.", "Method": "Structure-from-motion and correlation-based stereo techniques are used for recovering 3D geometric structure from images taken from distinct viewpoints. Photometric property extraction uses a low-dimensional linear space of the spatio-illumination volume, represented by a set of basis images.", "Outcome": "The model built can create realistic renderings from different viewpoints and illumination conditions, applicable in object recognition, virtual reality, and product advertisement.", "Future Impact": "N/A"} +{"id": "ddf8f49c-342a-4cd0-8b3b-b588af08ed0d", "Context": "Soboroff, Nicholas, and Cahan proposed a method for evaluating the performance of retrieval systems without relevance judgments, which correlated with actual evaluations using relevance judgments in the TREC competition.", "Key Idea": "The authors propose a new explanation for the phenomenon observed by Soboroff and colleagues by devising a simple measure to quantify the similarity of retrieval systems through the similarity of their retrieved results.", "Method": "The authors assess the average similarity of a system to other systems in a collection using the proposed similarity measure and compare the results to the methodology proposed by Soboroff et al.", "Outcome": "Evaluating retrieval systems according to average similarity yields results quite similar to those of Soboroff et al., with both techniques being highly correlated, effectively evaluating and ranking retrieval systems by popularity rather than performance.", "Future Impact": "N/A"} +{"id": "ded9a095-d94d-4ccd-8825-283ec4bb7093", "Context": "There is increasing interest in adopting UX within corporate environments, focusing on what competencies lead to effective UX design. There's a noted gap between UX pedagogy and professional practice.", "Key Idea": "This paper explores the evolution of design competence in UX practice, focusing on the transition from educational settings to professional environments through the lens of competence. It proposes a co-construction of identity between designers and their environments as they enter professional UX roles.", "Method": "A 12-week longitudinal study using surveys and interviews was conducted with students beginning internships and full-time positions in UX to document their perceived shift in competence.", "Outcome": "The study documents how participants' perception of their UX design competence evolves over time, influenced by various factors such as tool and representational knowledge, complexity, and corporate culture.", "Future Impact": "The study suggests opportunities for future research in building an understanding of competency in UX, particularly focusing on the early stages of UX practice based on the preliminary framing provided."} +{"id": "e10b883a-ded6-4b4e-9934-6daba2d2f2b2", "Context": "Traditional semantic hashing methods focus on generating binary codes for documents based on similarity in a keyword feature space, with two main limitations: unused tag information and inadequate reflection of semantic relationships beyond keyword matching.", "Key Idea": "The paper introduces Semantic Hashing using Tags and Topic Modeling (SHTTM), a novel approach that utilizes both tag information and probabilistic topic modeling to ensure hashing codes are consistent with tag information and preserve document topic/semantic similarity.", "Method": "The authors design a unified framework and an iterative coordinate descent procedure for learning optimal hashing codes. They conduct extensive empirical studies on four different datasets to demonstrate SHTTM's advantages.", "Outcome": "The SHTTM approach shows marked improvements over several state-of-the-art semantic hashing techniques in empirical studies, demonstrating the benefits of modeling tag information and utilizing topic modeling both separately and in combination.", "Future Impact": "N/A"} +{"id": "e827ee51-aafd-4f3b-99ca-60a5e289a555", "Context": "The paper addresses the challenge of Recognizing Textual Entailment (RTE) as part of the PASCAL 2005 challenge. This topic explores how to determine if one piece of text can logically be inferred from another.", "Key Idea": "The paper proposes a purely logical approach to recognize textual entailment by extracting atomic propositions from both the text and the hypothesis and representing them in a custom logical notation.", "Method": "To extract atomic propositions and represent them, the system utilizes the output of the Link Parser. For detecting independent entailment relations, the system incorporates Otter and WordNet.", "Outcome": "The approach of using atomic propositions and custom logical notation for recognizing textual entailment was successfully implemented and tested in the context of the PASCAL 2005 challenge.", "Future Impact": "N/A"} +{"id": "ea414544-c89c-4039-8227-23b11e9a1239", "Context": "Existing question answering systems that supplement Web search engines may struggle to provide concise single-snippet answers to definition questions.", "Key Idea": "The study introduces a practically unsupervised learning method that utilizes online encyclopedias and dictionaries to automatically generate a large dataset of positive and negative definition examples for training an SVM to distinguish between the two.", "Method": "The method was validated through experiments demonstrating its viability, comparing its performance with a system trained on questions and news articles from TREC.", "Outcome": "The proposed method outperformed the training system on TREC questions and news articles, significantly improving the search engine's ability to handle definition questions.", "Future Impact": "N/A"} +{"id": "e2466d98-98f9-41a8-8a38-e862adc2ed47", "Context": "A systematic literature review was done to form a richer context for IT Governance, highlighting the importance of IT Governance models for public organizations and the need for such a model.", "Key Idea": "The paper presents an IT Governance model based on COBIT 4.1 maturity model that can be adopted by both practitioners and researchers for public sector organizations.", "Method": "An empirical survey using a questionnaire based on the COBIT 4.1 maturity model was conducted to investigate IT Governance practice in multiple case studies from the Kingdom of Bahrain.", "Outcome": "The results enable public sector organizations to adopt an IT Governance model in a simple and dynamic manner, offering a basic structure to gain a better perspective on IT Governance processes and focusing decision-making.", "Future Impact": "The IT Governance model provides a basis for further research in IT Governance adoption models and bridges the gap between conceptual frameworks and real-life functioning governance."} +{"id": "e1a191db-6463-4be9-883d-dfc7fb05c5f4", "Context": "N/A", "Key Idea": "The paper introduces a novel dependency parsing method that enforces two structural properties on dependency trees: bounded block degree and well-nestedness, to better represent admissible dependency structures and connect dependency parsing to context-sensitive grammatical formalisms.", "Method": "The problem is cast as an Integer Linear Program solved with Lagrangian Relaxation, from which a heuristic and an exact method based on Branch-and-Bound search are derived.", "Outcome": "The proposed methods are found to be efficient and competitive with a baseline unconstrained parser, while successfully enforcing the structural properties in all cases.", "Future Impact": "N/A"} +{"id": "f23bdd28-b2d6-4a42-a56c-c9774f6451b5", "Context": "With the development of hypersonic vehicles such as X-51A and HTV-2, tracking these vehicles in near space has become a new challenge and area of interest.", "Key Idea": "The paper introduces a learning tracking algorithm specifically designed for hypersonic targets, particularly for the sliding jump maneuver, leveraging a Sine model to closely mimic this specific maneuver.", "Method": "The algorithm employs two Sine models with differing angular velocities within an IMM (Interactive Multiple Model) framework and adapts the sampling rate based on learned target tracking error characteristics. Its performance is compared against a single model algorithm and general IMM algorithms with fixed sampling rates through simulation experiments.", "Outcome": "Simulation experiments demonstrate that the proposed learning tracking algorithm can significantly improve tracking accuracy for hypersonic targets.", "Future Impact": "N/A"} +{"id": "f2c9f8c3-f9d1-4cae-b7ae-a919ada1daaf", "Context": "Access to online visual search engines requires the sharing of private user content -- the query images, posing a privacy concern.", "Key Idea": "The authors introduce the concept of a targeted mismatch attack for deep learning-based retrieval systems to generate an adversarial image that conceals the query image but leads to identical or very similar retrieval results.", "Method": "The authors design various loss functions for the adversarial image construction to evaluate the attacks on standard retrieval benchmarks, including considerations for unknown global pooling operations or unknown input resolution by the retrieval system.", "Outcome": "Successful attacks to partially unknown systems are demonstrated, showing how adversarial images leading to similar retrieval results as the original image can be generated.", "Future Impact": "N/A"} +{"id": "eb15ebe7-aa58-4a98-8f9e-939967c6359f", "Context": "The synchronization problem involves inferring unknown values attached to graph vertices, with edges labelled by the ratio of the incident vertices, for graphs where labels belong to a group. The problem extends to multi-graphs, those with multiple edges connecting the same pair of nodes, a scenario often encountered when modeling relationships between vertices with multiple measures.", "Key Idea": "The paper introduces MultiSynch, a synchronization algorithm for multi-graphs that approaches the problem through a constrained eigenvalue optimization, offering a solution applicable to any linear group.", "Method": "MultiSynch is validated empirically through experiments on both synthetic and real problems, showcasing its ability to handle multi-graph synchronization more effectively than the baseline approach of reducing multi-graphs to simple graphs by averaging their multi-edges.", "Outcome": "MultiSynch proves to be a more precise and accurate solution for the synchronization of multi-graphs compared to the baseline method of averaging multi-edges, as demonstrated through empirical evidence.", "Future Impact": "N/A"} +{"id": "e58b9947-7a3a-414d-a0e8-d6cf02ed7127", "Context": "With the integration of advances in RFID, wireless sensor networks, and Web-based services, physical things are becoming a crucial part of the ubiquitous Web, leading to a new research area of things recommendation in IoT.", "Key Idea": "The paper proposes a unified probabilistic framework for things recommendation in IoT by fusing information across relationships between users (social networks) and things (things correlations).", "Method": "The validation of the proposed approach is conducted on an IoT platform, utilizing the advantages of matrix factorization and exploiting the merits of social relationships and thing-thing correlations.", "Outcome": "The experimental results on the IoT platform demonstrate the feasibility and effectiveness of the proposed recommendation approach.", "Future Impact": "N/A"} +{"id": "f3cef657-1887-42a7-9e11-8fdf54f8fa90", "Context": "Catastrophic forgetting is a significant challenge in continual learning, where models tend to forget previous knowledge when learning new tasks.", "Key Idea": "The paper introduces kernel continual learning, a method that leverages the non-parametric nature of kernel methods with an episodic memory unit to tackle catastrophic forgetting and avoids task interference by learning task-specific classifiers based on kernel ridge regression.", "Method": "The authors deploy variational random features to learn a data-driven kernel for each task by formulating kernel continual learning as a variational inference problem, with the variational posterior distribution over the random Fourier basis inferred from the coreset of each task.", "Outcome": "Extensive evaluation on four benchmarks demonstrates that kernel continual learning is effective in tackling catastrophic forgetting and promise in using kernels for continual learning.", "Future Impact": "N/A"} +{"id": "fb538ce2-abf4-4bd8-b35c-1bfe3ab9f48e", "Context": "The complex word identification task focuses on identifying difficult words in a sentence for specific target audiences, which is crucial for lexical simplification to improve text readability.", "Key Idea": "To address this task, two systems were developed using Naive Bayes and Random Forest Classifiers, incorporating various lexical and semantic features to identify complex words.", "Method": "N/A", "Outcome": "The Naive Bayes classifier based system, after incorporating rule-based post-processing techniques, achieves the maximum G-score of 76.7%.", "Future Impact": "N/A"} +{"id": "f41bfd75-a6b0-44d7-a842-b117797ebd7e", "Context": "The paper addresses feature generation from various data representations like surface texts and parse trees for answer extraction, noting a gap in effectively utilizing more structured data representations.", "Key Idea": "The authors propose comparing three methods (feature vector, string kernel, and tree kernel) to represent syntactic features in Support Vector Machines for answer extraction.", "Method": "The authors conduct experiments on the TREC question answering task to evaluate the effectiveness of features generated from structured data representations versus those from surface text.", "Outcome": "The experiments demonstrate that features generated from more structured data representations significantly enhance performance compared to those generated from surface texts.", "Future Impact": "N/A"} +{"id": "f32d6bc3-d75e-4e84-8bfa-3c83578281dc", "Context": "Transaction log analysis is used to examine user commands and system responses during an online information search.", "Key Idea": "The paper focuses on using stochastic pattern developments within parsed user sessions, Markov chain analysis, and state transition probability matrices for transaction log analysis to evaluate and improve information systems.", "Method": "The methodology involves obtaining machine-readable transaction log tapes from online catalogs and analyzing them to discover system usage and actual user patterns when conducting an information search.", "Outcome": "The use of the methodology allowed for the discovery of patron use and system response patterns from several online public catalogs, showcasing its effectiveness in evaluating information systems.", "Future Impact": "The predictive power of the methodology could lead to the development of real-time aids for system design and improvement."} +{"id": "ffd14676-a525-479f-a74e-2c5d3a85c510", "Context": "Interest in parallel systems has been revived, focusing on computation through excitatory and inhibitory interactions in networks of neuron-like units, particularly for early stages of visual processing and the representation of small local fragments.", "Key Idea": "The paper tackles the challenge of representing shapes in parallel systems and proposes mechanisms for shape perception and visual attention, offering a novel interpretation of the Gestalt principle 'the whole is more than the sum of its parts'.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "f82f2e26-5437-4397-b781-50668ce5851b", "Context": "In current Statistical Machine Translation (SMT) systems, alignment models are trained independently of the translation models, and their parameters are not directly optimized for the translation task.", "Key Idea": "The paper introduces a novel framework for discriminative training of alignment models without the need for annotated link labels at the word level, using automatic translation metrics as the maximization criterion.", "Method": "The framework's effectiveness is evaluated through automatic translation evaluation metrics to assess the translation quality.", "Outcome": "An improvement in translation quality was observed when using the proposed framework, indicating that alignment models can be effectively optimized for the translation task without annotated data.", "Future Impact": "N/A"} +{"id": "f4f1bbaf-c1a2-44d5-8305-27235fa69d62", "Context": "The field of web mining, web science, data science, and data mining has shown a lot of interest in the analysis of (social) networks, emphasizing the need for handling the growing complexity of heterogeneous data.", "Key Idea": "This tutorial introduces feature-rich networks as a powerful modeling approach for capturing data and knowledge from multiple heterogeneous sources, focusing on multiplex and attributed networks to allow diverse analytical perspectives.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "It outlines future research directions in the emerging field of feature-rich networks, suggesting a unified perspective for further studies."} +{"id": "f32e53d5-c7f4-407e-a5cb-26fef230b5fd", "Context": "Matrix factorization techniques for recommendations, which learn latent embeddings of users and items from observed interaction data, commonly assume these embeddings are static or evolve homogeneously over time, which does not accurately reflect most real-world scenarios.", "Key Idea": "The paper introduces Dynamic Bayesian Logistic Matrix Factorization (DBLMF), a model that captures the heterogeneous evolution of user preferences and item attributes over time, using a diffusion process to model these dynamics within logistic matrix factorization.", "Method": "The authors validate DBLMF through extensive experiments on real datasets, comparing its performance against state-of-the-art methods. They also propose an efficient Bayesian inference algorithm to enhance the scalability of DBLMF on large datasets.", "Outcome": "The experiments demonstrate DBLMF's effectiveness in learning dynamic, heterogeneous embeddings for users and items that better capture their evolving interactions, outperforming state-of-the-art methods in recommendation tasks.", "Future Impact": "N/A"} +{"id": "fed7302a-43a7-412e-8ace-d07905e38c3c", "Context": "The main challenge of one-class collaborative filtering is how to interpret and model the latent signal from the missing class.", "Key Idea": "The paper introduces a novel Bayesian generative model for implicit collaborative filtering that differentiates between a user disliking an item and being unaware of it by treating the latent signal as an unobserved random graph.", "Method": "The model's efficacy is validated via large-scale distributed learning using stochastic gradient descent and mean field variational inference over random graph samples, with a fine-grained comparison against a state-of-the-art baseline on real-world data.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "f64fdfde-7e93-411b-865a-1e29d71c95b2", "Context": "Hierarchical topic models (HTMs) are a natural extension of flat topic models, offering better generalization and understanding by learning topics at different levels of abstraction. However, scalable systems for flat topic models struggle with HTMs due to their complex structures and susceptibility to local optima.", "Key Idea": "The paper introduces an efficient partially collapsed Gibbs sampling algorithm for the hierarchical latent Dirichlet allocation (hLDA) model, alongside an initialization strategy to mitigate issues with local optima in tree-structured models.", "Method": "The authors propose efficient data layout for vectorizing HTM and distributed data structures, including dynamic matrices and trees, to address new system challenges in scalable HTM systems. They empirically test this approach on a 131-million-document corpus with 28 billion tokens.", "Outcome": "The proposed system demonstrates an 87-fold efficiency increase over the prior open-source implementation for hLDA and successfully scales to thousands of CPU cores. It extracts 1,722 topics from a massive corpus with 50 machines in just 7 hours.", "Future Impact": "N/A"} +{"id": "feb75e1f-7838-48ca-9a78-cc31b717e5bf", "Context": "Most previous research on detecting fraudulent content on crowd-sourced review platforms like Yelp, TripAdvisor, and Foursquare focuses on supervised learning and textual or stylometry features. They often operate with a limited dataset and assume a limited threat model, mainly detecting fraud near sudden shifts in data or positive campaigns.", "Key Idea": "The authors propose OneReview, a system that finds fraudulent content on crowd-sourced review sites by leveraging cross-dataset analysis between independent review sites and employing textual, contextual, and metadata features to detect anomalies in a business's reputation across multiple sites.", "Method": "The method involves using Change Point Analysis on reviews from each business across different sites, employing a Change Point Analyzer to assess these points for discrepancies, and using supervised machine learning with a combination of features to identify fraudulent reviews among those deemed suspicious. They utilize data from Yelp and TripAdvisor for evaluation.", "Outcome": "The approach achieved 97% accuracy, 91% precision, and 90% recall through k-cross validation on a set of 5,655 labeled reviews. It successfully identified 61,983 reviews (about 8% of all reviews) as fraudulent, detected 3,980 businesses with fraudulent activities, and discovered several large astroturfing campaigns, demonstrating OneReview's effectiveness in detecting fraudulent campaigns.", "Future Impact": "N/A"} diff --git a/data/mixtral_0531.jsonl b/data/mixtral_0531.jsonl new file mode 100644 index 0000000..b29fbff --- /dev/null +++ b/data/mixtral_0531.jsonl @@ -0,0 +1,120 @@ +{"id": "01f161fe-dd40-45dd-89bd-fb1562771d73", "Context": "Event forecasting is an important task for applications such as automated analysis generation and resource allocation. However, capturing contextual information within event forecasting is challenging due to several factors such as uncertainty of context structure, high dimensional features, and adaptation of features over time.", "Key Idea": "The authors propose a novel graph convolutional network that extracts and learns graph representations from historical/prior event documents to predict the occurrence of future events and identify sequences of dynamic graphs as event context.", "Method": "The authors perform experiments on multiple real-world data sets to evaluate the proposed graph convolutional network against various state-of-the-art methods for social event prediction.", "Outcome": "Experimental results show that the proposed method is competitive against various state-of-the-art methods for social event prediction.", "Future Impact": "N/A"} +{"id": "07c3daea-a88c-4a67-9aac-20ef0ec62e79", "Context": "Non-pharmacological interventions, such as reminiscence and biographical cognitive stimulation practices, are common and effective for people with dementia. However, obtaining and maintaining biographical or personalized materials can be challenging.", "Key Idea": "The authors created a web platform that supports the work of psychologists in collecting and managing biographical materials for use in reminiscence and other biographical cognitive stimulation practices.", "Method": "The authors conducted a case study with one psychologist and three patients, using the platform for a period of two weeks.", "Outcome": "The results of the case study showed improvements in the collection of meaningful data about a person and in maintaining awareness of the therapy as a whole.", "Future Impact": "The platform has the potential to be widely adopted in the field of dementia care, improving the quality and efficiency of non-pharmacological interventions."} +{"id": "081d6673-3c7c-4aec-b101-cf55d75ac718", "Context": "Previous partial permutation synchronization (PPS) algorithms for multi-object matching involve computation-intensive and memory-demanding matrix operations which become intractable for large scale structure-from-motion datasets.", "Key Idea": "The authors propose a new algorithm, CEMP-Partial, for estimating the corruption levels of the observed partial permutations, which allows for the implementation of a nonconvex weighted projected power method without spectral initialization.", "Method": "The authors prove that under adversarial corruption, with certain assumptions, CEMP-Partial is able to exactly classify corrupted and clean partial permutations and demonstrate the state-of-the-art accuracy, speed and memory efficiency of the method on both synthetic and real datasets.", "Outcome": "The new PPS algorithm, MatchFAME, only involves sparse matrix operations and enjoys lower time and space complexities in comparison to previous PPS algorithms.", "Future Impact": "N/A"} +{"id": "0b31e456-4944-47e5-80ed-deaf6421c375", "Context": "The development of user interface design tools is based on Katou's (1986) method of verbal data collection, 'question-asking protocols.'", "Key Idea": "The authors propose 'directed dialogue protocols', an extension to the question-asking method that includes an experimental procedure of atomic tasks, interventions by the experimenter, and a technique for answering subject queries called sequential disclosure.", "Method": "N/A", "Outcome": "The proposed method has identified design choices which build learnability and usability into a product's user-interface.", "Future Impact": "N/A"} +{"id": "10c15fe5-c315-4b6d-8910-e6bc3279c817", "Context": "Most existing methods for incremental social event detection learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data and cannot memorize previously acquired knowledge.", "Key Idea": "The authors propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection that models complex social messages into unified social graphs to facilitate data utilization and explores the expressive power of GNNs for knowledge extraction.", "Method": "The authors use KPGNN to detect social events by adopting contrastive loss terms that cope with a changing number of event classes, leveraging the inductive learning ability of GNNs to efficiently detect events and extend its knowledge from previously unseen data, and using a mini-batch subgraph sampling strategy for scalable training.", "Outcome": "KPGNN requires no feature engineering, has few hyperparameters to tune and extensive experiment results demonstrate the superiority of KPGNN over various baselines.", "Future Impact": "N/A"} +{"id": "14b0ebd1-b654-4eed-bdd8-ebeb74250b15", "Context": "In few-shot relational triple extraction (FS-RTE), the goal is to extract relational triples from plain texts using only a few annotated samples. Previous work extracts all entities and then classifies their relations, which ignores the entity discrepancy between relations.", "Key Idea": "The authors propose a novel task decomposition strategy called Relation-then-Entity (RelATE) for FS-RTE. It first detects relations in a sentence and then extracts the corresponding head/tail entities of the detected relations.", "Method": "The authors propose a model called RelATE that builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and uses the annotated samples of the detected relations to extract the corresponding head/tail entities.", "Outcome": "The proposed model outperforms previous work by an absolute gain of 18.98% and 28.85% in F1 in two few-shot settings.", "Future Impact": "N/A"} +{"id": "16c1b4ae-73f8-4c23-8bdb-b931ade1baa5", "Context": "Various decision making tasks require selecting a preferred subset of items from a given set of feasible items, and recent work has considered methods for specifying such preferences based on the attribute values of individual elements within the set.", "Key Idea": "This paper proposes new algorithms for computing an optimal subset given a specification, based on direct set construction, and implicit enumeration as solutions to appropriate CSPs.", "Method": "The authors present new algorithms in each class and compare them empirically against previous results.", "Outcome": "The authors show that the problem is NP-hard in the general case and that heuristic search methods are necessary.", "Future Impact": "The proposed algorithms may improve or extend the current state-of-the-art in solving the problem of computing an optimal subset given a specification."} +{"id": "18f27ced-9f37-45d7-9b76-6663c349d408", "Context": "Most existing cross-modal retrieval approaches learn a common subspace in a joint manner, where the data from all modalities have to be involved during the whole training process, and the optimal parameters of different modality-specific transformations are dependent on each other. These approaches require the whole model to be retrained when handling samples from new modalities.", "Key Idea": "The authors propose a novel cross-modal retrieval method called Scalable Deep Multimodal Learning (SDML) that predefines a common subspace in which the between-class variation is maximized while the within-class variation is minimized. It trains modality-specific networks independently to transform the multimodal data into the predefined common subspace, which can be scalable to the number of modalities.", "Method": "The authors train m modality-specific networks for m modalities (one network for each modality) to transform the multimodal data into the predefined common subspace. They conduct comprehensive experimental results on four widely-used benchmark datasets to evaluate the effectiveness and efficiency of the proposed method.", "Outcome": "The proposed SDML method is effective and efficient in multimodal learning and outperforms the state-of-the-art methods in cross-modal retrieval.", "Future Impact": "N/A"} +{"id": "192f7803-df4d-40c0-b816-ba34339026b3", "Context": "Most current single image SR methods use empirical risk minimization, often with a pixel-wise mean squared error (MSE) loss. However, the outputs from such methods tend to be blurry, over-smoothed and generally appear implausible.", "Key Idea": "The authors introduce new methods for amortized MAP inference, where they calculate the MAP estimate directly using a convolutional neural network. They propose a neural network architecture that performs a projection to the affine subspace of valid SR solutions, and propose three methods to solve the optimization problem: Generative Adversarial Networks (GAN), denoiser-guided SR, and a baseline method using a maximum-likelihood-trained image prior.", "Method": "The authors perform experiments using the proposed amortized MAP inference methods on real image data and compare the results with the baseline method.", "Outcome": "The experiments show that the GAN based approach performs best on real image data.", "Future Impact": "The authors establish a connection between GANs and amortized variational inference as in e.g. variational autoencoders."} +{"id": "1946f496-f6cd-4736-8c30-a6ae70baa8b2", "Context": "Existing click models treat intrinsic relevance as an atomic query-document-specific parameter, which is solely estimated from historical clicks without using any content information about a document or relationship among the clicked/skipped documents under the same query. This leads to limitations in fully exploring the information about a document's relevance quality and making predictions for unseen documents.", "Key Idea": "The authors propose a novel Bayesian Sequential State model for click modeling that characterizes document content and dependencies among sequential click events within a query by a set of descriptive features via a probabilistic graphical model.", "Method": "The authors apply the posterior regularized Expectation Maximization algorithm for parameter learning and tailor the model to meet specific ranking-oriented properties, such as pairwise click preferences.", "Outcome": "Experiment results on a large set of real click logs demonstrate the effectiveness of the proposed model compared with several state-of-the-art click models.", "Future Impact": "N/A"} +{"id": "1b0e4045-d39b-4bea-8dec-e747f5c674f5", "Context": "Data uncertainty is inherent in applications such as sensor monitoring systems, location-based services, and biological databases. To manage this vast amount of imprecise information, probabilistic databases have been recently developed.", "Key Idea": "The authors propose two efficient algorithms for the discovery of frequent patterns and association rules from probabilistic data under the Possible World Semantics, which can be extended to discover maximal frequent patterns.", "Method": "The authors propose bottom-up and top-down algorithms for discovering frequent patterns and association rules from probabilistic data.", "Outcome": "The proposed algorithms were validated through extensive experiments using real and synthetic datasets.", "Future Impact": "N/A"} +{"id": "1dea5ec2-d311-4c03-bba5-e38d7a62fbd4", "Context": "The problem of spotting a set of signs occurring in videos with sequences of signs is tackled in this paper.", "Key Idea": "The authors propose to model the spatio-temporal signatures of a sign using an extension of sequential patterns that contain temporal intervals called Sequential Interval Patterns (SIP) and a novel multi-class classifier that organises different sequential interval patterns in a hierarchical tree structure called a Hierarchical SIP Tree (HSP-Tree).", "Method": "The authors evaluate the proposed method on both concatenated sequences of isolated signs and continuous sign sequences and show that the proposed method is superior in robustness and accuracy to a state of the art sign recogniser when applied to spotting a sequence of signs.", "Outcome": "The proposed HSP-Forest classifier is shown to be superior in robustness and accuracy to a state of the art sign recogniser when applied to spotting a sequence of signs.", "Future Impact": "N/A"} +{"id": "1e396f93-a73e-4d33-9a8e-56097a8c3c28", "Context": "Deep learning has yielded state-of-the-art performance on natural language processing tasks including named entity recognition (NER), but it typically requires large amounts of labeled data.", "Key Idea": "The authors demonstrate that the amount of labeled training data can be drastically reduced when deep learning is combined with active learning. They also propose a lightweight CNN-CNN-LSTM model for NER.", "Method": "The authors perform incremental active learning during the training process and apply the proposed model to standard datasets for NER.", "Outcome": "The proposed CNN-CNN-LSTM model achieves nearly state-of-the-art performance on standard datasets for NER while being computationally more efficient than best performing models. The authors are able to nearly match state-of-the-art performance with just 25% of the original training data.", "Future Impact": "N/A"} +{"id": "1e5571af-71cf-40b8-ba2a-18b42cae5b42", "Context": "Sequence mining is the task of extracting meaningful patterns from a sequence of data. Existing sequence mining techniques do not make use of background knowledge.", "Key Idea": "The authors propose a framework for knowledge-based sequence mining using Answer Set Programming (ASP) that allows for the incorporation of background knowledge.", "Method": "The authors demonstrate the ease of extracting condensed patterns using modular extensions of the basic ASP-based approach. They also show how ASP's preference handling capacities can be exploited for mining patterns of interest.", "Outcome": "An empirical study is provided that compares the proposed ASP-based sequence mining approach with a related sequence mining mechanism.", "Future Impact": "N/A"} +{"id": "1ef9b762-e9be-46c5-ad19-090fe16200c4", "Context": "In human-robot interaction, accurately deriving pointing information from a corresponding gesture is an important issue.", "Key Idea": "The authors propose a novel approach that takes into account prior information about the location of possible pointing targets and uses Dempster-Shafer theory of evidence to fuse information from head pose and hand pointing orientation.", "Method": "The authors perform detailed experimental results that validate the effectiveness of the method in realistic application setups.", "Outcome": "The proposed method is able to accurately derive pointing information from a corresponding gesture in realistic application setups.", "Future Impact": "N/A"} +{"id": "2065b977-7782-4981-ad70-3121a2315687", "Context": "In automatic speech recognition, unsupervised language model adaptation is a common problem. The existing methods use unigram LSA for unsupervised LM adaptation.", "Key Idea": "The authors propose a correlated bigram LSA approach for unsupervised LM adaptation which is trained using efficient variational EM and smoothed using fractional Kneser-Ney smoothing. The authors also address the scalability issue to large training corpora by bootstrapping of bigram LSA from unigram LSA.", "Method": "The authors train the correlated bigram LSA model using efficient variational EM and smoothed it using the proposed fractional Kneser-Ney smoothing. The authors address the scalability issue to large training corpora by bootstrapping of bigram LSA from unigram LSA. For LM adaptation, unigram and bigram LSA are integrated into the background N-gram LM via marginal adaptation and linear interpolation respectively.", "Outcome": "Experimental results on the Mandarin RT04 test set show that applying unigram and bigram LSA together yields 6%-8% relative perplexity reduction and 2.5% relative character error rate reduction which is statistically significant compared to applying only unigram LSA. On the large-scale evaluation on Arabic, 3% relative word error rate reduction is achieved which is also statistically significant.", "Future Impact": "N/A"} +{"id": "206d2d53-dbaf-4a2f-810d-856309d8eb83", "Context": "Existing head reenactment systems rely on explicit pose representations, such as facial landmarks or 3D head pose parameters, which are often obtained through external pose estimation algorithms.", "Key Idea": "The authors propose a neural head reenactment system driven by a learned latent pose representation, which is capable of predicting the foreground segmentation alongside the RGB image.", "Method": "The authors train the reenactment system using only image reconstruction losses. The authors test the learned descriptors on pose-related tasks such as keypoint prediction and pose-based retrieval.", "Outcome": "The proposed system is able to reproduce mimics of the driving person and perform cross-person reenactment.", "Future Impact": "The learned descriptors can be used for other pose-related tasks, such as keypoint prediction and pose-based retrieval."} +{"id": "24e29617-a320-450a-aaa5-19d8700d74b7", "Context": "The need for an efficient parallel distributed algorithm for matrix completion that can handle large-scale data and is scalable.", "Key Idea": "The authors propose a new decentralized algorithm called NOMAD (Non-locking, stOchastic Multi-machine algorithm for Asynchronous and Decentralized matrix completion) that uses asynchronous, non-blocking communication between processors and asynchronously transfers the ownership of a variable between processors in a decentralized fashion.", "Method": "The authors perform extensive empirical evaluation of NOMAD on commodity hardware and on a HPC cluster in both multi-core and distributed memory settings.", "Outcome": "NOMAD outperforms synchronous algorithms that require explicit bulk synchronization after every iteration, as shown in the empirical evaluation.", "Future Impact": "NOMAD can potentially be used in various applications that require efficient and scalable matrix completion, such as in recommendation systems, image and video processing, and sensor networks."} +{"id": "29dd9fd3-6c98-4e4b-b70c-0474ff361419", "Context": "Building a deformable shape model for a new species is not always possible due to the lack of 3D data, and there is no accurate shape model available for birds.", "Key Idea": "The authors propose a method to capture new species using an articulated template and images of that species, and learn a shape space that captures variation both among species and within each species from image evidence.", "Method": "The authors fit the articulated template to each training sample and use a low-dimensional embedding to learn a 3D shape space that better reflects the phylogenetic relationships among birds than learned perceptual features.", "Outcome": "The authors learn models of multiple species from the CUB dataset, and contribute new species-specific and multi-species shape models that are useful for downstream reconstruction tasks.", "Future Impact": "N/A"} +{"id": "2a220303-8653-497f-b2b5-c829583c2714", "Context": "Counting and sampling directed acyclic graphs (DAGs) from a Markov equivalence class are fundamental tasks in graphical causal analysis. These tasks have been considered open problems in the area.", "Key Idea": "The authors present polynomial-time algorithms for counting and sampling DAGs from a Markov equivalence class, solving a long-standing open problem.", "Method": "The authors perform experiments comparing their algorithms to state-of-the-art methods.", "Outcome": "The experimental results show that the proposed algorithms significantly outperform state-of-the-art methods.", "Future Impact": "N/A"} +{"id": "2b5cc037-4841-4fb1-85ff-673230198be1", "Context": "Matrix completion is the problem of recovering an incomplete matrix of rank r with columns arriving online over time. This problem is widely applied to recommendation systems, computer vision, and system identification. Prior work in this area has focused on developing algorithms for the noiseless case, but there is a need for algorithms that can tolerate a large amount of noise.", "Key Idea": "The authors propose two algorithms for matrix completion that can tolerate noise and are provable under two realistic noise models. The algorithms are adaptive and perform well experimentally in both synthetic and real-world datasets.", "Method": "The authors present an algorithm for bounded deterministic noise that returns a matrix of small error with sample complexity almost as small as the best prior results in the noiseless case. The authors also present an algorithm for sparse random noise that recovers an μ0-incoherent matrix with high probability and small sample complexity.", "Outcome": "The first algorithm achieves strong guarantee under bounded deterministic noise and the second algorithm achieves strong guarantee under sparse random noise. Both algorithms perform well experimentally in both synthetic and real-world datasets.", "Future Impact": "N/A"} +{"id": "2eb66e5a-472c-4db1-b02f-47fe5eb7e71e", "Context": "Previous work on generative models for text generation rely on syntactic and semantic information such as constituency parses or paraphrase pairs. However, this information can be difficult to obtain and may not always be available.", "Key Idea": "The authors propose a generative model, called QKVAE, that exhibits disentangled latent representations of syntax and semantics without relying on any external syntactic or semantic information, but instead uses the inductive bias found in attention-based architectures such as Transformers.", "Method": "The authors use Attention in the decoder of QKVAE to read latent variables where one latent variable infers keys while another infers values. The authors run experiments on latent representations and syntax/semantics transfer to show that QKVAE displays clear signs of disentangled syntax and semantics.", "Outcome": "The experiments show that QKVAE displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer.", "Future Impact": "N/A"} +{"id": "31ab88b5-e66f-4b69-98b6-7a470dce9875", "Context": "Conditional neural networks play an important role in a number of sequence-to-sequence modeling tasks, including personalized sound enhancement (PSE), speaker dependent automatic speech recognition (ASR), and generative modeling such as text-to-speech synthesis. In conditional neural networks, the output of a model is often influenced by a conditioning vector, in addition to the input. Common approaches of conditioning include input concatenation or modulation with the conditioning vector, which comes at a cost of increased model size.", "Key Idea": "The authors propose a novel approach of neural network conditioning by learning intermediate layer activations based on the conditioning vector.", "Method": "The authors systematically explore and evaluate the proposed conditioning method on the task of PSE and personalized ASR in single speaker scenarios.", "Outcome": "The proposed method is able to produce conditional models with comparable or better quality, while decreasing model sizes.", "Future Impact": "The proposed conditioning method has a broad applicability across a number of application domains and can make the models ideal candidates for resource-efficient on-device deployment."} +{"id": "36d7073a-d006-4a4a-850f-ca7a3f0ca19b", "Context": "Many important problems can be framed as learning from graph data, but there is a lack of a general framework for learning convolutional neural networks for arbitrary graphs with discrete and continuous node and edge attributes.", "Key Idea": "The authors propose a framework for learning convolutional neural networks for arbitrary graphs, by extracting locally connected regions from graphs.", "Method": "The authors demonstrate the proposed framework on established benchmark data sets and compare the learned feature representations with state-of-the-art graph kernels.", "Outcome": "The learned feature representations are competitive with state-of-the-art graph kernels and their computation is highly efficient.", "Future Impact": "N/A"} +{"id": "370141c7-e1bb-4010-9938-efcad6cf2e62", "Context": "Innovations in neural architectures have fostered significant breakthroughs in language modeling and computer vision, but these novel architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized.", "Key Idea": "The authors propose GradInit, an automated and architecture agnostic method for initializing neural networks. GradInit adjusts the norm of each network layer so that a single step of SGD or Adam with prescribed hyperparameters results in the smallest possible loss value.", "Method": "The authors introduce a scalar multiplier variable in front of each parameter block and optimize these variables using a simple numerical scheme.", "Outcome": "GradInit accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. It also improves the stability of the original Transformer architecture for machine translation, enabling training it without learning rate warmup using either Adam or SGD under a wide range of learning rates and momentum coefficients.", "Future Impact": "N/A"} +{"id": "37da719b-8db9-4523-8a4a-3f2eef238978", "Context": "Fragile watermarking methods often use authentication bits to detect tampering, but this can lead to false positives and negatives. Robust spread spectrum based watermarking schemes have been proposed but do not allow for image recovery.", "Key Idea": "The authors propose a fragile watermarking method using self-embedding for tampered image recovery that does not use authentication bits, and combines robust spread spectrum based watermarking, block based embedding, and DCT based compression.", "Method": "The authors perform simulations to test the recovery performance of the proposed watermarking method.", "Outcome": "The proposed watermarking method is able to recover tampered images with high fidelity.", "Future Impact": "The proposed watermarking method could be used in applications where image integrity needs to be ensured and tampered images need to be recovered."} +{"id": "3ac464c2-2214-4bf2-a6b6-03da2498cb03", "Context": "Query scheduling in database management systems has received renewed attention due to the rise of the DaaS model for database deployment, with most research focusing on scheduling algorithms.", "Key Idea": "This paper investigates the use of histograms describing the distribution of likely query execution times as input to the query scheduler and proposes a novel distribution-based scheduling algorithm called Shepherd.", "Method": "The authors perform extensive experimentation with both synthetic and TPC workloads to compare the performance of Shepherd with state-of-the-art point-based methods.", "Outcome": "The results show that Shepherd substantially outperforms state-of-the-art point-based methods.", "Future Impact": "N/A"} +{"id": "3b076d81-3ed0-4d57-84e8-4145f67052bd", "Context": "Referring expressions and other object descriptions can lead to unwanted conversational implicatures if they do not follow the Local Brevity, No Unnecessary Components, and Lexical Preference preference rules.", "Key Idea": "The authors propose a polynomial time generation algorithm that incorporates the Local Brevity, No Unnecessary Components, and Lexical Preference preference rules.", "Method": "The authors compare the computational complexity of incorporating the preference rules into a generation algorithm against alternative formalizations of conversational implicature.", "Outcome": "The authors find that incorporating the preference rules into a generation algorithm is polynomial time, while some alternative formalizations of conversational implicature make the generation task NP-Hard.", "Future Impact": "N/A"} +{"id": "3b6e7572-2ef8-4565-a3b7-301a3fd38acd", "Context": "Backward locking and update locking are well-known sources of inefficiency in backpropagation that prevent from concurrently updating layers. Several works have suggested using local error signals to train network blocks asynchronously to overcome these limitations.", "Key Idea": "The authors propose a differentiable search algorithm named SEDONA to automate the process of finding the best configuration for local training, including how to decouple network blocks and which auxiliary networks to use for each block.", "Method": "The authors perform experiments on VGG and ResNet variants with CIFAR-10, Tiny-ImageNet and ImageNet datasets. They compare the performance of the proposed algorithm to end-to-end backpropagation and other state-of-the-art greedy-learning methods.", "Outcome": "The proposed algorithm can consistently discover transferable decoupled architectures for VGG and ResNet variants. It significantly outperforms the ones trained with end-to-end backpropagation and other state-of-the-art greedy-learning methods in CIFAR-10, Tiny-ImageNet and ImageNet. It also reports up to 2× speedup over backpropagation in total training time.", "Future Impact": "N/A"} +{"id": "4164fd9a-7ab8-4013-b416-a205231f10f2", "Context": "Plan recognition is the problem of inferring the goals and plans of an agent after observing its behavior. Recent approaches use slightly modified planning algorithms to solve this problem without a plan library, assuming actions have deterministic effects and both agent and observer have complete information about the initial state.", "Key Idea": "This paper extends the approach to probabilistic plan recognition by defining the probability of a partially observed execution given a goal in terms of the cost difference of achieving the goal under two conditions: complying with the observations, and not complying with them.", "Method": "The authors use two calls to a classical planner to compute the cost difference and the posterior goal probabilities, without modifying the planner in any way.", "Outcome": "The authors consider a number of examples to illustrate the quality, flexibility, and scalability of the approach.", "Future Impact": "N/A"} +{"id": "432f5702-7b7d-4995-812e-40925f1a18dd", "Context": "Multiple output prediction is a central problem in machine learning, where y is high dimensional and x is either low or high dimensional. The one vs. all approach and several sophisticated multiple output prediction methods have been used.", "Key Idea": "The authors propose a new method for multiple output prediction that involves selecting a small subset of yL dimensions of y and modeling (i) x → yL and (ii) yL → y.", "Method": "The authors perform experiments on multilabel classification and multivariate regression datasets using the proposed method.", "Outcome": "The proposed method outperforms the one vs. all approach as well as several sophisticated multiple output prediction methods.", "Future Impact": "N/A"} +{"id": "438caf1c-5c7e-4283-a3cd-bbab302df185", "Context": "Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively.", "Key Idea": "The authors propose an acquisition function, CAL (Contrastive Active Learning), that selects contrastive examples, i.e. data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods.", "Method": "The authors compare the proposed acquisition function, CAL, with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. The authors also conduct an extensive ablation study of their method and analyze all actively acquired datasets.", "Outcome": "The proposed acquisition function, CAL, performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data.", "Future Impact": "N/A"} +{"id": "45b76955-9670-4664-939c-f5a61eb597df", "Context": "The state of the art in auditing group fairness in ranked lists assumes a logarithmic loss in importance as a function of the rank, and does not account for varying user behaviors or non-binary protected attributes.", "Key Idea": "The authors propose a novel metric for auditing group fairness in ranked lists that models user attention through parametrization and allows non-binary protected attributes.", "Method": "The authors perform three simulated fairness audits using the proposed metric to show that determining fairness of a ranked output necessitates knowledge of the end-users of the particular service.", "Outcome": "The proposed metric is able to better address the human factors inherent in the problem by measuring the whole sociotechnical system, consisting of a ranking algorithm and individuals using it.", "Future Impact": "N/A"} +{"id": "48bacac1-1ca9-4be8-90e6-470596de0e26", "Context": "Recognizing polarity in sentiment analysis requires a list of polar words and phrases. Many studies have investigated (semi-) unsupervised methods of learning polarity of words and phrases.", "Key Idea": "The authors propose a method of building a sentiment analysis lexicon by using structural clues to extract polar sentences from Japanese HTML documents and developing these structural clues to achieve high precision.", "Method": "The authors use a massive collection of HTML documents and extract polar sentences using structural clues. These polar sentences are then used to build a lexicon for sentiment analysis.", "Outcome": "N/A", "Future Impact": "The proposed method of building a sentiment analysis lexicon can be used for other languages and can be applied to other NLP tasks such as opinion mining and subjectivity analysis."} +{"id": "49309d5a-5959-4f8f-ae30-9fd2350f0cbc", "Context": "Existing compression techniques for deep convolutional neural networks (CNNs) struggle to be computationally friendly, despite excelling at reducing model sizes. This is due to the enormous amount of memory and compute resources required by CNNs.", "Key Idea": "The authors present a novel quantization strategy called focused quantization based on power-of-two values that exploits the weight distributions after fine-grained pruning. The method dynamically discovers the most effective numerical representation for weights in layers with varying sparsities.", "Method": "The authors replace multiplications in quantized CNNs with bit-shift operations for efficient inference and couple it with lossless encoding to build a compression pipeline. The authors evaluate the proposed method on ResNet-50 and ResNet-18.", "Outcome": "The proposed method achieves a 18.08x compression ratio (CR) in ResNet-50 with only 0.24% loss in top-5 accuracy, outperforming existing compression methods. The authors found that a fully compressed ResNet-18 is not only higher in CR and top-5 accuracy but also more hardware efficient.", "Future Impact": "N/A"} +{"id": "4e6f8004-9384-4c5c-8d7f-265410a290df", "Context": "Deep learning methods have been proposed for completing partial data from shape acquisition setups, but these methods only complete the partial shape with a single output, ignoring the ambiguity when reasoning the missing geometry.", "Key Idea": "The authors pose a multimodal shape completion problem, in which they seek to complete the partial shape with multiple outputs by learning a one-to-many mapping. The authors develop a method that completes the partial shape via conditional generative modeling, without requiring paired training data.", "Method": "The authors evaluate the approach on several datasets that contain varying forms of shape incompleteness, and compare among several baseline methods and variants of their methods qualitatively and quantitatively.", "Outcome": "The proposed method demonstrates the merit of completing partial shapes with both diversity and quality.", "Future Impact": "N/A"} +{"id": "4fdcceeb-f50f-4c4f-8b92-5985498114f8", "Context": "Recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images is a challenging problem. The object undergoes uniform expansion, three-dimensional shearing about an unknown point in space, and rigid motion. Feature correspondence over multiple frames is assumed.", "Key Idea": "The authors propose a novel solution to the problem of recovering the three-dimensional motion of a non-rigid object from a sequence of stereo images using algebraic geometry, the commutative algebra software package MACAULAY, and the Fortran polynomial continuation program POLSYS.", "Method": "The authors reduce the problem of recovering the three-dimensional motion uniquely to the solution of a set of homogeneous polynomial equations. The authors use the MACAULAY and POLSYS software packages to solve the polynomial equations.", "Outcome": "The proposed solution is shown to determine the motion uniquely with only two (stereo) snapshots and four points correspondence.", "Future Impact": "N/A"} +{"id": "532e797a-4b72-488a-80e4-03713d3c8435", "Context": "The traditional idea of using linear low-order or low-rank shape model for the task of Non-Rigid Structure-from-Motion (NRSfM) has been used. However, this method may not be effective for long monocular video sequences observing a non-rigid object performing recurrent and possibly repetitive dynamic action.", "Key Idea": "The authors propose a new method for NRSfM that exploits the property of shape recurrency, which is a generalized rigidity, and reduces NRSfM problems to rigid ones provided that certain recurrency condition is satisfied.", "Method": "The authors develop efficient algorithms for automatic recurrency detection, as well as camera view clustering via a rigidity-check to implement this idea as a practical approach.", "Outcome": "Experiments on both simulated sequences and real data demonstrate the effectiveness of the method.", "Future Impact": "This paper offers a novel perspective on rethinking structure-from-motion, and the authors hope it will inspire other new problems in the field."} +{"id": "56992082-e04e-4a8b-a985-abfea27fc2e0", "Context": "Dynamic network pruning is a method to achieve runtime acceleration by dynamically determining the inference paths based on different inputs. However, previous methods directly generate continuous decision values for each weight channel, which cannot reflect a clear and interpretable pruning process.", "Key Idea": "The authors propose a method that explicitly models the discrete weight channel selections, which encourages more diverse weights utilization and achieves more sparse runtime inference paths. The authors also propose a novel adversarial example detection algorithm by discriminating the runtime decision features.", "Method": "The authors perform experiments on CIFAR10 and ImageNet datasets. They observe the differences in the layerwise decisions between normal and adversarial examples. They also apply the proposed dynamic network to adversarial example detection.", "Outcome": "The proposed dynamic network achieves higher prediction accuracy under the similar computing budgets on CIFAR10 and ImageNet datasets compared to traditional static pruning methods and other dynamic pruning approaches. The proposed adversarial detection algorithm can significantly improve the state-of-the-art detection rate across multiple attacks.", "Future Impact": "The proposed method can help to build an interpretable and robust model for adversarial example detection."} +{"id": "58279154-e623-46d2-a431-cc409e094e2c", "Context": "Previous approaches to analogy-driven theorem proving are limited and not cognitively adequate. Analogy is typically used to prove theorems by finding similarities between two problems.", "Key Idea": "The authors propose a model of analogy-driven theorem proving that works at the level of proof-plans and uses a source proof-plan to guide the construction of a proof-plan for the target problem, including a reformulation of the source proof-plan.", "Method": "The authors processed several well-known theorems using their analogy-driven proof-plan construction model.", "Outcome": "The proposed model was able to prove several theorems that could not be proven analogically by previous approaches.", "Future Impact": "N/A"} +{"id": "5a3da6ef-67b0-41bc-a994-fc5ff455a27b", "Context": "Online reviews play a crucial role in electronic commerce. However, pervasive spam reviews can mislead customers and defame decent stores. Existing methods did not examine a great portion of singleton reviews which can almost determine a store's rating and impression.", "Key Idea": "The authors propose to detect spam attacks via unusually correlated temporal patterns. They identify and construct multidimensional time series based on aggregate statistics to depict and mine such correlations and propose a hierarchical algorithm to robustly detect time windows where attacks are likely to have happened.", "Method": "The authors conduct experiments to detect singleton review attacks by mapping the problem to an abnormally correlated pattern detection problem.", "Outcome": "The proposed method is effective in detecting singleton review attacks. The authors discover that singleton review is a significant source of spam reviews and largely affects the ratings of online stores.", "Future Impact": "N/A"} +{"id": "5b47ca6b-ff9b-429c-adb1-ad9a171eea0e", "Context": "The problem of building style-adapted maximum entropy language models for speech recognition is addressed, where a large corpus of written language data and a small corpus of speech transcripts are given.", "Key Idea": "The authors investigate a recently proposed Bayesian adaptation method for building style-adapted maximum entropy language models.", "Method": "The authors perform experiments comparing the proposed Bayesian adaptation method to linear interpolation.", "Outcome": "The proposed Bayesian adaptation method outperforms linear interpolation in the experiments.", "Future Impact": "N/A"} +{"id": "5b9f94f9-d93f-455d-a110-007ad67ada6d", "Context": "Existing specification languages for tree-based grammars fail to adequately support identifier management.", "Key Idea": "The authors propose XMG (eXtensible Meta-Grammar) as a solution to the problem, which provides a sophisticated treatment of identifiers.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "XMG can potentially be used in various natural language processing tasks that require tree-based grammars and identifier management."} +{"id": "5c268324-e160-489d-9722-6b59d5e3471b", "Context": "The abstract discusses the problem of lack of accessibility on the web, particularly for marginalized groups such as visually-impaired users.", "Key Idea": "The authors propose a new approach for achieving web accessibility through self-interested web designers and strategic knowledge acquisition mechanisms.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "This approach may lead to a more accessible web, particularly for visually-impaired users, by incentivizing web designers to prioritize accessibility and providing them with the necessary knowledge to do so through mechanisms for eliciting knowledge from marginalized groups."} +{"id": "5d87de73-77d4-4efd-b8e1-d7561b13f69f", "Context": "The task of preposition sense disambiguation is a challenge in NLP, and previous systems have struggled to achieve high accuracy.", "Key Idea": "The authors propose a supervised classification approach that uses linguistically motivated features derived from both sides of the preposition and utilizes the phrase structure.", "Method": "The authors evaluate their system on the SemEval 2007 Preposition Sense Disambiguation datasets and compare its results to those of the systems participating in the workshop. They test the system with five different classifiers.", "Outcome": "The proposed system outperforms the best system in the SemEval task, achieving an increased accuracy.", "Future Impact": "N/A"} +{"id": "5dccca98-2b58-47e3-9b8f-3b1888aa3976", "Context": "The heterogeneity of today's IT environments and the increasing demands from mobile users are major obstacles for the creation of real-time data warehouses.", "Key Idea": "The authors propose a new middleware paradigm called Space-based computing that offers a level of abstraction superior to conventional middleware solutions and seamless integration of mobile devices using open standards.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The proposed Space-based computing paradigm has the potential to mobilize heterogeneous database and realize the real-time data warehouse vision."} +{"id": "5e1f387c-d883-4d1f-8397-e4a533a3387b", "Context": "Hyperspectral imaging is used in various fields but it is affected by different sources of degradation, and the lack of accurate ground-truth makes restoration tasks challenging. Training deep neural networks for restoration is difficult in contrast to traditional RGB imaging problems.", "Key Idea": "The authors propose a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data.", "Method": "The authors perform experiments on various denoising benchmarks to show the computational efficiency and the performance of the proposed method.", "Outcome": "The proposed method is computationally efficient and significantly outperforms the state of the art on various denoising benchmarks.", "Future Impact": "N/A"} +{"id": "68f8d058-1403-4066-b3d3-a8a2836b35e1", "Context": "Imperfect vector training labels with registration uncertainty are common in applications such as streamline classification on Earth imagery or tissue segmentation on medical imagery. Existing research often focuses on uncertainty in label class semantics or characterizes label registration uncertainty at the pixel level.", "Key Idea": "The authors propose a deep learning framework that can quantify and reduce the registration uncertainty of training labels as well as train neural network parameters simultaneously.", "Method": "The authors propose a registration-uncertainty-aware loss function and design an iterative uncertainty reduction algorithm by re-estimating the posterior of true vector label locations distribution based on a Gaussian process.", "Outcome": "Evaluations on real-world datasets in National Hydrography Dataset refinement show that the proposed approach significantly outperforms several baselines in the registration uncertainty estimations performance and classification performance.", "Future Impact": "N/A"} +{"id": "69aacc53-6730-4db0-b420-9a45b96a642e", "Context": "The problem of finding the configuration of a collection of geometric bodies to satisfy a set of given constraints has been suggested to be solved efficiently by symbolically reasoning about geometry using a degrees of freedom analysis and plan fragments.", "Key Idea": "This paper presents a method for automatically synthesizing plan fragments using first principles about geometric bodies, actions, and topology.", "Method": "The authors use first principles about geometric bodies, actions, and topology to synthesize plan fragments.", "Outcome": "The authors show that the synthesized plan fragments can efficiently find the configuration of a collection of geometric bodies to satisfy a set of given constraints.", "Future Impact": "The proposed method for automatically synthesizing plan fragments has the potential to improve the efficiency and accuracy of geometric reasoning and constraint satisfaction systems."} +{"id": "6dc39f88-d613-4ec0-b70d-d5daa6f3643c", "Context": "Data analysis systems typically execute aggregation queries and produce results that may not be exact, but rather an approximation of the true value. These approximations may not have deterministic bounds.", "Key Idea": "The authors propose BitGourmet, a novel data analysis system that supports deterministic approximate query processing (DAQ), which executes aggregation queries and produces deterministic bounds that are guaranteed to contain the true value.", "Method": "The authors demonstrate BitGourmet by dividing each column vertically, bit-by-bit, and evaluating queries on subsets of these bit vectors. The system uses a specialized query processing engine and a scenario-specific query optimizer that relies on quality and cost models to decide the optimal bit selection and execution plan.", "Outcome": "The proposed system realizes a trade-off between result quality and execution time, making data analysis more interactive.", "Future Impact": "N/A"} +{"id": "6de74297-fb80-448f-b7ae-41f8d9701044", "Context": "Previous literatures use different imaging models to describe central catadioptric cameras and fisheye cameras separately.", "Key Idea": "The authors present a unified imaging model for both central catadioptric cameras and fisheye cameras.", "Method": "The authors show that the proposed unified model can cover some existing models for fisheye cameras and fit well for many actual fisheye cameras used in previous literatures. The authors perform experiments of calibration from some central catadioptric and fisheye images.", "Outcome": "The authors confirm the validity and usefulness of the proposed unified model by experimental results.", "Future Impact": "The proposed unified model enables the use of existing calibration methods for central catadioptric cameras to be directly applied to fisheye cameras and makes the metric calibration from single fisheye image only using projections of lines possible."} +{"id": "6ecf725b-661e-4897-8169-22d71826d0e8", "Context": "Existing Open Information Extraction (OIE) systems lack adaptability to different OIE tasks, which have different requirements.", "Key Idea": "The authors propose a new adaptable and efficient OIE system, OIE@OIA, which follows the methodology of Open Information eXpression (OIX) and parses a sentence to an Open Information Annotation (OIA) Graph.", "Method": "The authors implement an end-to-end OIA generator by annotating a dataset and designing an efficient learning algorithm for the complex OIA graph. They adapt the OIE@OIA system to accomplish three popular OIE tasks.", "Outcome": "The proposed OIE@OIA system achieves new SOTA performances on three popular OIE tasks, showing the great adaptability of the system. It also shows a significant advantage in terms of efficiency as it needs much fewer training samples compared to other end-to-end OIE baselines.", "Future Impact": "N/A"} +{"id": "73e353a8-e0d6-466f-af93-6fccf38fcb18", "Context": "Previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modelling the inherent information flows at the turn level.", "Key Idea": "The authors propose to discover information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer. They then introduce a new approach that learns to predict reasoning paths over this semantic graph.", "Method": "The authors construct a semantic graph based on lexical components in each question and answer. They then introduce a path prediction model that predicts a path from the current turn through past dialogue turns that contain additional visual cues to answer the current question. The reasoning model sequentially processes both visual and textual information through this reasoning path.", "Outcome": "The experimental results demonstrate the effectiveness of the proposed method and provide insights on how models use semantic dependencies in a dialogue context to retrieve visual cues.", "Future Impact": "N/A"} +{"id": "754f792b-fcf3-42f4-bb2c-5c1dcce21d7a", "Context": "The abstract argues for the use of formal meaning representations for natural language and identifies the problem of identifying the possible antecedents of anaphoric expressions as a challenge for such representations.", "Key Idea": "The authors propose the use of a specific structural property of formal meaning representations that facilitates the identification of possible antecedents of anaphoric expressions, allowing for a richer understanding of anaphora.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The proposed approach to formal meaning representations has the potential to significantly improve the ability to understand and interpret anaphora in natural language, with potential applications in natural language processing and understanding."} +{"id": "7b0671d7-c8f6-4e81-828b-c73958a6a63a", "Context": "There is a lack of interactive exhibits about medieval music that focus on education, entertainment, and historic authenticity.", "Key Idea": "The authors describe the design experience and lessons learned from creating an interactive exhibit about medieval music that relies on audio as its only feedback channel and focuses on educational value, entertainment aspects, and historic authenticity.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The paper provides valuable insights for designers and researchers in the field of interactive exhibits and education on how to balance educational value, entertainment aspects, and historic authenticity."} +{"id": "7b21425c-a2b7-4d19-b030-a8350b2a7a80", "Context": "Seq2Seq models are a popular choice for set generation tasks such as entity typing and dialogue emotion tagging. However, these models treat a set as a sequence and do not fully leverage its key properties, namely order-invariance and cardinality.", "Key Idea": "The authors propose a novel algorithm for effectively sampling informative orders over the combinatorial space of label orders. The authors jointly model the set cardinality and output by prepending the set size and taking advantage of the autoregressive factorization used by Seq2Seq models.", "Method": "The authors train a Seq2Seq model on augmented data with signals of order-invariance and cardinality, without any additional annotations.", "Outcome": "Training a Seq2Seq model on this augmented data gets an average relative improvement of 20% on four benchmark datasets across various models: BART, T5, and GPT-3.", "Future Impact": "N/A"} +{"id": "7c065e41-7c2e-430e-a918-6a776037bf50", "Context": "The paper summarizes the discussion of the SIGMOD 2005 panel on Databases and Information Retrieval: Rethinking the Great, which aimed to discuss whether data management systems architectures should be redesigned to merge Database (DB) and Information Retrieval (IR) technologies.", "Key Idea": "N/A", "Method": "N/A", "Outcome": "The panel had very high attendance and generated lively discussions.", "Future Impact": "N/A"} +{"id": "7d850699-212d-466b-976f-0afed0653fef", "Context": "The Helping Our Own (HOO) 2011 Shared Task focused on identifying determiner and preposition errors in non-native English essays. The HOO 2012 Shared Task extended this to include the Cambridge Learner Corpus FCE Dataset.", "Key Idea": "The authors extend their n-gram-based data-driven prediction approach and incorporate word clustering to improve error detection. They also develop a missing determiner detector.", "Method": "The authors perform experiments on the HOO 2012 Shared Task dataset, focusing on three error categories: missing determiner, incorrect determiner, and incorrect preposition.", "Outcome": "The proposed system improved the approach by incorporating word clustering and developing a missing determiner detector.", "Future Impact": "N/A"} +{"id": "802a5b78-a022-4d38-bfb3-f28eee4ef89a", "Context": "Neural network based dependency parsing has been recently studied, which can effectively alleviate the problems of data sparsity and feature engineering by using dense features. However, it is still a challenge to model the complicated syntactic and semantic compositions of the dense features in neural network based methods.", "Key Idea": "The authors propose two heterogeneous gated recursive neural networks, Tree-GRNN and DAG-GRNN, to model the feature combinations for the trees in stack and the feature combinations of the nodes whose dependency relations have not been built yet, respectively.", "Method": "The authors integrate Tree-GRNN and DAG-GRNN to automatically learn the compositions of the dense features for transition-based dependency parsing.", "Outcome": "Experiment results on two prevalent benchmark datasets (PTB3 and CTB5) show the effectiveness of the proposed model.", "Future Impact": "N/A"} +{"id": "80698baf-89cb-4a50-9f5c-0c74151b798b", "Context": "Peer grading in MOOCs (Massive Open Online Courses) is a common practice, where students grade their peers' assignments. However, the motivation of students to perform well during peer grading is not well understood.", "Key Idea": "The authors propose a study to examine how students in a MOOC might be motivated to do a better job during peer grading.", "Method": "The authors conduct a controlled study involving more than one thousand students in a popular MOOC. They ask two specific questions: (1) When a student knows that his or her own peer grading efforts are being examined by peers, does this knowledge alone tend to motivate the student to do a better job when grading assignments? And (2) when a student not only knows that his or her own peer grading efforts are being examined by peers, but he or she is also given a number of other peer grading efforts to evaluate (so the peer graders see how other peer graders evaluate assignments), do both of these together tend to motivate the student to do a better job when grading assignments?", "Outcome": "The study finds strong statistical evidence that ``grading the graders'' tends to increase the quality of peer grading.", "Future Impact": "N/A"} +{"id": "816707fd-9214-4435-ac40-b2655e55c9d0", "Context": "Pretrained language models have achieved superhuman performance on many benchmarks, creating a need for harder tasks.", "Key Idea": "The authors introduce CoDA21, a benchmark that measures natural language understanding capabilities of pretrained language models by evaluating their ability to align definitions with contexts.", "Method": "The authors evaluate the performance of pretrained language models on CoDA21, a benchmark that requires a deep understanding of contexts and definitions, including complex inference and world knowledge.", "Outcome": "The authors find that there is a large gap between human and pretrained language model performance on CoDA21, suggesting that CoDA21 measures an aspect of natural language understanding that is not sufficiently covered in existing benchmarks.", "Future Impact": "CoDA21 could be used to improve pretrained language models by providing a challenging benchmark that measures an aspect of natural language understanding that is not sufficiently covered in existing benchmarks."} +{"id": "8737b031-f77c-4f32-8a68-4be9b0c9ecf8", "Context": "Training a deep network to perform semantic segmentation requires large amounts of labeled data. Researchers have investigated the use of synthetic data, which can be labeled automatically, but a network trained on synthetic data performs relatively poorly on real images.", "Key Idea": "The authors propose a new approach to handle synthetic images that does not require seeing any real images at training time. The authors build on the observation that foreground and background classes are not affected in the same manner by the domain shift.", "Method": "The authors perform experiments on Cityscapes and CamVid datasets with models trained on synthetic data only, using a detection-based manner to handle the foreground classes.", "Outcome": "The authors' experiments evidence the effectiveness of their approach on Cityscapes and CamVid datasets.", "Future Impact": "N/A"} +{"id": "8ae36735-e4ac-48da-bd2b-5538a6a00a73", "Context": "Relational machine learning (RML) is a framework for predicting preferences in large scale networks by jointly modeling user labels and relational structure. However, existing RML approaches do not fully utilize unlabeled instances and have expensive collective inference procedures.", "Key Idea": "This paper proposes a method to overcome these limitations by implementing a maximum entropy constraint on the inference step, correcting bias, and outlining a massively scalable variational inference algorithm for large scale relational network domains.", "Method": "The authors analyze the effect of full semi-supervised RML and implement a maximum entropy constraint on the inference step. They also outline a massively scalable variational inference algorithm for large scale relational network domains and extend it to incorporate the maximum entropy constraint.", "Outcome": "The proposed method demonstrates improvement over a variety of baselines on seven real-world datasets, including large scale networks with over five million edges.", "Future Impact": "N/A"} +{"id": "8aedb046-2f51-4229-bc19-ea6db98355cb", "Context": "Researchers used to describe the inter-connectivity among websites with a HostGraph and adopted the random walk model in the HostGraph to get the ranks of websites. However, the random walk over such a HostGraph is not reasonable because it is not in accordance with the browsing behavior of web surfers.", "Key Idea": "The authors propose a novel method named AggregateRank, which mathematically proves that the probability of visiting a website by the random web surfer should be equal to the sum of the PageRank values of the pages inside that website and can approximate the sum of PageRank accurately with a lower computational complexity than PageRank.", "Method": "The authors mathematically prove the AggregateRank method and evaluate it through theoretical analysis and experiments.", "Outcome": "Both theoretical analysis and experimental evaluation show that AggregateRank is a better method for ranking websites than previous methods.", "Future Impact": "N/A"} +{"id": "927df1bd-273a-4088-8c56-2e79cac37072", "Context": "k-means is a popular data processing algorithm, and a proper initialization of k-means is crucial for obtaining a good final solution. The k-means++ initialization algorithm obtains an initial set of centers that is provably close to the optimum solution, but it is sequential in nature.", "Key Idea": "The authors propose a new algorithm called k-means|| which obtains a nearly optimal solution after a logarithmic number of passes in parallel, outperforming k-means++ in both sequential and parallel settings.", "Method": "The authors prove that the proposed initialization algorithm k-means|| obtains a nearly optimal solution after a logarithmic number of passes and then show that in practice a constant number of passes suffices.", "Outcome": "Experimental evaluation on real-world large-scale data demonstrates that k-means|| outperforms k-means++ in both sequential and parallel settings.", "Future Impact": "N/A"} +{"id": "9292bc08-a70f-4adb-a6d7-920728893a39", "Context": "The paper is based on propositional argumentation systems and monotonic logics. An assumption-based argumentation framework is a special case of this.", "Key Idea": "The authors introduce a number of propositional argumentation systems obtained by gradually extending the underlying language and associated monotonic logics, and show a stronger argumentation system in a full classical language to be equivalent to a system of causal reasoning.", "Method": "The authors use a system of causal reasoning to extend the underlying language and associated monotonic logics.", "Outcome": "The authors establish a correspondence between the argumentation system and the system of causal reasoning.", "Future Impact": "The implications of this correspondence for the respective nonmonotonic theories of argumentation and causal reasoning are discussed."} +{"id": "93bf12a1-2174-43e3-9cd4-c2b8aeed2f93", "Context": "Recent approaches in unsupervised domain mapping involve learning a mapping between two unmatched datasets A and B, by learning both GAB and GBA simultaneously.", "Key Idea": "The authors propose a new method for unsupervised domain mapping that learns GAB without learning GBA, by maintaining the distance between a pair of samples.", "Method": "The authors perform experiments to learn a mapping that maintains the distance between a pair of samples and between different parts of the same sample before and after mapping.", "Outcome": "The proposed method allows for one-sided mapping learning and leads to preferable numerical results over existing circularity-based constraint.", "Future Impact": "N/A"} +{"id": "9748b061-590a-4602-9015-a68e879ffced", "Context": "HCI researchers are looking into using liquid-based materials to create novel interfaces, but existing liquid-based smart material printing systems have limitations such as low printing resolution, small range of printable materials, and lack of customizability.", "Key Idea": "The authors present a design strategy for HCI researchers to build and customize a liquid-based smart material printing platform, called xPrint, with off-the-shelf or easy-to-machine parts. The design includes a magnetic assembly-based modular hardware and an open-source, highly customizable software design and simulation platform.", "Method": "The authors introduce the system design in detail and demonstrate the material variability and customizability through three use cases.", "Outcome": "xPrint has a large range of printable materials from synthesized polymers to natural micro-organism-living cells with a printing resolution from 10μm up to 5mm (droplet size).", "Future Impact": "N/A"} +{"id": "977698a6-56ef-4e57-94c5-b3a68a451a80", "Context": "Emotion cause extraction aims to identify the reasons behind a certain emotion expressed in text. It is a much more difficult task compared to emotion classification.", "Key Idea": "The authors propose a new approach which considers emotion cause identification as a reading comprehension task in QA and uses a new mechanism to store relevant context in different memory slots to model context information.", "Method": "The authors use deep memory networks for emotion cause extraction and evaluate the performance of their proposed approach on a recently released emotion cause dataset.", "Outcome": "The proposed method achieves the state-of-the-art performance on the emotion cause dataset, outperforming a number of competitive baselines by at least 3.01% in F-measure.", "Future Impact": "N/A"} +{"id": "9821d5f7-72b0-4841-a54f-d2af4a04ea3a", "Context": "Inverse rendering is the problem of estimating surface properties of an object from images. In the case of surface texture, it can be approximated by a linear basis. A dichromatic reflectance model is assumed and the challenge is to estimate spherical harmonic illumination coefficients and texture parameters.", "Key Idea": "The authors propose a method to estimate spherical harmonic illumination coefficients and texture parameters in a specular invariant colour subspace by solving a system of bilinear equations. They focus on the case of faces, where both shape and texture can be efficiently described by a linear statistical model.", "Method": "The authors fit a 3D morphable model to a single colour image, accounting for both non-Lambertian specular reflectance and complex illumination of the same light source colour. They solve only convex optimization problems.", "Outcome": "The authors are able to recover statistical texture model parameters with an accuracy comparable to more computationally expensive analysis-by-synthesis approaches.", "Future Impact": "N/A"} +{"id": "9999e1df-f439-4f62-bd03-337e494e9da3", "Context": "Data type encapsulation schemes typically express 'space' operations in a way that is not natural to the base language operators.", "Key Idea": "The authors propose a data type encapsulation scheme that results in a conceptual separation of operators and procedure calls in the base language, leading to a language of considerable expressive power.", "Method": "The authors have implemented the proposed scheme and provide several examples to demonstrate its use.", "Outcome": "N/A", "Future Impact": "The proposed data type encapsulation scheme has the potential to simplify the process of expressing 'space' operations in terms of base language operators, making it easier for developers to write more expressive code."} +{"id": "99bef0ca-923c-480d-a89e-9a2f27e02157", "Context": "Most real-world games and many recreational games are games of incomplete information. Over the last dozen years, abstraction has emerged as a key enabler for solving large incomplete-information games.", "Key Idea": "The authors present a method for solving large incomplete-information games by abstracting the game to generate a smaller, abstract game that is strategically similar to the original game, computing an approximate equilibrium in the abstract game and mapping the strategy back to the original game.", "Method": "The authors review key developments in the field of abstraction for solving large incomplete-information games, present algorithms for information and action abstraction and discuss reverse mapping of opponent's actions into the abstraction.", "Outcome": "N/A", "Future Impact": "The authors discuss other topics of current and future research in the field of abstraction for solving large incomplete-information games."} +{"id": "9b95b592-1562-4ef9-b0ed-e0655fadc73b", "Context": "Current methods for surface matching are limited in their ability to handle large deformations and arbitrary topology changes.", "Key Idea": "The authors propose a new method for surface matching that uses a geodesic distance evolution scheme on a 3-manifold to handle large deformations and topological changes.", "Method": "The authors set up a partial differential equation governing the propagation of surfaces at equal geodesic distance from the given original surface. Using an eulerian formulation with level-sets, it gives stable numerical algorithms for computing distance maps.", "Outcome": "The proposed method for surface matching is able to obtain matching paths as the orbits of the vector field defined as the sum of two distance maps' gradient values, and it can handle the case of large deformation and topological changes.", "Future Impact": "N/A"} +{"id": "9cd7e7e1-8893-4db6-8327-48f098187699", "Context": "Online social networks, such as Twitter, have become an important alternative information channel to traditional media during natural disasters. However, the large volume of messages can lead to information overload for end users.", "Key Idea": "The authors propose the development of an automatic classifier of tweets to filter relevant and non-relevant information during natural disasters.", "Method": "The authors use a dataset from the Chilean earthquake of 2010 to build and validate a ground truth, and test the effect of class imbalance and dimensionality reduction over 5 classifiers.", "Outcome": "The authors show that the performance of the classifiers is affected by class imbalance and dimensionality reduction, providing important considerations for building these systems.", "Future Impact": "N/A"} +{"id": "9f641e23-3886-4ac6-b65e-28db86ee48be", "Context": "Google unveiled the generalized second price (GSP) auction nearly fifteen years ago. The Vickrey-Clarke-Groves (VCG) auction would have been the proper choice according to theoretical accounts, but GSP has succeeded spectacularly.", "Key Idea": "Advertisers' preferences map to a model called value maximization, and for value maximizers, GSP is the truthful auction. This implies an axiomatization of GSP which can be applied much more broadly than the simple model for which GSP was originally designed.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The axiomatization of GSP can be applied much more broadly than the simple model for which GSP was originally designed. It can recover the folklore definition of GSP when applied to arbitrary single-parameter domains."} +{"id": "a10dafe9-6093-47f0-8429-7b62c46566ea", "Context": "Enterprise mashup scenarios often involve feeds derived from data created primarily for eye consumption, such as email, news, calendars, blogs, and web feeds. These data sources can test the capabilities of current data mashup products, as the attributes needed to perform join, aggregation, and other operations are often buried within unstructured feed text.", "Key Idea": "The authors present the integration of SystemT, an information extraction system from IBM Research, with IBM's InfoSphere MashupHub, which enables the conversion of unstructured text into structured information that can facilitate mashup operations.", "Method": "The authors demonstrate how to build domain-specific annotators with SystemT's declarative rule language, AQL, and how to use these annotators to combine structured and unstructured information in an enterprise mashup.", "Outcome": "N/A", "Future Impact": "N/A"} +{"id": "a1739057-ef00-4b01-9c26-4ab2b5d5708e", "Context": "With the increasing use of data mining tools and techniques, it is envisioned that a Knowledge Discovery and Data Mining System (KDDMS) will have to support and optimize for complex mining queries, specifically in the case of a sequence of queries and multiple simultaneous queries.", "Key Idea": "The authors present a systematic mechanism to optimize for these scenarios, targeting the class of mining queries involving frequent pattern mining on one or multiple datasets by proposing a system architecture and new algorithms.", "Method": "The authors implement and evaluate their system with both real and synthetic datasets, comparing it with systems that do not support caching or optimize for multiple queries.", "Outcome": "The experimental results show that the proposed techniques can achieve a speedup of up to a factor of 9.", "Future Impact": "N/A"} +{"id": "a5bc5ab8-70fb-4dff-824b-7606228e44a9", "Context": "N/A", "Key Idea": "EzMath is a new notation for embedding mathematical expressions in Web pages.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "EzMath has the potential to make it easier for people to add mathematical expressions to web pages."} +{"id": "a722b600-3725-4738-a47b-435aebd63e13", "Context": "Today's market for smart home devices has quickly evolved to include products that monitor, automate, and present themselves as human, which can lead to emergent problems with privacy in people's homes.", "Key Idea": "The authors propose a design philosophy for intelligent agents in the smart home that prioritizes privacy and respect for the user, as an alternative to the current ways that these devices are built.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The proposed design philosophy represents the first steps towards a more respectful future for smart home devices, and the application of this philosophy to the design of privacy empowering technologies represents a potential improvement for the field."} +{"id": "ad69e31c-2c7f-4db0-916c-3deccaab37fd", "Context": "Go is an ancient board game that has posed unique opportunities and challenges for AI and machine learning, particularly in learning a good evaluation function in a scalable way.", "Key Idea": "The authors propose a machine learning approach that focuses on learning the propensity of local patterns in Go from a library of games. The local tactical information is fed into a recursive neural network, derived from a Bayesian network architecture, to produce local outputs representing local territory ownership probabilities.", "Method": "The authors train the system using 9 × 9 amateur game data and test it on a 19 × 19 professional game data.", "Outcome": "A system trained using only 9 × 9 amateur game data performs surprisingly well on a test set derived from 19 × 19 professional game data.", "Future Impact": "The authors suggest possible directions for further improvements, such as incorporating more sophisticated features and learning from more data."} +{"id": "b0ad60d7-3a3a-42c2-acbc-fba55e708ba0", "Context": "In natural language, the meaning of a lexeme often varies due to the specific surrounding context. Computational approaches to natural language processing rely on a reliable, long-range-context-dependent representation of the meaning of each lexeme that appears in a given sentence.", "Key Idea": "The authors propose a new general technique that produces a context-dependent u0027meaningu0027 representation for a lexeme in a specific surrounding context.", "Method": "The authors perform experiments with a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases.", "Outcome": "The authors' proposed method represents the u0027meaningu0027 of a lexeme in a specific context by a list of semantically replaceable elements, which can be used to compare the u0027meaningu0027 of conceptual units in different contexts and also can serve as features for machine learning approaches to classify semantic roles and relationships.", "Future Impact": "N/A"} +{"id": "b14f2bc1-607f-4d8b-a731-b4afdf30a633", "Context": "Current explanation datasets often employ synthetic data with simple reasoning structures and cannot express more complex reasoning processes, such as the rebuttal to a reasoning step and the degree of certainty of the evidence.", "Key Idea": "The authors propose a comprehensive logical reasoning explanation form that includes three main components: (1) The condition of rebuttal that the reasoning node can be challenged; (2) Logical formulae that uncover the internal texture of reasoning nodes; (3) Reasoning strength indicated by degrees of certainty.", "Method": "The authors evaluate the current best models' performance on this new explanation form.", "Outcome": "The experimental results show that generating reasoning graphs remains a challenging task for current models, even with the help of giant pre-trained language models.", "Future Impact": "N/A"} +{"id": "b391a193-83e3-4f11-801f-1842647d626e", "Context": "Many studies have employed graph-based deep learning methods to exploit the dependencies among facial action units (AU) for the AU detection task. However, the dependencies among AUs in real-world data are often noisy and uncertain.", "Key Idea": "The authors propose an uncertain graph neural network (UGN) to learn the probabilistic mask that simultaneously captures both the individual dependencies among AUs and the uncertainties. The authors also propose an adaptive weighted loss function based on the epistemic uncertainties to adaptively vary the weights of the training samples.", "Method": "The authors conduct experiments on two benchmark datasets, BP4D and DISFA, to demonstrate the effectiveness of the proposed UGN and adaptive weighted loss function.", "Outcome": "The proposed UGN and adaptive weighted loss function achieve the state-of-the-art performance on the two benchmark datasets.", "Future Impact": "N/A"} +{"id": "b6b29c8a-7c8c-444c-b434-2ff9e166d9aa", "Context": "Offline Reinforcement Learning (RL) aims to learn policies from previously collected datasets without exploring the environment. Directly applying off-policy algorithms to offline RL usually fails due to the extrapolation error caused by the out-of-distribution (OOD) actions.", "Key Idea": "The authors propose a purely uncertainty-driven offline algorithm, Pessimistic Bootstrapping for offline RL (PBRL), that conducts uncertainty quantification via the disagreement of bootstrapped Q-functions, and performs pessimistic updates by penalizing the value function based on the estimated uncertainty. The authors further propose a novel OOD sampling method to tackle the extrapolating error.", "Method": "The authors show that such OOD sampling and pessimistic bootstrapping yields provable uncertainty quantifier in linear MDPs, thus providing the theoretical underpinning for PBRL.", "Outcome": "Extensive experiments on D4RL benchmark show that PBRL has better performance compared to the state-of-the-art algorithms.", "Future Impact": "N/A"} +{"id": "bd84a1fd-ee85-4ac9-a6ae-e534b6013506", "Context": "The abstract does not provide enough context, but in general, convolutional networks are a type of neural network commonly used for image classification and other tasks. They are often viewed from the perspective of linear operators or differential equations.", "Key Idea": "The authors propose a harmonic decomposition of convolutional networks, which are expansions into sums of elementary functions of increasing order. These elementary functions are related to the spherical harmonics.", "Method": "The authors establish harmonic decompositions of convolutional networks and use it to characterize the integral operators associated with convolutional networks.", "Outcome": "The harmonic decompositions allow the authors to obtain statistical bounds for convolutional networks.", "Future Impact": "The harmonic decomposition proposed by the authors can potentially be used for better understanding the properties of convolutional networks and improve their performance."} +{"id": "c268a190-6974-4190-8f48-db5dcbda8bc8", "Context": "The data management ecosystem faces a challenge in handling the variety of data, which comes in multiple formats such as relational and (semi-)structured data (i.e. XML). Traditional databases are limited in their ability to handle different types of data formats as they are designed for a single type of data format.", "Key Idea": "The authors propose a multi-model processing framework for relational and semi-structured data (i.e. XML) and a worst-case optimal join algorithm that can guarantee that the intermediate results are no larger than the worst-case join results.", "Method": "The authors perform experiments to compare the running time and intermediate result size of their multi-model algorithm with baseline join methods.", "Outcome": "Preliminary results show that the multi-model algorithm significantly outperforms the baseline join methods in terms of running time and intermediate result size.", "Future Impact": "N/A"} +{"id": "c37e0f9f-1654-4f1f-b812-bf6f67c0c840", "Context": "Deep Learning based AI systems are susceptible to adversarial attacks, which are small perturbations added to inputs that can cause misclassification. Developing defenses against such adversarial attacks is an active research area.", "Key Idea": "The authors propose a novel statistical approach for adversarial detection in image classification, based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class.", "Method": "The authors make use of various statistical distances such as ED, MMD for adversarial detection and analyze the performance of each metric on MNIST and CIFAR-10 datasets.", "Outcome": "The authors experimentally show that their approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size and the degree of adversarial perturbation.", "Future Impact": "N/A"} +{"id": "c3aa9543-0695-42b4-99af-d033e3912801", "Context": "The real estate industry has not attracted much attention from the KDD community despite its enormous size and prominence. This is because the real estate industry did not appreciate the value of data science methods until recently and the Data Science community was not aware of challenging real estate problems.", "Key Idea": "This tutorial provides an introduction to real estate for data scientists and outlines a spectrum of data science problems in the real estate industry, many of which are being tackled by new prop-tech companies.", "Method": "The tutorial presents concrete examples from three real estate companies: Airbnb, Cherre, and Compass, where the authors work.", "Outcome": "N/A", "Future Impact": "This tutorial could lead to more data science research being applied to the real estate industry, and more data scientists becoming interested in working on real estate problems."} +{"id": "c5c22c6f-e9e9-4b0a-ac57-d3baa100033b", "Context": "Stochastic convex optimization problems with a large number of linear constraints arise from SDP-relaxations of combinatorial problems, which involve a number of constraints that is polynomial in the problem dimension. Prior works require full passes over all constraints which can be computationally expensive.", "Key Idea": "The authors propose two novel conditional gradient-based methods for solving structured stochastic convex optimization problems with a large number of linear constraints that only process a subset of the constraints at each iteration.", "Method": "The authors' algorithms rely on variance reduction and smoothing used in conjunction with conditional gradient steps, and are accompanied by rigorous convergence guarantees.", "Outcome": "N/A", "Future Impact": "Preliminary numerical experiments are provided for illustrating the practical performance of the methods."} +{"id": "c676aecf-7468-4258-bb41-22bc1811bc3a", "Context": "Community search, or finding a connected subgraph containing given query nodes in a social network, is a fundamental problem. Most existing community search models focus on the internal cohesiveness of a community.", "Key Idea": "The authors propose a new community search model, DMCS, that finds a community with high modularity, meaning dense connections inside communities and sparse connections to nodes outside the community.", "Method": "The authors prove that the DMCS problem is NP-hard and present new algorithms that run in log-linear time to the graph size. They conduct extensive experimental studies in real-world and synthetic networks.", "Outcome": "The proposed DMCS algorithm achieves up to 8.5 times higher accuracy in terms of NMI than baseline algorithms.", "Future Impact": "N/A"} +{"id": "ca53b2c4-2912-4515-aae6-938c3f268a60", "Context": "The problem of detecting epidemic tendency by mining search logs, and the need for an algorithm that can select epidemic related queries/terms.", "Key Idea": "The authors propose an algorithm that uses click-through information to select epidemic related queries/terms, and a method to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search logs using linear regression.", "Method": "The authors perform experiments to test the effectiveness of the algorithm in finding ERTs which obtain a high correlation value with epidemic occurrences.", "Outcome": "The proposed algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. The proposed method performs better when combining different ERTs than using single ERT.", "Future Impact": "N/A"} +{"id": "ccae9338-7379-4af1-8fea-8945ba429c5c", "Context": "Object tracking is a critical and challenging problem in computer vision. More and more researchers are paying attention to applying deep learning to extract powerful features for better tracking accuracy.", "Key Idea": "The authors propose a novel triplet loss function in a Siamese network framework for object tracking, instead of using pairwise loss for training.", "Method": "The authors perform theoretical analysis by comparing gradients and back-propagation, and apply the proposed triplet loss to three real-time trackers based on Siamese network. The authors evaluate the proposed method on several popular tracking benchmarks.", "Outcome": "The proposed method achieves superior tracking performance compared to the baseline trackers and comparable accuracy to recent state-of-the-art real-time trackers.", "Future Impact": "N/A"} +{"id": "ce74316d-c5dc-47f1-b0c4-0591bc3fb4b6", "Context": "Several methods have been proposed to evaluate queries over a native XML DBMS, where the queries specify both path and keyword constraints. These methods broadly consist of graph traversal approaches and approaches based on information-retrieval style inverted lists.", "Key Idea": "The authors propose a strategy that combines the two forms of auxiliary indexes, structure indexes and inverted lists, and a query evaluation algorithm for branching path expressions based on this strategy.", "Method": "The authors perform experiments over the Niagara XML DBMS to evaluate the benefit of integrating the two forms of indexes. They also consider algorithmic issues in evaluating path expression queries when the notion of relevance ranking is incorporated.", "Outcome": "The results show that by integrating the proposed techniques with the Threshold Algorithm, instance optimal algorithms to push down top k computation are obtained.", "Future Impact": "N/A"} +{"id": "db900c02-9a35-4a83-aa19-15b763259100", "Context": "Quantum computing is a powerful computational paradigm with applications in several fields, including machine learning. Deep learning, and in particular Convolutional Neural Networks (CNN), have become essential for applications in signal processing and image recognition. Quantum deep learning remains a challenging problem as it is difficult to implement non-linearities with quantum unitaries.", "Key Idea": "The authors propose a quantum algorithm for evaluating and training deep convolutional neural networks, called QCNN, that reproduces the outputs of classical CNNs and allows for non-linearities and pooling operations.", "Method": "The authors present numerical simulations for the classification of the MNIST dataset to provide practical evidence for the efficiency of the QCNN.", "Outcome": "The QCNN reproduces completely the outputs of the classical CNN and allows for non-linearities and pooling operations. The QCNN is in particular interesting for deep networks and could allow new frontiers in the image recognition domain.", "Future Impact": "The QCNN could allow for many more convolution kernels, larger kernels, high dimensional inputs, and high depth input channels in the image recognition domain."} +{"id": "dca09f6f-b63a-42a1-9eb5-fbfa45bc6389", "Context": "Activity logs collected from wearable devices are a promising source of data for personalized exercise scheduling, workout recommendation, and heart rate anomaly detection. However, such data are heterogeneous, noisy, diverse in scale and resolution, and have complex interdependencies.", "Key Idea": "The authors propose FitRec, a context-aware LSTM-based model that captures the personalized and temporal patterns of fitness data by considering two levels of context information: context within a specific activity, and context across a user's activity history.", "Method": "The authors evaluate the proposed model on a novel dataset containing over 250 thousand workout records coupled with hundreds of millions of parallel sensor measurements and metadata. They demonstrate that the model is able to learn contextual, personalized, and activity-specific dynamics of users' heart rate profiles during exercise.", "Outcome": "The proposed model outperforms baselines on several personalized recommendation tasks.", "Future Impact": "N/A"} +{"id": "dd282632-ee41-45da-add8-d68d89c57e2d", "Context": "Classification activation map (CAM) is a crucial mechanism for weakly supervised object localization (WSOL) which uses the classification structure to generate pixel-wise localization maps. However, CAM directly uses the classifier trained on image-level features to locate objects, making it prefer to discern global discriminative factors rather than regional object cues.", "Key Idea": "The authors propose a plug-and-play mechanism called BagCAMs to better project a well-trained classifier for the localization task without refining or re-training the baseline structure. The authors adopt a proposed regional localizer generation (RLG) strategy to define a set of regional localizers and then derive them from a well-trained classifier.", "Method": "The authors conduct experiments to improve the performance of baseline WSOL methods and evaluate the proposed BagCAMs on three WSOL benchmarks.", "Outcome": "Experiments indicate that adopting the proposed BagCAMs can improve the performance of baseline WSOL methods to a great extent and obtains state-of-the-art performance on three WSOL benchmarks.", "Future Impact": "N/A"} +{"id": "ddf8f49c-342a-4cd0-8b3b-b588af08ed0d", "Context": "Soboroff and Cahan proposed a method for evaluating the performance of retrieval systems without relevance judgments, which is found to be correlated with actual evaluations using relevance judgments in the TREC competition.", "Key Idea": "The authors propose an explanation for this phenomenon by devising a simple measure for quantifying the similarity of retrieval systems and demonstrating that evaluating retrieval systems according to average similarity yields results quite similar to the methodology proposed by Soboroff et al., and these two techniques are in fact highly correlated.", "Method": "The authors use a simple measure for quantifying the similarity of retrieval systems by assessing the similarity of their retrieved results. Then, given a collection of retrieval systems and their retrieved results, they use this measure to assess the average similarity of a system to the other systems in the collection.", "Outcome": "The authors find that evaluating retrieval systems according to average similarity yields results quite similar to the methodology proposed by Soboroff et al., and these two techniques are in fact highly correlated.", "Future Impact": "This study highlights the importance of understanding the underlying mechanism of evaluation methods and potential future work could be to propose new evaluation methods that are not affected by popularity bias."} +{"id": "ded9a095-d94d-4ccd-8825-283ec4bb7093", "Context": "There is increasing interest in the adoption of UX within corporate environments, and there is a need to understand what competencies translate into effective UX design.", "Key Idea": "This paper proposes a co-construction of identity between the designer and their environment, where various factors such as tool and representational knowledge, complexity, and corporate culture influence perceptions of competence in UX over time.", "Method": "A 12-week longitudinal data collection, including surveys and interviews, documents the shift in students' and early professionals' perception of competence as they begin internships and full-time positions in UX.", "Outcome": "The study shows that students' perception of competence in UX shifts over time and is influenced by various factors such as tool and representational knowledge, complexity, and corporate culture.", "Future Impact": "The study suggests opportunities for future research in building a deeper understanding of competency in UX based on this preliminary framing of early UX practice."} +{"id": "dfba01c5-a632-4394-8607-9a32f20e526c", "Context": "Current systems do not automatically build a scene model that contains both 3D geometric information and photometric information under various illumination conditions.", "Key Idea": "The authors propose a system that automatically builds a scene model from real images, which contains both 3D geometric information of the scene structure and its photometric information under various illumination conditions.", "Method": "The authors use structure-from-motion and correlation-based stereo techniques to match pixels between images of different viewpoints and to reconstruct the scene in 3D space. The photometric property is extracted from images taken under different illumination conditions by computing a low-dimensional linear space of the spatio-illumination volume, and is represented by a set of basis images.", "Outcome": "The proposed system can be used to create realistic renderings from different viewpoints and illumination conditions. Applications include object recognition, virtual reality and product advertisement.", "Future Impact": "N/A"} +{"id": "e10b883a-ded6-4b4e-9934-6daba2d2f2b2", "Context": "The problem of designing efficient and effective solutions for large scale similarity search is important. Existing semantic hashing methods represent data examples as compact binary codes through semantic hashing, which has produced promising results with fast search speed and low storage cost. However, these methods model document relationships based on similarity in a keyword feature space, which does not fully reflect semantic relationships that go beyond keyword matching and does not exploit tag information.", "Key Idea": "The authors propose a novel hashing approach, Semantic Hashing using Tags and Topic Modeling (SHTTM), that incorporates both tag information and similarity information from probabilistic topic modeling. A unified framework is designed for ensuring hashing codes to be consistent with tag information by a formal latent factor model and preserving the document topic/semantic similarity.", "Method": "An iterative coordinate descent procedure is proposed for learning the optimal hashing codes. An extensive set of empirical studies on four different datasets has been conducted to demonstrate the advantages of the proposed SHTTM approach against several other state-of-the-art semantic hashing techniques.", "Outcome": "Experimental results indicate that the modeling of tag information and utilizing topic modeling are beneficial for improving the effectiveness of hashing separately, while the combination of these two techniques in the unified framework obtains even better results.", "Future Impact": "N/A"} +{"id": "e1a191db-6463-4be9-883d-dfc7fb05c5f4", "Context": "Dependency parsing is the task of analyzing the grammatical structure of a sentence and representing it as a tree, where nodes represent words and edges represent grammatical relationships between words. Previous methods for dependency parsing do not enforce certain structural properties that can be useful for better representing the set of admissible dependency structures in treebanks and connecting dependency parsing to context-sensitive grammatical formalisms.", "Key Idea": "The authors propose a novel dependency parsing method that enforces two structural properties on dependency trees: bounded block degree and well-nestedness.", "Method": "The authors cast the problem of enforcing these structural properties on dependency trees as an Integer Linear Program and solve it using Lagrangian Relaxation. They then derive a heuristic and an exact method based on a Branch-and-Bound search.", "Outcome": "The proposed methods are efficient and competitive compared to a baseline unconstrained parser, while enforcing structural properties in all cases.", "Future Impact": "The proposed method can lead to better representation of admissible dependency structures in treebanks and provide a way to connect dependency parsing to context-sensitive grammatical formalisms."} +{"id": "e2466d98-98f9-41a8-8a38-e862adc2ed47", "Context": "The study is motivated by the importance of Information Technology (IT) Governance models for public organizations, and a gap in the literature for an IT Governance model that can be adopted by both practitioners and researchers.", "Key Idea": "The authors propose a new IT Governance model that can be adopted by public sector organizations in a simple and dynamic manner.", "Method": "The authors conducted a systematic literature review and an empirical survey using a questionnaire based on the COBIT 4.1 maturity model to investigate IT Governance practice in multiple case studies from the Kingdom of Bahrain.", "Outcome": "The proposed model provides a basic structure of a concept, allowing organizations to gain a better perspective on IT Governance processes and providing a clear focus for decision-making attention.", "Future Impact": "The proposed model forms a basis for further research in IT Governance adoption models and bridges the gap between conceptual frameworks, real life and functioning governance."} +{"id": "e58b9947-7a3a-414d-a0e8-d6cf02ed7127", "Context": "With recent advances in RFID, wireless sensor networks, and Web-based services, physical things are becoming an integral part of the emerging ubiquitous Web. The paper focuses on the things recommendation problem in Internet of Things (IoT).", "Key Idea": "The authors propose a unified probabilistic-based framework that fuses information across relationships between users (i.e., users' social network) and things (i.e., things correlations) to make more accurate recommendations.", "Method": "The authors validate the proposed approach based on an Internet of Things platform and perform experiments to demonstrate its feasibility and effectiveness.", "Outcome": "The proposed approach not only inherits the advantages of matrix factorization but also exploits the merits of social relationships and thing-thing correlations.", "Future Impact": "N/A"} +{"id": "e827ee51-aafd-4f3b-99ca-60a5e289a555", "Context": "The PASCAL 2005 Recognizing Textual Entailment challenge aims to recognize textual entailment, which is the ability to determine if a piece of text implies a given hypothesis.", "Key Idea": "The authors propose a logical approach to recognize textual entailment by extracting atomic propositions from both the text and the entailment hypothesis and expressing them in a custom logical notation.", "Method": "The authors use the output of Link Parser to extract propositions and encode them into a logical notation. To detect independent entailment relations, the system relies on the use of Otter and WordNet.", "Outcome": "The proposed method is used to test the practicability of a purely logical approach for recognizing textual entailment.", "Future Impact": "N/A"} +{"id": "ea414544-c89c-4039-8227-23b11e9a1239", "Context": "Question answering systems that supplement Web search engines struggle to produce single-snippet answers to definition questions.", "Key Idea": "The authors propose a practically unsupervised learning method that uses on-line encyclopedias and dictionaries to generate positive and negative definition examples to train an SVM.", "Method": "The authors perform experiments to compare the proposed method to the alternative of training the system on questions and news articles from TREC.", "Outcome": "The proposed method outperforms the alternative and helps the search engine handle definition questions significantly better.", "Future Impact": "N/A"} +{"id": "eb15ebe7-aa58-4a98-8f9e-939967c6359f", "Context": "Synchronization refers to the problem of inferring the unknown values attached to vertices of a graph where edges are labelled with the ratio of the incident vertices, and labels belong to a group. The synchronization problem on multi-graphs arises when multiple measures are available to model the relationship between two vertices, such as when different sensors measure the same quantity or when the original graph is partitioned into sub-graphs that are solved independently.", "Key Idea": "The authors present MultiSynch, a synchronization algorithm for multi-graphs that is based on a principled constrained eigenvalue optimization, which is a general solution that can cope with any linear group.", "Method": "The authors prove empirically that the baseline solution of reducing multi-graphs to simple ones by averaging their multi-edges falls short and present MultiSynch algorithm for multi-graphs synchronization.", "Outcome": "MultiSynch algorithm is shown to be profitably usable both on synthetic and real problems.", "Future Impact": "N/A"} +{"id": "f23bdd28-b2d6-4a42-a56c-c9774f6451b5", "Context": "With the development of hypersonic vehicles in near space, tracking for them is becoming a new task and hotspot. Current tracking algorithms may not be effective for hypersonic targets, especially for the sliding jump maneuver.", "Key Idea": "The authors propose a learning tracking algorithm that uses a Sine model and the Interacting Multiple Model (IMM) algorithm. The algorithm also learns the target tracking error characteristics to adjust the sampling rate adaptively.", "Method": "The authors compare the proposed algorithm with the single accurate model algorithm and general IMM algorithms with fixed sampling rate. The authors conduct simulation experiments to prove the effectiveness of the proposed algorithm.", "Outcome": "The proposed algorithm improves the tracking accuracy effectively compared to the single accurate model algorithm and general IMM algorithms with fixed sampling rate.", "Future Impact": "N/A"} +{"id": "f2c9f8c3-f9d1-4cae-b7ae-a919ada1daaf", "Context": "Access to online visual search engines implies sharing of private user content - the query images.", "Key Idea": "The authors introduce the concept of targeted mismatch attack for deep learning based retrieval systems, which generates an adversarial image that leads to identical or very similar retrieval results as the user intended query but looks nothing like it.", "Method": "The authors transfer attacks to partially unknown systems by designing various loss functions for the adversarial image construction, including loss functions for unknown global pooling operation or unknown input resolution by the retrieval system.", "Outcome": "The authors evaluate the attacks on standard retrieval benchmarks and compare the results retrieved with the original and adversarial image.", "Future Impact": "The study highlights the potential risks of private user content sharing in online visual search engines and the need for more robust retrieval systems."} +{"id": "f32d6bc3-d75e-4e84-8bfa-3c83578281dc", "Context": "Transaction log analysis is a methodology used to examine user commands and system responses in online information searches. It has been used to aid in the evaluation of information systems and improve existing and future systems.", "Key Idea": "This paper proposes a methodology for monitoring and evaluating information systems using transaction log analysis. It uses stochastic pattern developments within parsed user sessions, mathematical models utilizing Markov chain analysis and state transition probability matrices.", "Method": "The authors obtained machine-readable transaction log tapes from online catalogs and analyzed them using the proposed methodology.", "Outcome": "The results obtained from the analysis of patron use and system response patterns from several online public catalogs have been presented in the paper.", "Future Impact": "The predictive power of the methodology may allow real-time aids to be developed and assist in system design."} +{"id": "f32e53d5-c7f4-407e-a5cb-26fef230b5fd", "Context": "Matrix factorization has been widely adopted for recommendation by learning latent embeddings of users and items from observed user-item interaction data. However, previous methods usually assume the learned embeddings are static or homogeneously evolving with the same diffusion rate.", "Key Idea": "The authors propose a novel dynamic matrix factorization model, named Dynamic Bayesian Logistic Matrix Factorization (DBLMF), which aims to learn heterogeneous user and item embeddings that are drifting with inconsistent diffusion rates.", "Method": "The authors extend logistic matrix factorization to model the probability a user would like to interact with an item at a given timestamp, and a diffusion process to connect latent embeddings over time. An efficient Bayesian inference algorithm has also been proposed to make DBLMF scalable on large datasets.", "Outcome": "The effectiveness of the proposed method has been demonstrated by extensive experiments on real datasets, compared with the state-of-the-art methods.", "Future Impact": "N/A"} +{"id": "f3cef657-1887-42a7-9e11-8fdf54f8fa90", "Context": "Catastrophic forgetting is a major problem in continual learning, which is the ability of a model to learn new tasks without forgetting previously learned ones.", "Key Idea": "The authors propose a new continual learning approach called kernel continual learning, which leverages the non-parametric nature of kernel methods and an episodic memory unit to tackle catastrophic forgetting.", "Method": "The authors use an episodic memory unit that stores a subset of samples for each task to learn task-specific classifiers based on kernel ridge regression. They also introduce variational random features to learn a data-driven kernel for each task and formulate kernel continual learning as a variational inference problem.", "Outcome": "The proposed method is evaluated on four benchmarks and is shown to be effective in tackling catastrophic forgetting.", "Future Impact": "The proposed method can be used to improve continual learning in various applications such as robotics, natural language processing, and computer vision."} +{"id": "f41bfd75-a6b0-44d7-a842-b117797ebd7e", "Context": "In answer extraction, features are generated from surface texts, but the process of generating features from more structured data representations, such as parse trees, is not well-studied.", "Key Idea": "The authors propose and compare three methods for generating features from parse trees: feature vector, string kernel, and tree kernel, for use in Support Vector Machines.", "Method": "The authors perform experiments on the TREC question answering task using features generated from surface texts and parse trees using the proposed methods.", "Outcome": "The results show that features generated from more structured data representations, such as parse trees, significantly improve the performance based on features generated from surface texts.", "Future Impact": "N/A"} +{"id": "f4f1bbaf-c1a2-44d5-8305-27235fa69d62", "Context": "In the field of web mining and web science, as well as data science and data mining, there has been a lot of interest in the analysis of (social) networks. With the growing complexity of heterogeneous data, feature-rich networks have emerged as a powerful modeling approach.", "Key Idea": "The tutorial presents a unified perspective on feature-rich networks, focusing on different modeling approaches such as multiplex and attributed networks.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The tutorial outlines important principles, methods, tools and future research directions in this emerging field of feature-rich networks."} +{"id": "f64fdfde-7e93-411b-865a-1e29d71c95b2", "Context": "Large-scale topic models are used in many practical applications for feature extraction and dimensionality reduction. Hierarchical topic models (HTMs) are an extension of flat topic models that can learn topics of different levels of abstraction, but existing scalable systems for flat topic models cannot handle HTMs due to their complicated data structures and susceptibility to local optima.", "Key Idea": "The authors propose an efficient partially collapsed Gibbs sampling algorithm for hierarchical latent Dirichlet allocation (hLDA), an initialization strategy to deal with local optima, and efficient data layout and distributed data structures for building scalable systems for HTMs.", "Method": "The authors implement their proposed methods and evaluate their performance on a 131-million-document corpus with 28 billion tokens. They compare their implementation to the previous open-source implementation for hLDA.", "Outcome": "The authors' system is 87 times more efficient than the previous open-source implementation for hLDA, and can scale to thousands of CPU cores. They are able to extract 1,722 topics from the corpus with 50 machines in just 7 hours.", "Future Impact": "N/A"} +{"id": "f82f2e26-5437-4397-b781-50668ce5851b", "Context": "In present Statistical Machine Translation (SMT) systems, alignment is trained in a previous stage as the translation model, and alignment model parameters are not tuned in function of the translation task.", "Key Idea": "The authors propose a novel framework for discriminative training of alignment models with automated translation metrics as maximization criterion, where alignments are optimized for the translation task without the need for link labels at the word level.", "Method": "The authors evaluate the framework in terms of automatic translation evaluation metrics.", "Outcome": "The authors observe an improvement in translation quality.", "Future Impact": "N/A"} +{"id": "fb538ce2-abf4-4bd8-b35c-1bfe3ab9f48e", "Context": "The complex word identification task is important for lexical simplification, which helps improve the readability of texts with challenging words.", "Key Idea": "The authors developed two systems using Naive Bayes and Random Forest classifiers to identify complex words based on various lexical and semantic features.", "Method": "The authors used Naive Bayes and Random Forest classifiers to identify complex words in a sentence. They incorporated rule-based post-processing techniques to improve the performance of the Naive Bayes classifier.", "Outcome": "The Naive Bayes classifier based system achieves the maximum G-score of 76.7%", "Future Impact": "N/A"} +{"id": "feb75e1f-7838-48ca-9a78-cc31b717e5bf", "Context": "Previous work on finding fraudulent reviews on crowd-sourced review platforms mainly focuses on supervised machine learning and textual and stylometry features. These works have limited ground truth data and assume a limited threat model.", "Key Idea": "The authors propose OneReview, a system that finds fraudulent content on a crowd-sourced review site by leveraging correlations with other independent review sites and the use of textual and contextual features. It utilizes Change Point Analysis method and supervised machine learning to detect anomalous changes in a business’ reputation across multiple review sites and identify fraudulent reviews.", "Method": "The authors evaluated their approach using data from two reviewing websites, Yelp and TripAdvisor, and obtained Yelp reviews through the Yelp Data Challenge. They used Change Point Analysis method on the reviews of every business independently on every website and then used their proposed Change Point Analyzer to evaluate change-points, detect those that do not match across the websites, and identify them as suspicious. They used supervised machine learning, utilizing a combination of textual and metadata features to locate fraudulent reviews among the suspicious reviews.", "Outcome": "The authors obtained 97% (+/- 0.01) accuracy, 91% (+/- 0.03) precision and 90% (+/- 0.06) recall in detecting fraudulent reviews. They classified 8% of all reviews as fraudulent. They also detected fraudulent campaigns and identified 3,980 businesses with fraudulent reviews and 14,910 suspected spam.", "Future Impact": "N/A"} +{"id": "fed7302a-43a7-412e-8ace-d07905e38c3c", "Context": "In one-class collaborative filtering, interpreting and modeling the latent signal from the missing class is a challenge.", "Key Idea": "The authors propose a novel Bayesian generative model for implicit collaborative filtering that delineates the odds of a user disliking an item from simply being unaware of it. The latent signal is treated as an unobserved random graph connecting users with items they might have encountered.", "Method": "The authors demonstrate large-scale distributed learning through a combination of stochastic gradient descent and mean field variational inference over random graph samples.", "Outcome": "The proposed model is fine-grained compared to a state-of-the-art baseline on real-world data.", "Future Impact": "N/A"} +{"id": "ffd14676-a525-479f-a74e-2c5d3a85c510", "Context": "There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units. Representing shapes in parallel systems is a complex issue.", "Key Idea": "This paper considers the difficulties involved in representing shapes in parallel systems and suggests ways of overcoming them. It provides a mechanism for shape perception and visual attention.", "Method": "N/A", "Outcome": "N/A", "Future Impact": "The paper's proposed mechanism for shape perception and visual attention can allow a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts."} +{"id": "adfd8058-64b3-4062-953c-034b732e2fa0", "Context": "Current video compression schemes are based on complex algorithms such as H.264, which may not be efficient in coding certain types of video sequences.", "Key Idea": "The authors propose a video compression scheme based on texture synthesis through Directional Empirical Mode Decomposition (DEMD) algorithm. The proposed scheme decompose P and B-frames into Intrinsic Mode Function (IMF) image and its residue, and only the first level IMF image for P and B frames are coded.", "Method": "The authors perform wavelet decomposition over residual image and use energy level at the HH band as a decision criterion for number of decomposition to be performed for optimum synthesis. The authors also demonstrate the effectiveness of the algorithm in multi-resolution parametric modeling of image data and scalable coding of IMF parameters.", "Outcome": "The proposed scheme demonstrates significant compression with acceptable quality.", "Future Impact": "N/A"} \ No newline at end of file diff --git a/massw/__init__.py b/massw/__init__.py new file mode 100644 index 0000000..366295c --- /dev/null +++ b/massw/__init__.py @@ -0,0 +1 @@ +"""Init file for massw package.""" diff --git a/massw/api/__init__.py b/massw/api/__init__.py new file mode 100644 index 0000000..14e8999 --- /dev/null +++ b/massw/api/__init__.py @@ -0,0 +1 @@ +"""Init.""" diff --git a/massw/api/api_gpt.py b/massw/api/api_gpt.py new file mode 100644 index 0000000..1dd6a20 --- /dev/null +++ b/massw/api/api_gpt.py @@ -0,0 +1,633 @@ +"""Async API calls to OpenAI GPTs. + +Reference: https://github.com/SpellcraftAI/oaib + +```python +# Pass API key, or set env var `OPENAI_API_KEY` +batch = Batch(tpm=10000, azure=AzureConfig()) + +# Add chat completion requests +for i in range(5): + await batch.add( + "chat.completions.create", + model="gpt-3.5-turbo", + messages=[{"role": "user", "content": "say hello"}] + ) + +# Run the batch processing for chat completions +chat = await batch.run() + +# Add embedding requests +for i in range(5): + await batch.add( + "embeddings.create", + model="text-embedding-3-large", + input="hello world" + ) + +# Run the batch processing for embeddings +embeddings = await batch.run() +``` +""" +import os +import signal +import asyncio +import ast +from asyncio import (FIRST_COMPLETED, CancelledError, Event, Lock, Queue, + QueueEmpty, create_task, gather, sleep, + wait, wait_for) +from asyncio import TimeoutError as AsyncTimeoutError +from datetime import datetime +from os import environ +from time import time +from types import SimpleNamespace +from typing import Coroutine, List, Set, Dict + +import openai +import pandas as pd +from tqdm.auto import tqdm + + +async def add_messages_and_run(batch, messages, model): + """Help async function to add messages to the batch and run it.""" + for pid, msg in messages: + await batch.add("chat.completions.create", + model=model, + messages=msg, + metadata={"pid": pid}) + return await batch.run() + + +def prompts_to_raw_output_gpt(messages, model, tpm): + """ + Simulate API call to a GPT-based model. + + Return chat results using an asynchronous batch process. + """ + batch = Batch(tpm=tpm, azure=AzureConfig()) + loop = asyncio.get_event_loop() + results = loop.run_until_complete(add_messages_and_run(batch, + messages, + model)) + return results + + +def raw_output_to_dict_gpt(output_path: str) -> Dict[str, str]: + """ + Load and convert raw output from the GPT model into a dictionary. + + Args: + output_path (str): Path to the model output CSV file. + + Returns: + Dict[str, str]: A dictionary mapping pid to processed output. + """ + output_dict = {} + task_output = pd.read_csv(output_path, sep="\t") + for _, row in task_output.iterrows(): + if 'result' in row and row['result']: + output = ast.literal_eval( + row['result'])["choices"][0]["message"]["content"] + else: + output = "" + output_dict[row['pid']] = output + + return output_dict + + +class AzureConfig(SimpleNamespace): + """Azure OpenAI API configuration.""" + + def __init__( + self, + azure_endpoint: str = environ.get("AZURE_OPENAI_ENDPOINT"), + api_version: str = environ.get("AZURE_OPENAI_VERSION"), + api_key: str = environ.get("AZURE_OPENAI_KEY"), + ): + """Initialize AzureConfig with the Azure OpenAI API credentials.""" + super().__init__(azure_endpoint=azure_endpoint, + api_version=api_version, + api_key=api_key) + + def display_endpoint(self): + """Display the Azure endpoint.""" + print(f"Endpoint: {self.azure_endpoint}") + + def is_configured(self): + """Check if all necessary API configuration details are provided.""" + return all([self.azure_endpoint, self.api_version, self.api_key]) + + +def getattr_dot(obj: any, index: str): + """Get an attribute of an object using dot notation.""" + parts = index.split('.') + for part in parts: + obj = getattr(obj, part, None) + if obj is None: + return None + return obj + + +async def cancel_all(tasks: set): + """Cancel all tasks in a set.""" + tasks.discard(None) + for task in tasks: + task.cancel() + + try: + await gather(*tasks) + except CancelledError: + pass + + +async def race(tasks: Set[Coroutine]): + """Run tasks concurrently and return when the first one completes.""" + tasks = map(create_task, tasks) + done, pending = await wait(tasks, return_when=FIRST_COMPLETED) + + await cancel_all(pending) + results = await gather(*done) + + return results[0] + + +async def close_queue(queue: Queue): + """Close the queue and wait for all tasks to complete.""" + while not queue.empty(): + queue.get_nowait() + queue.task_done() + + return await queue.join() + + +def get_limits(headers): + """Get the rate limits from the response headers.""" + rpm = headers.get("x-ratelimit-limit-requests") + tpm = headers.get("x-ratelimit-limit-tokens") + + return int(rpm), int(tpm) + + +class Batch: + """ + An instance for managing batch requests to the OpenAI API. + + Parameters + ---------- + rpm : int, default: `500` + The maximum number of requests allowed per minute. Defaults to lowest + tier. + tpm : int, default: `10_000` + The maximum number of tokens allowed per minute. Defaults to lowest + tier. + workers : int, default: `8` + The number of concurrent workers to process the requests. + safety : float, default: `0.1` + The safety factor to apply to the token per minute calculation. + Defaults to `0.1`, which means the engine will wait + until the current TPM drops below 90% of the limit, + to prevent going over. This is necessary because + we don't know how many tokens a response will contain before we get it. + loglevel : int, default: `1` + If set to 0, suppresses the progress bar and logging output. + If set to 1, logs include metadata only. + If set to 2, logs include both data and + metadata for each request. + timeout : int, default: `60` + The maximum time to wait for a single request to complete, in seconds. + api_key : str, default: `os.environ.get("OPENAI_API_KEY")` + The API key used for authentication with the OpenAI API. If not + provided, the class attempts to use an API_KEY constant defined + elsewhere. + log_path : str, default: `"oaib.txt"` + The file path for logging the progress and errors of batch processing. + Defaults to "oaib.txt". + **client_args + Additional keyword arguments to pass to the OpenAI client. + """ + + def __init__(self, + rpm: int = -1, + tpm: int = 10_000, + workers: int = 8, + safety: float = 0.1, + loglevel: int = 1, + timeout: int = 60, + azure=None, + api_key: str = None, + logdir: str = "oaib.txt", + index: List[str] = None, + **client_kwargs): + """Init method for the Batch class.""" + api_key = api_key or (os.environ.get("AZURE_OPENAI_KEY") + if azure else os.environ.get("OPENAI_API_KEY")) + + if not api_key: + raise ValueError( + "No OpenAI API key found.\ + Please provide an `api_key` parameter or\ + set the `OPENAI_API_KEY` environment variable." + ) + + if loglevel > 2: + raise ValueError( + f"Allowable `loglevel` values are 0, 1, or 2; found {loglevel}" + ) + + self.rpm = rpm + self.tpm = tpm + self.safety = safety + self.loglevel = loglevel + self.timeout = timeout + self.logdir = logdir + self.index = index + self.azure = None + + if rpm == -1: + self.rpm = tpm // 1000 * 6 + + if azure: + azure = vars(azure) + self.azure = azure + + self.client = openai.AsyncAzureOpenAI( + **{ + **azure, "api_key": api_key + }, **client_kwargs) + else: + self.client = openai.AsyncOpenAI(api_key=api_key, **client_kwargs) + + self.__num_workers = workers + + self.__lock = Lock() + self.__queue = Queue() + self.__stopped = Event() + + self.__workers = set() + self.__processing = set() + self.__callbacks = set() + + self.__current = SimpleNamespace(rpm=0, tpm=0) + self.__totals = SimpleNamespace(requests=0, tokens=0, queued=0) + self.__progress = SimpleNamespace(main=None, rpm=None, tpm=None) + + # Send stop event on SIGINT. + signal.signal(signal.SIGINT, + lambda code, stack: create_task(self.stop(code, stack))) + + def __clear_log(self): + """Clear the log file.""" + with open(self.logdir, "w", encoding="utf-8") as file: + file.write("") + + def log(self, *messages, worker: int = None, loglevel: int = None): + """Log messages to the log file.""" + if (loglevel or self.loglevel) > 0: + now = datetime.now() + timestamp = now.strftime("%Y-%m-%d %H:%M:%S") + + for message in messages: + prefix = f"WORKER {worker}" if worker else "MAIN" + message = " | ".join([prefix.rjust(8), timestamp, message]) + + with open(self.logdir, "a", encoding="utf-8") as file: + file.write(message + "\n") + + async def _cleanup(self): + """ + Ensure the stop event is set. + + All workers and processing tasks are + cancelled. Also closes the progress bar and queue. + """ + # Stop was triggered, tasks done: final tick. + self._tick() + + if self.succeeded: + self.log("WAITING FOR CLOCK") + await wait([self.__clock]) + else: + self.log("CANCELLING ALL TASKS") + await cancel_all({ + self.__clock, + *self.__processing, + *self.__callbacks, + *self.__workers, + }) + + await close_queue(self.__queue) + + self.__current = SimpleNamespace(rpm=0, tpm=0) + self.__totals = SimpleNamespace(requests=0, tokens=0, queued=0) + + for bar_progress in vars(self.__progress).values(): + if isinstance(bar_progress, tqdm): + bar_progress.close() + + self.__workers.clear() + + def _tick(self): + now = time() + seconds = now - self._start + minutes = seconds / 60 + + if self._last_tick and now - self._last_tick < pd.Timedelta("1s"): + return + + self.__current.rpm = self.__totals.requests // minutes + self.__current.tpm = self.__totals.tokens // minutes + + self.__progress.main.n = self.__totals.requests + self.__progress.main.total = self.__totals.queued + + self.__progress.rpm.n = self.__current.rpm + self.__progress.rpm.total = self.rpm + + self.__progress.tpm.n = self.__current.tpm + self.__progress.tpm.total = self.tpm + + self.__progress.main.refresh() + self.__progress.tpm.refresh() + self.__progress.rpm.refresh() + + if self.__stopped.is_set(): + self.__progress.main.set_description( + "✅ DONE" if self.succeeded else "🛑 STOPPED", refresh=True) + + async def _watch(self): + while True: + await sleep(0.1) + self._tick() + + if self.__stopped.is_set(): + break + + async def _process(self, request, i=None): + endpoint, func, kwargs, metadata = request + + if self.loglevel == 1: + log_content = f"{metadata}" + else: + log_content = f"{metadata} | {kwargs}" + + self.log(f"PROCESSING | {log_content}", worker=i) + + try: + [response] = await wait_for(gather(func(**kwargs)), + timeout=self.timeout) + + except AsyncTimeoutError: + self.log(f"TIMEOUT | {self.timeout}s | {kwargs}", worker=i) + return + + except Exception as e: + self.log(f"PROCESSING ERROR | {e}", worker=i) + return + + headers = response.headers + response = response.parse() + tokens = response.usage.total_tokens + + row = pd.DataFrame([{ + **metadata, "endpoint": endpoint, + **kwargs, "result": response.model_dump() + }]) + + # Store one copy of response headers - for use by Auto subclass. + if self._headers is None: + self.log(f"HEADERS | {dict(headers)}") + self._headers = headers + + self.__totals.requests += 1 + self.__totals.tokens += tokens + + self.output = pd.concat([self.output, row], ignore_index=True) + self.log(f"PROCESSED | {kwargs}", worker=i) + + if self._callback: + callback = create_task(self._callback(row)) + self.__callbacks.add(callback) + callback.add_done_callback( + lambda _: self.__callbacks.remove(callback)) + + def _next(self, i): + try: + self.log(f"REQUESTS: {self.__queue.qsize()}", worker=i) + request = self.__queue.get_nowait() + self.__queue.task_done() + + except QueueEmpty: + self.log("EMPTY QUEUE", worker=i) + + if self._listening: + self.log("LISTENING", worker=i) + return True + + return False + + processing = create_task(self._process(request, i)) + self.__processing.add(processing) + processing.add_done_callback( + lambda _: self.__processing.remove(processing)) + + return True + + async def __worker(self, i): + while True: + async with self.__lock: + if self.__stopped.is_set(): + break + + proceed = self._next(i) + if not proceed: + break + + self.__progress.main.set_description("🟢 RUNNING", + refresh=True) + + now = time() + avg_tpr = (now - self._start) / (self.__totals.requests or 1) + + # The RPM does not need a safety threshold because it is known + # in advance, but we still apply a 1% reduction to minimize + # going over on small timescales. + # effective_rpm = 0.99 * self.rpm + effective_tpm = (1 - self.safety) * self.tpm + rpm_delay = 60 / self.rpm + + start = now + while self.__current.tpm + avg_tpr >= effective_tpm \ + and not self.__stopped.is_set(): + self.__progress.main.set_description("🟡 WAITING", + refresh=True) + + await sleep(0.1) + end = time() + + remaining = rpm_delay - (end - start) + if remaining > 0: + await sleep(remaining) + + def __setup(self, callback=None, listening=False): + self.output = pd.DataFrame() + self.succeeded = False + + # For use by Auto(Batch) subclass. + self._headers = None + # If true, clearing the queue will not stop the engine. + self._listening = listening + # Callback to run on each processed request. + self._callback = callback + + self._start = time() + self._last_tick = None + + self.__clear_log() + self.__stopped.clear() + self.__clock = create_task(self._watch()) + + self.__workers = { + create_task(self.__worker(i)) + for i in range(self.__num_workers) + } + + silence = self.loglevel == 0 + + self.__progress.main = tqdm(total=self.__queue.qsize(), + unit='req', + dynamic_ncols=True, + disable=silence) + + self.__progress.rpm = tqdm( + desc="RPM", + total=self.rpm, + unit='rpm', + dynamic_ncols=True, + disable=silence, + bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt}") + + self.__progress.tpm = tqdm( + desc="TPM", + total=self.tpm, + unit='tpm', + dynamic_ncols=True, + disable=silence, + bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt}") + + if self.azure: + self.log(f"USING AZURE | {self.azure}") + + async def listen(self, callback=None): + """ + Listen for incoming requests and processes them as they arrive. + + This method is non-blocking and + can be used to process requests in real-time. + """ + self.__setup(callback, True) + await self.__stopped.wait() + await self.stop() + + async def run(self, callback=None): + """ + Initiate the processing of all queued requests. + + Manage requests according to the specified rate limits. + This method waits for the processing to complete or stop conditionally. + + Returns + ------- + output : pandas.DataFrame + A DataFrame containing the results of the processed requests. + """ + if not self.__queue.qsize() >= 1: + raise ValueError( + "Engine cannot run without at least one job scheduled") + + # start = time() + # Run setup and wait for a stop condition. + self.__setup(callback) + await race({self.__queue.join(), self.__stopped.wait()}) + + # If the run was successful, it needs to be stopped. Finish processing + # existing requests first. + if not self.__stopped.is_set(): + self.log("FINISHING PROCESSING | 5 second timeout") + await gather(*self.__processing) + await gather(*self.__callbacks) + await gather(*self.__workers) + await self.stop() + + if self.index: + self.log("INDEX | Setting index") + self.output.set_index(self.index, inplace=True) + self.output.sort_index(inplace=True) + + self.log("RETURNING OUTPUT") + # print(f"\nRun took {time() - start:.2f}s.\n") + return self.output + + async def stop(self, code=0, stack=None): + """ + Stop the processing of requests. + + Parameters + ---------- + code : int, default: `None` + Set an exit code. By default, + it is set to 1 if the run was interrupted before completing, + 0 otherwise. + stack : never + Reserved for future use, + e.g., for passing exception stack information. Currently not used. + + Returns + ------- + success : bool + False if the run was cancelled or interrupted, + True if it completed successfully. + """ + self.succeeded = code == 0 + + self.log(f"STOP EVENT | Exit code {code}") + if stack: + self.log(f"STACK INFO\n\n{stack}\n") + + self.__stopped.set() + await self._cleanup() + + async def add(self, + endpoint="chat.completions.create", + metadata: dict = {}, + **kwargs): + """ + Schedules an API request to be added to the batch processing queue. + + Parameters + ---------- + endpoint : str, default: `"chat.completions.create"` + The OpenAI API endpoint to call, + e.g., 'chat.completions.create' or 'embeddings.create'. + metadata : dict, default: `None` + A dictionary containing additional data to be added + to this observation row in the DataFrame. + **kwargs + Keyword arguments to pass to the OpenAI API endpoint function. + Common kwargs include 'model' and input parameters, + like 'messages' for 'chat.completions.create' + or 'input' for 'embeddings.create'. + + Returns + ------- + None + """ + # Read the client method. + func = getattr_dot(self.client.with_raw_response, endpoint) + + # Add the request to the queue. + request = (endpoint, func, kwargs, metadata) + model = kwargs.get("model") + await self.__queue.put(request) + + self.__totals.queued += 1 + self.log(f"QUEUED | {model}") diff --git a/massw/api/api_mistral.py b/massw/api/api_mistral.py new file mode 100644 index 0000000..69ae704 --- /dev/null +++ b/massw/api/api_mistral.py @@ -0,0 +1,91 @@ +"""Api module for interacting with the Mistral model.""" +import urllib.request +import json +import pandas as pd +import time +import os +from typing import Dict + + +def prompts_to_raw_output_mistral(messages): + """Process prompts using the specified Mistral model endpoint.""" + final_results = pd.DataFrame(columns=['pid', 'output']) + + url = os.environ.get("MISTRAL_API_URL") + api_key = os.environ.get("MISTRAL_API_KEY") + print(f"Using Mistral API at {url}") + print(f"Using Mistral API key: {api_key}") + headers = { + 'Content-Type': 'application/json', + 'Authorization': f'Bearer {api_key}', + 'azureml-model-deployment': 'mistralai-mixtral-8x7b-instru-7' + } + + for pid, msg in messages: + response_df = handle_mistral_model(url, headers, msg, {"pid": pid}) + final_results = pd.concat([final_results, + response_df], ignore_index=True) + + return final_results + + +def raw_output_to_dict_mistral(output_path: str) -> Dict[str, str]: + """ + Load and convert raw output from the Mistral model into a dictionary. + + Args: + output_path (str): Path to the model output CSV file. + + Returns: + Dict[str, str]: A dictionary mapping pid to processed output. + """ + output_dict = {} + task_output = pd.read_csv(output_path, sep="\t", + converters={'result': lambda x: json.loads(x) + if x else None}) + for _, row in task_output.iterrows(): + output = row["output"] + output_dict[row['pid']] = output + return output_dict + + +def handle_mistral_model(url, headers, messages, entry): + """Handle the Mistral model API request.""" + output_df = pd.DataFrame(columns=['pid', 'output']) + max_retries = 5 + retries = 0 + while retries < max_retries: + try: + data = { + "input_data": { + "input_string": messages, + "parameters": { + "temperature": 0, + "top_p": 0.9, + "do_sample": True, + "max_new_tokens": 200, + "return_full_text": True + } + } + } + body = str.encode(json.dumps(data)) + req = urllib.request.Request(url, body, headers) + print(f"{req = }") + with urllib.request.urlopen(req) as response: + result_json = json.loads(response.read()) + output_df = output_df.append({"pid": entry["pid"], + "output": result_json}, + ignore_index=True) + break + except urllib.error.HTTPError as error: + print(f"The request failed with status code: {error.code}") + retries += 1 + time.sleep(2) + # mistral has a werid excetion, need to change below + # To avoid "Catching too general exception". + except Exception as e: + print(f"An unexpected error occurred: {str(e)}") + retries += 1 + time.sleep(2) + + return output_df diff --git a/massw/data.py b/massw/data.py new file mode 100644 index 0000000..eaeac7a --- /dev/null +++ b/massw/data.py @@ -0,0 +1,157 @@ +"""Data loading and processing functions.""" + +from dataclasses import dataclass +import numpy as np + +import os +import pandas as pd +import jsonlines as jl +from typing import List, Union + +from massw.download import download_dataset + +PROJECT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) +DATA_DIR = os.path.join(PROJECT_DIR, "data") + + +@dataclass +class Workflow: + """Workflow dataclass.""" + + id: str # Unique ID + title: str # Title of the publication + abstract: str # Abstract of the publication + year: int # Year of publication + venue: str # Publication venue + context: str # Context of the workflow + key_idea: str # Key idea of the workflow + method: str # Method used in the workflow + outcome: str # Outcome of the workflow + projected_impact: str # Projected impact of the workflow + data: dict = None # Additional data + + +class MASSWDataset: + """MASSW dataset class.""" + + def __init__(self, data: pd.DataFrame, metadata: pd.DataFrame): + """Initialize the dataset. + + Args: + data (pd.DataFrame): The data containing the workflows. + metadata (pd.DataFrame): The metadata containing the workflow + information. + """ + self.merged_data = data.join(metadata.set_index("id"), on="id") + self.merged_data = self.merged_data.set_index("id") + self.merged_data = self.merged_data.fillna(np.nan).replace([np.nan], + [None]) + + def __len__(self): + """Return the number of workflows in the dataset.""" + return len(self.merged_data) + + def _get_by_position(self, pos: int) -> Workflow: + """Return a workflow by its position. + + Args: + pos (int): The position of the workflow. + + Returns: + Workflow: The workflow object. + """ + row = self.merged_data.iloc[pos] + return Workflow(id=self.merged_data.index[pos], + title=row["title"], + abstract=row["abstract"], + year=row["year"], + venue=row["venue"], + context=row["context"], + key_idea=row["key_idea"], + method=row["method"], + outcome=row["outcome"], + projected_impact=row["projected_impact"], + data=row["data"]) + + def _get_by_unique_id(self, id_my: str) -> Workflow: + """Return a workflow by its unique ID. + + Args: + id_my (str): The unique ID of the workflow. + + Returns: + Workflow: The workflow object. + """ + row = self.merged_data.loc[id_my] + return Workflow(id=id_my, + title=row["title"], + abstract=row["abstract"], + year=row["year"], + venue=row["venue"], + context=row["context"], + key_idea=row["key_idea"], + method=row["method"], + outcome=row["outcome"], + projected_impact=row["projected_impact"], + data=row["data"]) + + def __getitem__(self, key): + """Return a workflow by its unique ID or position.""" + if isinstance(key, int): + return self._get_by_position(key) + if isinstance(key, str): + return self._get_by_unique_id(key) + raise TypeError("Invalid key type.") + + def __iter__(self): + """Return an iterator over the workflows.""" + for idx in range(len(self)): + yield self[idx] + + def __repr__(self): + """Return a string representation of the dataset.""" + return f"MASSWDataset({len(self)} workflows)" + + def search(self, + query: str, + return_ids=False) -> Union[List[Workflow], List[str]]: + """Search for workflows containing a query string by title. + + Args: + query (str): The query string to search for. + return_ids (bool): Whether to return the IDs of the workflows. + If true, returns a list of IDs. + Otherwise, returns a list of Workflows. + + Returns: + list: A list of workflows or IDs containing the query string. + """ + mask = self.merged_data["title"].str.contains(query, + case=False, + na=False) + ids = mask[mask].index + if return_ids: + return ids.tolist() + return [self._get_by_unique_id(id) for id in ids] + + +def load_massw(version: str = "v1") -> MASSWDataset: + """Load the massw dataset. + + Args: + version (str): The version of the dataset to load. + + Returns: + MASSWDataset: The MASSW dataset object. + """ + data_path = os.path.join(DATA_DIR, f"massw_{version}.tsv") + metadata_path = os.path.join(DATA_DIR, f"massw_metadata_{version}.jsonl") + if not os.path.exists(data_path) or not os.path.exists(metadata_path): + download_dataset(version) + data = pd.read_csv(data_path, sep="\t") + metadata = [] + with jl.open(metadata_path) as f: + for line in f: + metadata.append(line) + metadata = pd.DataFrame(metadata) + return MASSWDataset(data, metadata) diff --git a/massw/download.py b/massw/download.py new file mode 100644 index 0000000..fddf713 --- /dev/null +++ b/massw/download.py @@ -0,0 +1,50 @@ +"""Data loading and processing utilities.""" +import os +import sys + +import wget + +# Setting the project directory relative to this script's location +PROJECT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) + + +def download_dataset(version="v1"): + """Download the dataset from remote storage.""" + urls = { + "v1": { + "massw_metadata_v1.jsonl": + "https://www.dropbox.com/scl/fi/r2jlil9lj0ypo2fpl3fxa/\ + massw_metadata_v1.jsonl?rlkey=ohnriak63x4ekyli25naajp0q&dl=1", + "massw_v1.tsv": + "https://www.dropbox.com/scl/fi/ykkrpf269fikuchy429l7/\ + massw_v1.tsv?rlkey=mssrbgz3k8adij1moxqtj34ie&dl=1", + } + } + try: + files = urls[version] + except KeyError as e: + raise ValueError( + f"Invalid version: {version}.\ + Choose from {list(urls.keys())}") from e + for filename, url in files.items(): + print(f"Downloading {filename}...") + # Constructing the output path + out_path = os.path.join(PROJECT_DIR, "data", filename) + if os.path.exists(out_path): + print(f"{filename} already exists. Skipping download.") + continue + wget.download(url, out=out_path, bar=bar_progress) + + +def bar_progress(current, total, width=80): + """Display a progress bar for the download.""" + progress_message = f"Downloading: {current / total * 100:.0f}% \ + [{current} / {total}] bytes" + # Don't use print() as it will print in new line every time. + width = min(width, 100) + sys.stdout.write("\r" + progress_message) + sys.stdout.flush() + + +if __name__ == "__main__": + download_dataset() diff --git a/massw/llm_sim_eval.txt b/massw/llm_sim_eval.txt new file mode 100644 index 0000000..a1e3e14 --- /dev/null +++ b/massw/llm_sim_eval.txt @@ -0,0 +1,21 @@ +You are an expert in Computer Science with a specialization in text analysis, +you are tasked to evaluate the semantic similarity between pairs of abstractive +summarizations from scientific papers. Each summarization pertains to the same +aspect (Context, Key Idea, Method, Outcome, or Projected Impact) of the same +paper. + +For each pair of summarizations, classify the aspect, and assign a similarity score +on a scale from 1 to 10, where 1 indicates completely dissimilar and 10 +indicates identical content. Before scoring, include a brief justification for +your score. You should output your results in JSON format as shown in the example. + +Example Input: +Input 1: The experiments demonstrated a 20% increase in efficiency, confirming the proposed model's effectiveness. +Input 2: Results show that the new model outperforms existing ones by improving efficiency by approximately 20%. + +Example JSON Output: +{ + "aspect": "Outcome", + "score": 9, + "justification": "Both texts describe similar measurable improvements in efficiency, closely aligning in their depiction of the model's effectiveness." +} diff --git a/massw/metrics.py b/massw/metrics.py new file mode 100644 index 0000000..4c27674 --- /dev/null +++ b/massw/metrics.py @@ -0,0 +1,363 @@ +"""Compute cosine similarity, ROUGE, BLEU, METEOR, and BERTScore metrics. + +See example usage in the `__main__` block at the end of the file. +""" + +import json +import asyncio +from typing import List, Union + +import evaluate +import numpy as np +from sentence_transformers import SentenceTransformer +from massw.api.api_gpt import AzureConfig, Batch + +LLM_SIM_PROMPT = """ +You are an expert in Computer Science with a specialization in text analysis, +you are tasked to evaluate the semantic similarity between pairs of abstractive +summarizations from scientific papers. Each summarization pertains to the same +aspect (Context, Key Idea, Method, Outcome, or Projected Impact) of the same +paper. + +For each pair of summarizations, classify the aspect, +and assign a similarity score on a scale from 1 to 10, +where 1 indicates completely dissimilar and 10 +indicates identical content. Before scoring, include a brief justification for +your score. +You should output your results in JSON format as shown in the example. + +Example Input: +Input 1: The experiments demonstrated a 20% increase in efficiency, +confirming the proposed model's effectiveness. +Input 2: Results show that the new model outperforms existing ones +by improving efficiency by approximately 20%. + +Example JSON Output: +{ + "aspect": "Outcome", + "score": 9, + "justification": "Both texts describe similar measurable improvements + in efficiency, closely aligning in their depiction + of the model's effectiveness." +} +""" + + +class LLMSimilarity: + """Evaluate the similarity between two texts using a language model.""" + + def __init__(self, model_name: str = "gpt-4"): + """Initialize the language model similarity evaluator.""" + assert model_name in ["gpt-4", "gpt-35-turbo"] + self.model_name = model_name + self.tpm = {"gpt-4": 4000, "gpt-35-turbo": 40000}[model_name] + + def generate_prompt(self, text_1: str, text_2: str): + """Generate the prompt for the language model.""" + user_prompt = f"Input 1: {text_1}\nInput 2: {text_2}" + messages = [ + { + "role": "system", + "content": LLM_SIM_PROMPT + }, + { + "role": "user", + "content": user_prompt + }, + ] + return messages + + async def _compute(self, + predictions: List[str], + references: List[List[str]]): + """Compute the similarity between predictions and references.""" + batch = Batch(tpm=self.tpm, azure=AzureConfig()) + if isinstance(references[0], list): + new_predictions = [] + for pred, refs in zip(predictions, references): + new_predictions.extend([pred] * len(refs)) + new_references = [ref for refs in references for ref in refs] + predictions, references = new_predictions, new_references + # Compute similarity between each prediction and reference + for i, (pred, ref) in enumerate(zip(predictions, references)): + messages = self.generate_prompt(pred, ref) + await batch.add(endpoint="chat.completions.create", + response_format={"type": "json_object"}, + model=self.model_name, + messages=messages, + metadata={"id": i}) + results = await batch.run() + results = results.sort_values("id") + outputs = results["result"].apply( + lambda x: x["choices"][0]["message"]["content"]) + try: + scores = outputs.apply(lambda x: float(json.loads(x)["score"])) + except ValueError as e: + print(e) + scores = None + scores = scores.dropna() + scores = scores.to_numpy() + try: + scores = scores.reshape(len(predictions), -1) + return {"llm_sim": float(np.mean(np.max(scores, axis=1)))} + except ValueError as e: + print(e) + return {"llm_sim": float(np.mean(scores))} + + def compute(self, predictions: List[str], references: List[List[str]]): + """Compute the similarity between predictions and references.""" + loop = asyncio.get_event_loop() + results = loop.run_until_complete( + self._compute(predictions, references)) + return results + + +class CosineSimilarity: + """Compute cosine similarity between two ordered list of texts.""" + + def __init__(self): + """Initialize the SentenceTransformer model.""" + self.encoder = SentenceTransformer( + 'intfloat/multilingual-e5-large-instruct') + + def get_detailed_instruct(self, query: str) -> str: + """Generate a detailed instruct for the query.""" + return f"Instruct: Retrieve semantically similar text.\nQuery: {query}" + + def get_embeddings(self, texts: List[str], is_query: bool): + """Compute embeddings for the given texts.""" + if is_query: + texts = [self.get_detailed_instruct(query) for query in texts] + embeddings = self.encoder.encode( + texts, + convert_to_numpy=True, + normalize_embeddings=True, + ) + return embeddings + + def compute( + self, + predictions: List[str], + references: Union[List[str], List[List[str]]], + ): + """Compute the cosine similarity between predictions and references.""" + length = len(predictions) + if isinstance(references[0], list): + new_predictions = [] + for pred, refs in zip(predictions, references): + new_predictions.extend([pred] * len(refs)) + new_references = [ref for refs in references for ref in refs] + predictions, references = new_predictions, new_references + predictions_embeddings = self.get_embeddings(predictions, + is_query=True) + references_embeddings = self.get_embeddings(references, is_query=False) + # Compute pairwise cosine similarity + cosine_similarities = [] + for pred, ref in zip(predictions_embeddings, references_embeddings): + cosine_similarities.append(np.dot(pred, ref)) # Already normalized + cosine_similarities = np.array(cosine_similarities) + cosine_similarities = cosine_similarities.reshape(length, -1) + cosine_similarities = np.max(cosine_similarities, axis=1) + return {"cosine": float(np.mean(cosine_similarities))} + + +class NAHit: + """Compute the precision, recall, and F1 score for N/A hit metric.""" + + def is_na(self, s: str): + """Check if the string is N/A.""" + if s.lower() in ["n/a", "na", "not applicable"]: + return True + if len(s.split()) < 3: + return True + return False + + def compute( + self, + predictions: List[str], + references: Union[List[str], List[List[str]]], + ): + """Compute the precision, recall, and F1 score for N/A hit metric.""" + predictions_na = [self.is_na(pred) for pred in predictions] + if isinstance(references[0], list): + references_na = [] + references_na = [ + all(self.is_na(ref) for ref in refs) for refs in references + ] + else: + references_na = [self.is_na(ref) for ref in references] + precision = evaluate.load("precision") + recall = evaluate.load("recall") + f1 = evaluate.load("f1") + return { + "precision": + precision.compute( + predictions=predictions_na, + references=references_na, + )["precision"], + "recall": + recall.compute( + predictions=predictions_na, + references=references_na, + )["recall"], + "f1": + f1.compute( + predictions=predictions_na, + references=references_na, + )["f1"], + "pred_ratio": + sum(predictions_na) / len(predictions_na), + "ref_ratio": + sum(references_na) / len(references_na), + } + + +cs = CosineSimilarity() +llm_sim = LLMSimilarity() +bertscore = evaluate.load("bertscore") +bleurt = evaluate.load("bleurt", + module_type="metric", + checkpoint="BLEURT-20-D12", + config_name="BLEURT-20-D12") +rouge = evaluate.load("rouge") +bleu = evaluate.load("bleu") +meteor = evaluate.load("meteor") +nahit = NAHit() + + +def compute_metrics(predictions: List[str], + references: List[List[str]], + metric_names=None): + """Compute cosine similarity, ROUGE, BLEU, METEOR, and BERTScore.""" + if metric_names is None: + metric_names = [ + "cosine", + "rouge", + "bleu", + "meteor", + "bleurt", + "bertscore", + "nahit", + "llm_sim" + ] + metrics = {} + if "nahit" in metric_names: + metrics["nahit"] = nahit.compute( + predictions=predictions, + references=references, + ) + # Remove N/A predictions and references + if isinstance(references[0], list): + references_na = [ + all(nahit.is_na(ref) for ref in refs) for refs in references + ] + else: + references_na = [nahit.is_na(ref) for ref in references] + predictions_na = [nahit.is_na(pred) for pred in predictions] + both_not_na = [ + not pred_na and not ref_na + for pred_na, ref_na in zip(predictions_na, references_na) + ] + predictions = [ + pred for pred, not_na in zip(predictions, both_not_na) if not_na + ] + references = [ + ref for ref, not_na in zip(references, both_not_na) if not_na + ] + + metric_computation_functions = { + "cosine": cs, + "llm_sim": llm_sim, + "rouge": rouge, + "bleu": bleu, + "meteor": meteor, + "bertscore": bertscore, + "bleurt": bleurt + } + + for metric_name in metric_names: + if metric_name in metric_computation_functions: + if metric_name == "bertscore": + score = metric_computation_functions[metric_name].compute( + predictions=predictions, + references=references, + lang="en" + ) + metrics[metric_name] = { + "precision": np.array(score["precision"]).mean(), + "recall": np.array(score["recall"]).mean(), + "f1": np.array(score["f1"]).mean() + } + elif metric_name == "bleurt": + if isinstance(references[0], list): + grouped_references = list(zip(*references)) + scores = [] + for grouped_reference in grouped_references: + score = \ + metric_computation_functions[metric_name].compute( + predictions=predictions, + references=grouped_reference) + scores.append(score["scores"]) + scores = np.array(scores) # (num_refs, num_preds) + score = np.mean(np.max(scores, axis=0)) + else: + score = metric_computation_functions[metric_name].compute( + predictions=predictions, + references=references) + score = np.mean(score["scores"]) + metrics[metric_name] = {"bleurt": score} + else: + metrics[metric_name] = \ + metric_computation_functions[metric_name].compute( + predictions=predictions, + references=references + ) + + return metrics + + +def flatten_metrics(metric_dict: dict): + """Flatten the metric dictionary for easy display.""" + flat_metrics = {} + if "meteor" in metric_dict: + flat_metrics["METEOR"] = metric_dict["meteor"]["meteor"] + if "cosine" in metric_dict: + flat_metrics["Cosine Embedding"] = metric_dict["cosine"]["cosine"] + if "bleu" in metric_dict: + flat_metrics["BLEU"] = metric_dict["bleu"]["bleu"] + flat_metrics["Precision-1"] = metric_dict["bleu"]["precisions"][0] + flat_metrics["Precision-2"] = metric_dict["bleu"]["precisions"][1] + flat_metrics["Length Ratio"] = metric_dict["bleu"]["length_ratio"] + if "rouge" in metric_dict: + flat_metrics["ROUGE-1"] = metric_dict["rouge"]["rouge1"] + flat_metrics["ROUGE-2"] = metric_dict["rouge"]["rouge2"] + if "nahit" in metric_dict: + flat_metrics["N/A-precision"] = metric_dict["nahit"]["precision"] + flat_metrics["N/A-recall"] = metric_dict["nahit"]["recall"] + flat_metrics["N/A-f1"] = metric_dict["nahit"]["f1"] + flat_metrics["N/A in pred"] = metric_dict["nahit"]["pred_ratio"] + flat_metrics["N/A in ref"] = metric_dict["nahit"]["ref_ratio"] + if "bertscore" in metric_dict: + flat_metrics["BERTScore-precision"] = metric_dict["bertscore"][ + "precision"] + flat_metrics["BERTScore-recall"] = metric_dict["bertscore"]["recall"] + flat_metrics["BERTScore-f1"] = metric_dict["bertscore"]["f1"] + if "bleurt" in metric_dict: + flat_metrics["BLEURT"] = metric_dict["bleurt"]["bleurt"] + if "llm_sim" in metric_dict: + flat_metrics["LLM Similarity"] = metric_dict["llm_sim"]["llm_sim"] + return flat_metrics + + +if __name__ == "__main__": + predictions_demo = ["The cat sat on the mat.", "The dog ate my homework."] + references_demo = [["The cat sat on the mat.", "The cat sat on the desk."], + ["The dog ate my homework.", "The dog ate my lunch."]] + + # Compute metrics + metrics_demo = compute_metrics(predictions=predictions_demo, + references=references_demo, + metric_names=["llm_sim"]) + + # Print results + print(json.dumps(metrics_demo, indent=2)) diff --git a/notebooks/evaluate.ipynb b/notebooks/evaluate.ipynb new file mode 100644 index 0000000..253b169 --- /dev/null +++ b/notebooks/evaluate.ipynb @@ -0,0 +1,1312 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "Evaluate the quality of extracted content by comparing it with the human references." + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "You try to use a model that was created with version 2.4.0.dev0, however, your version is 2.3.1. This might cause unexpected behavior or errors. In that case, try to update to the latest version.\n", + "\n", + "\n", + "\n", + "/Users/jimmy/miniforge3/envs/llm/lib/python3.8/site-packages/huggingface_hub/file_download.py:1132: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.\n", + " warnings.warn(\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "INFO:tensorflow:Reading checkpoint /Users/jimmy/.cache/huggingface/metrics/bleurt/BLEURT-20-D12/downloads/extracted/2b0bd60025f714bf0eca857470aa967f784a446243ab3666b88cb6794a07c374/BLEURT-20-D12.\n", + "INFO:tensorflow:Config file found, reading.\n", + "INFO:tensorflow:Will load checkpoint BLEURT-20-D12\n", + "INFO:tensorflow:Loads full paths and checks that files exists.\n", + "INFO:tensorflow:... name:BLEURT-20-D12\n", + "INFO:tensorflow:... bert_config_file:bert_config.json\n", + "INFO:tensorflow:... max_seq_length:512\n", + "INFO:tensorflow:... vocab_file:None\n", + "INFO:tensorflow:... do_lower_case:None\n", + "INFO:tensorflow:... sp_model:sent_piece\n", + "INFO:tensorflow:... dynamic_seq_length:True\n", + "INFO:tensorflow:Creating BLEURT scorer.\n", + "INFO:tensorflow:Creating SentencePiece tokenizer.\n", + "INFO:tensorflow:Creating SentencePiece tokenizer.\n", + "INFO:tensorflow:Will load model: /Users/jimmy/.cache/huggingface/metrics/bleurt/BLEURT-20-D12/downloads/extracted/2b0bd60025f714bf0eca857470aa967f784a446243ab3666b88cb6794a07c374/BLEURT-20-D12/sent_piece.model.\n", + "INFO:tensorflow:SentencePiece tokenizer created.\n", + "INFO:tensorflow:Creating Eager Mode predictor.\n", + "INFO:tensorflow:Loading model.\n", + "INFO:tensorflow:BLEURT initialized.\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "INFO:tensorflow:BLEURT initialized.\n", + "[nltk_data] Downloading package wordnet to /Users/jimmy/nltk_data...\n", + "[nltk_data] Package wordnet is already up-to-date!\n", + "[nltk_data] Downloading package punkt to /Users/jimmy/nltk_data...\n", + "[nltk_data] Package punkt is already up-to-date!\n", + "[nltk_data] Downloading package omw-1.4 to /Users/jimmy/nltk_data...\n", + "[nltk_data] Package omw-1.4 is already up-to-date!\n" + ] + } + ], + "source": [ + "import sys\n", + "sys.path.append(\"..\")\n", + "\n", + "import nest_asyncio\n", + "nest_asyncio.apply()\n", + "\n", + "import pandas as pd\n", + "import jsonlines as jl\n", + "\n", + "from massw.metrics import compute_metrics, flatten_metrics" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "metadata": {}, + "outputs": [], + "source": [ + "aspects = [\"context\", \"key_idea\", \"method\", \"outcome\", \"future_impact\"]\n", + "annotation_path = \"../data/annotation_0531.jsonl\"\n", + "prediction_models = [\"gpt4\", \"gpt35\", \"mistral\"]\n", + "prediction_paths = {model: f\"../data/{model}_0531.jsonl\" for model in prediction_models}" + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Length of annotation records: 240\n", + "Length of predictions for gpt4: 120\n", + "Length of predictions for gpt35: 120\n", + "Length of predictions for mistral: 120\n" + ] + } + ], + "source": [ + "predictions = {}\n", + "\n", + "with jl.open(annotation_path) as f:\n", + " annotations = []\n", + " for line in f:\n", + " annotations.append({\n", + " \"id\": line[\"id\"],\n", + " \"texts\": line[\"displayed_text\"],\n", + " \"context\": line[\"label_annotations\"][\"Multi-aspect Summary\"][\"Context\"],\n", + " \"key_idea\": line[\"label_annotations\"][\"Multi-aspect Summary\"][\"Key idea\"],\n", + " \"method\": line[\"label_annotations\"][\"Multi-aspect Summary\"][\"Method\"],\n", + " \"outcome\": line[\"label_annotations\"][\"Multi-aspect Summary\"][\"Outcome\"],\n", + " \"future_impact\": line[\"label_annotations\"][\"Multi-aspect Summary\"][\"Future Impact\"],\n", + " })\n", + "\n", + "for model in prediction_models:\n", + " prediction_path = prediction_paths[model]\n", + " with jl.open(prediction_path) as f:\n", + " predictions[model] = []\n", + " for line in f:\n", + " predictions[model].append({\n", + " \"id\": line[\"id\"],\n", + " \"context\": line[\"Context\"],\n", + " \"key_idea\": line[\"Key Idea\"],\n", + " \"method\": line[\"Method\"],\n", + " \"outcome\": line[\"Outcome\"],\n", + " \"future_impact\": line[\"Future Impact\"],\n", + " })\n", + "\n", + "print(\"Length of annotation records:\", len(annotations))\n", + "for model in prediction_models:\n", + " print(f\"Length of predictions for {model}:\", len(predictions[model]))" + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Length of annotation records after filtering: 240\n" + ] + } + ], + "source": [ + "# Preserve annotations labeled by at least 2 annotators\n", + "id_counts = pd.Series([a[\"id\"] for a in annotations]).value_counts()\n", + "annotations = [a for a in annotations if id_counts[a[\"id\"]] >= 2]\n", + "print(\"Length of annotation records after filtering:\", len(annotations))" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Number of common records: 120\n" + ] + } + ], + "source": [ + "# Find intersection of annotations and predictions\n", + "annotation_ids = set([a[\"id\"] for a in annotations])\n", + "prediction_ids = {model: set([p[\"id\"] for p in predictions[model]]) for model in prediction_models}\n", + "common_ids = annotation_ids\n", + "for model in prediction_models:\n", + " common_ids = common_ids.intersection(prediction_ids[model])\n", + "print(\"Number of common records:\", len(common_ids))" + ] + }, + { + "cell_type": "code", + "execution_count": 22, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Computing metrics for gpt4\n", + "Computing metrics for gpt4 on context\n", + "Computing metrics for gpt4 on key_idea\n", + "Computing metrics for gpt4 on method\n", + "Computing metrics for gpt4 on outcome\n", + "Computing metrics for gpt4 on future_impact\n", + "Computing metrics for gpt35\n", + "Computing metrics for gpt35 on context\n", + "Computing metrics for gpt35 on key_idea\n", + "Computing metrics for gpt35 on method\n", + "Computing metrics for gpt35 on outcome\n", + "Computing metrics for gpt35 on future_impact\n", + "Computing metrics for mistral\n", + "Computing metrics for mistral on context\n", + "Computing metrics for mistral on key_idea\n", + "Computing metrics for mistral on method\n", + "Computing metrics for mistral on outcome\n", + "Computing metrics for mistral on future_impact\n" + ] + } + ], + "source": [ + "import warnings\n", + "warnings.filterwarnings(\"ignore\")\n", + "# Compute all metrics\n", + "# Model as prediction source, and annotation as ground truth references\n", + "\n", + "metrics = {}\n", + "\n", + "refs = {}\n", + "for aspect in aspects:\n", + " refs[aspect] = []\n", + " for idx in common_ids:\n", + " refs[aspect].append([a[aspect] for a in annotations if a[\"id\"] == idx])\n", + "\n", + "for model in prediction_models:\n", + " print(f\"Computing metrics for {model}\")\n", + " metrics[model] = {}\n", + " model_predictions = {}\n", + " for aspect in aspects:\n", + " model_predictions[aspect] = []\n", + " for idx in common_ids:\n", + " found = [p[aspect] for p in predictions[model] if p[\"id\"] == idx]\n", + " assert len(found) == 1\n", + " model_predictions[aspect].append(found[0])\n", + "\n", + " for aspect in aspects:\n", + " print(f\"Computing metrics for {model} on {aspect}\")\n", + " metrics[model][aspect] = compute_metrics(\n", + " predictions=model_predictions[aspect],\n", + " references=refs[aspect],\n", + " metric_names=[\"nahit\"]\n", + " # metric_names=[\"bleurt\", \"cosine\", \"bertscore\", \"rouge\", \"bleu\"]\n", + " )" + ] + }, + { + "cell_type": "code", + "execution_count": 23, + "metadata": {}, + "outputs": [], + "source": [ + "import pandas as pd\n", + "\n", + "# View metrics\n", + "for model in prediction_models:\n", + " for aspect in aspects:\n", + " metrics[model][aspect] = flatten_metrics(metrics[model][aspect])" + ] + }, + { + "cell_type": "code", + "execution_count": 26, + "metadata": {}, + "outputs": [], + "source": [ + "# Set float precision\n", + "pd.set_option('display.precision', 3)\n", + "gpt4_df = pd.DataFrame(metrics[\"gpt4\"]).rename(columns={\"future_impact\": \"projected_impact\"})\n", + "gpt35_df = pd.DataFrame(metrics[\"gpt35\"]).rename(columns={\"future_impact\": \"projected_impact\"})\n", + "mistral_df = pd.DataFrame(metrics[\"mistral\"]).rename(columns={\"future_impact\": \"projected_impact\"})" + ] + }, + { + "cell_type": "code", + "execution_count": 27, + "metadata": {}, + "outputs": [ + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
N/A-precision1.0000.0000.5000.2670.932
N/A-recall0.5830.0000.4210.3640.923
N/A-f10.7370.0000.4570.3080.928
N/A in pred0.1170.0080.1330.1250.858
N/A in ref0.2000.0080.1580.0920.867
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "N/A-precision 1.000 0.000 0.500 0.267 0.932\n", + "N/A-recall 0.583 0.000 0.421 0.364 0.923\n", + "N/A-f1 0.737 0.000 0.457 0.308 0.928\n", + "N/A in pred 0.117 0.008 0.133 0.125 0.858\n", + "N/A in ref 0.200 0.008 0.158 0.092 0.867" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
N/A-precision0.00.0001.0000.5831.000
N/A-recall0.00.0000.1050.6360.346
N/A-f10.00.0000.1900.6090.514
N/A in pred0.00.0000.0170.1000.300
N/A in ref0.20.0080.1580.0920.867
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "N/A-precision 0.0 0.000 1.000 0.583 1.000\n", + "N/A-recall 0.0 0.000 0.105 0.636 0.346\n", + "N/A-f1 0.0 0.000 0.190 0.609 0.514\n", + "N/A in pred 0.0 0.000 0.017 0.100 0.300\n", + "N/A in ref 0.2 0.008 0.158 0.092 0.867" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
N/A-precision1.0000.0000.6670.2500.951
N/A-recall0.0420.0000.4210.3640.750
N/A-f10.0800.0000.5160.2960.839
N/A in pred0.0080.0080.1000.1330.683
N/A in ref0.2000.0080.1580.0920.867
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "N/A-precision 1.000 0.000 0.667 0.250 0.951\n", + "N/A-recall 0.042 0.000 0.421 0.364 0.750\n", + "N/A-f1 0.080 0.000 0.516 0.296 0.839\n", + "N/A in pred 0.008 0.008 0.100 0.133 0.683\n", + "N/A in ref 0.200 0.008 0.158 0.092 0.867" + ] + }, + "metadata": {}, + "output_type": "display_data" + } + ], + "source": [ + "display(gpt4_df)\n", + "display(gpt35_df)\n", + "display(mistral_df)" + ] + }, + { + "cell_type": "code", + "execution_count": 16, + "metadata": {}, + "outputs": [ + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
Cosine Embedding0.9400.9440.8940.9310.916
BLEURT0.6070.5820.5100.6030.611
BERTScore-f10.9340.9280.9080.9330.933
BLEU0.3840.3750.1970.3550.282
ROUGE-10.6040.5720.4500.5960.563
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "Cosine Embedding 0.940 0.944 0.894 0.931 0.916\n", + "BLEURT 0.607 0.582 0.510 0.603 0.611\n", + "BERTScore-f1 0.934 0.928 0.908 0.933 0.933\n", + "BLEU 0.384 0.375 0.197 0.355 0.282\n", + "ROUGE-1 0.604 0.572 0.450 0.596 0.563" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
Cosine Embedding0.9340.9360.8950.9280.876
BLEURT0.5970.5750.5100.6080.498
BERTScore-f10.9340.9270.9100.9340.905
BLEU0.5240.4390.1970.4520.170
ROUGE-10.6350.5820.4450.6260.371
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "Cosine Embedding 0.934 0.936 0.895 0.928 0.876\n", + "BLEURT 0.597 0.575 0.510 0.608 0.498\n", + "BERTScore-f1 0.934 0.927 0.910 0.934 0.905\n", + "BLEU 0.524 0.439 0.197 0.452 0.170\n", + "ROUGE-1 0.635 0.582 0.445 0.626 0.371" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
Cosine Embedding0.9440.9490.9050.9330.917
BLEURT0.6450.6360.5540.6740.635
BERTScore-f10.9460.9430.9200.9480.936
BLEU0.5900.5560.2950.6650.384
ROUGE-10.6930.6620.5090.7070.599
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "Cosine Embedding 0.944 0.949 0.905 0.933 0.917\n", + "BLEURT 0.645 0.636 0.554 0.674 0.635\n", + "BERTScore-f1 0.946 0.943 0.920 0.948 0.936\n", + "BLEU 0.590 0.556 0.295 0.665 0.384\n", + "ROUGE-1 0.693 0.662 0.509 0.707 0.599" + ] + }, + "metadata": {}, + "output_type": "display_data" + } + ], + "source": [ + "rows = [\"Cosine Embedding\", \"BLEURT\", \"BERTScore-f1\", \"BLEU\", \"ROUGE-1\"]\n", + "# rows = [\"Cosine Embedding\", \"BLEURT\", \"BERTScore-f1\", \"BLEU\", \"ROUGE-1\", \"LLM Similarity\"]\n", + "gpt4_df.loc[rows].to_csv(\"gpt4_quality.csv\", float_format=\"%.3f\")\n", + "gpt35_df.loc[rows].to_csv(\"gpt35_quality.csv\", float_format=\"%.3f\")\n", + "mistral_df.loc[rows].to_csv(\"mistral_quality.csv\", float_format=\"%.3f\")\n", + "\n", + "display(gpt4_df.loc[rows])\n", + "display(gpt35_df.loc[rows])\n", + "display(mistral_df.loc[rows])" + ] + }, + { + "cell_type": "code", + "execution_count": 11, + "metadata": {}, + "outputs": [], + "source": [ + "# Since every paper has two annotations, we can split the annotations into two\n", + "# groups by their ids.\n", + "annotations.sort(key=lambda x: x[\"id\"])\n", + "annotations_1 = annotations[::2]\n", + "annotations_2 = annotations[1::2]\n", + "\n", + "# Use one as reference and the other as prediction\n", + "cross_val_metrics = {}\n", + "for aspect in aspects:\n", + " cross_val_metrics[aspect] = flatten_metrics(\n", + " compute_metrics(\n", + " predictions=[a[aspect] for a in annotations_1],\n", + " references=[a[aspect] for a in annotations_2],\n", + " metric_names=[\"bleurt\", \"cosine\", \"bertscore\", \"rouge\", \"bleu\"]))\n" + ] + }, + { + "cell_type": "code", + "execution_count": 12, + "metadata": {}, + "outputs": [ + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
Cosine Embedding0.9350.9440.9000.9360.941
BLEU0.5940.4640.3570.6080.642
Precision-10.6940.5630.5240.6990.757
Precision-20.6080.4720.3730.6200.690
Length Ratio0.9911.1331.0311.0890.953
ROUGE-10.7030.6370.5400.7370.748
ROUGE-20.6330.5460.3810.6610.686
BERTScore-precision0.9420.9340.9220.9470.959
BERTScore-recall0.9430.9440.9260.9540.951
BERTScore-f10.9420.9380.9240.9500.955
BLEURT0.6560.6180.5590.6710.742
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "Cosine Embedding 0.935 0.944 0.900 0.936 0.941\n", + "BLEU 0.594 0.464 0.357 0.608 0.642\n", + "Precision-1 0.694 0.563 0.524 0.699 0.757\n", + "Precision-2 0.608 0.472 0.373 0.620 0.690\n", + "Length Ratio 0.991 1.133 1.031 1.089 0.953\n", + "ROUGE-1 0.703 0.637 0.540 0.737 0.748\n", + "ROUGE-2 0.633 0.546 0.381 0.661 0.686\n", + "BERTScore-precision 0.942 0.934 0.922 0.947 0.959\n", + "BERTScore-recall 0.943 0.944 0.926 0.954 0.951\n", + "BERTScore-f1 0.942 0.938 0.924 0.950 0.955\n", + "BLEURT 0.656 0.618 0.559 0.671 0.742" + ] + }, + "execution_count": 12, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "human_df = pd.DataFrame(cross_val_metrics).rename(columns={\"future_impact\": \"projected_impact\"})\n", + "human_df" + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "metadata": {}, + "outputs": [ + { + "data": { + "text/html": [ + "
\n", + "\n", + "\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
contextkey_ideamethodoutcomeprojected_impact
Cosine Embedding0.9350.9440.9000.9360.941
BLEURT0.6560.6180.5590.6710.742
BERTScore-f10.9420.9380.9240.9500.955
BLEU0.5940.4640.3570.6080.642
ROUGE-10.7030.6370.5400.7370.748
\n", + "
" + ], + "text/plain": [ + " context key_idea method outcome projected_impact\n", + "Cosine Embedding 0.935 0.944 0.900 0.936 0.941\n", + "BLEURT 0.656 0.618 0.559 0.671 0.742\n", + "BERTScore-f1 0.942 0.938 0.924 0.950 0.955\n", + "BLEU 0.594 0.464 0.357 0.608 0.642\n", + "ROUGE-1 0.703 0.637 0.540 0.737 0.748" + ] + }, + "metadata": {}, + "output_type": "display_data" + } + ], + "source": [ + "human_df.loc[rows].to_csv(\"human_agreement.csv\", float_format=\"%.3f\")\n", + "\n", + "display(human_df.loc[rows])" + ] + }, + { + "cell_type": "code", + "execution_count": 21, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "ID: ffd14676-a525-479f-a74e-2c5d3a85c510\n", + "[gpt4]\n", + "Context: Interest in parallel systems has been revived, focusing on computation through excitatory and inhibitory interactions in networks of neuron-like units, particularly for early stages of visual processing and the representation of small local fragments.\n", + "Key Idea: The paper tackles the challenge of representing shapes in parallel systems and proposes mechanisms for shape perception and visual attention, offering a novel interpretation of the Gestalt principle 'the whole is more than the sum of its parts'.\n", + "Method: N/A\n", + "Outcome: N/A\n", + "Future Impact: N/A\n", + "[gpt35]\n", + "Context: There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units. This paper considers the difficulties involved in representing shapes in parallel systems.\n", + "Key Idea: The authors suggest ways of representing shapes in parallel systems which provides a mechanism for shape perception and visual attention, and allows a novel interpretation of the Gestalt slogan.\n", + "Method: N/A\n", + "Outcome: N/A\n", + "Future Impact: The proposed mechanism for shape perception and visual attention can be extended to other domain-specific problems involving the representation of complex shapes such as in medical imaging, robotics, and autonomous driving.\n", + "[mistral]\n", + "Context: There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units. Representing shapes in parallel systems is a complex issue.\n", + "Key Idea: This paper considers the difficulties involved in representing shapes in parallel systems and suggests ways of overcoming them. It provides a mechanism for shape perception and visual attention.\n", + "Method: N/A\n", + "Outcome: N/A\n", + "Future Impact: The paper's proposed mechanism for shape perception and visual attention can allow a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.\n", + "[Human]\n", + "Context: In parallel systems, visual processing by having individual network units represent hypotheses about how small local fragments of the visual input should be interpreted and encode knowledge about the constraints between local interpretations. But the representational issues are more complex higher up in the visual system.\n", + "Key Idea: The author provides a mechanism for shape perception and visual attention which allows a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.\n", + "Method: N/A\n", + "Outcome: N/A\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: There has been a recent revival of interest in parallel systems in which computation is performed by excitatory and inhibitory interactions within a network of relatively simple, neuronlike units. However, higher up than small local fragments, the representational issues are complex in the visual system.\n", + "Key Idea: The authors suggest ways of overcoming the difficulties involved in representing shapes in parallel systems. They provide a mechanism for shape perception and visual attention.\n", + "Method: N/A\n", + "Outcome: The proposed mechanism allows a novel interpretation of the Gestalt slogan that the whole is more than the sum of its parts.\n", + "Future Impact: N/A\n", + "\n", + "\n", + "ID: 73e353a8-e0d6-466f-af93-6fccf38fcb18\n", + "[gpt4]\n", + "Context: Video-grounded dialogues require reasoning over dialogue context in a multi-turn setting, which is more complex than traditional visual question answering. Previous approaches mostly use dialogue context as simple text input without modelling the turn-level information flows.\n", + "Key Idea: The proposed approach discovers information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer, introducing a novel method to learn reasoning paths over this semantic graph.\n", + "Method: The method involves predicting a reasoning path over the semantic graph from the current turn through past dialogue turns containing additional visual cues needed to answer the current question. The reasoning model processes both visual and textual information through this path, and the propagated features are used to generate the answer.\n", + "Outcome: The experimental results demonstrate the effectiveness of the proposed method in video-grounded dialogues by showing how it efficiently uses semantic dependencies in dialogue to retrieve visual cues.\n", + "Future Impact: The method provides insights on how models can leverage semantic dependencies in dialogues to retrieve visual information, suggesting potential applications in enhancing video-grounded dialogue systems.\n", + "[gpt35]\n", + "Context: Video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting. Dialogue context is mostly used as a simple text input without modeling the inherent information flows at the turn level in previous approaches.\n", + "Key Idea: The authors propose to discover information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer, and introducing a new approach that learns to predict reasoning paths over this semantic graph.\n", + "Method: The authors construct semantic graphs based on dialogue context, and propose a model that predicts reasoning paths through the graph. Additionally, their reasoning model processes both visual and textual information through this reasoning path and the propagated features are used to generate the answer. \n", + "Outcome: Experimental results demonstrate the effectiveness of the proposed method.\n", + "Future Impact: N/A\n", + "[mistral]\n", + "Context: Previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modelling the inherent information flows at the turn level.\n", + "Key Idea: The authors propose to discover information flows among dialogue turns through a semantic graph constructed based on lexical components in each question and answer. They then introduce a new approach that learns to predict reasoning paths over this semantic graph.\n", + "Method: The authors construct a semantic graph based on lexical components in each question and answer. They then introduce a path prediction model that predicts a path from the current turn through past dialogue turns that contain additional visual cues to answer the current question. The reasoning model sequentially processes both visual and textual information through this reasoning path.\n", + "Outcome: The experimental results demonstrate the effectiveness of the proposed method and provide insights on how models use semantic dependencies in a dialogue context to retrieve visual cues.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: Compared to traditional visual question answering, video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting. Previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modelling the inherent information flows at the turn level.\n", + "Key Idea: The authors propose to introduce a new approach that learns to predict reasoning paths over this semantic graph for video-grounded dialogues.\n", + "Method: N/A\n", + "Outcome: The experimental results demonstrate the effectiveness of t he proposed method.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: Video-grounded dialogues require additional reasoning over dialogue context to answer questions in a multi-turn setting, but previous approaches to video-grounded dialogues mostly use dialogue context as a simple text input without modeling the inherent information flows at the turn level.\n", + "Key Idea: The author introduces an approach that a path prediction model predicts a path from the current turn through past dialogue turns that contain additional visual cues to answer the current question, and than a reasoning model processes both visual and textual information through this reasoning path and the propagated features are used to generate the answer.\n", + "Method: N/A\n", + "Outcome: The proposed method is demonstrated as effective, and insights on how models use semantic dependencies in a dialogue context to retrieve visual cues is are provided.\n", + "Future Impact: N/A\n", + "\n", + "\n", + "ID: 14b0ebd1-b654-4eed-bdd8-ebeb74250b15\n", + "[gpt4]\n", + "Context: In few-shot relational triple extraction (FS-RTE), traditional approaches follow an entity-then-relation paradigm which first extracts all entities and then classifies their relations, overlooking the entity discrepancy between different relations.\n", + "Key Idea: The authors propose a novel task decomposition strategy for FS-RTE called Relation-then-Entity, which first detects the relations in a sentence and then extracts the corresponding head/tail entities of these relations. A model named RelATE is further introduced to implement this strategy.\n", + "Method: RelATE uses a dual-level attention mechanism to aggregate relation-relevant information for detecting relation occurrences and utilizes the annotated samples of the detected relations for extracting the corresponding head/tail entities.\n", + "Outcome: The RelATE model significantly outperforms previous works in FS-RTE, showcasing an absolute gain of 18.98% and 28.85% in F1 in two different few-shot settings.\n", + "Future Impact: N/A\n", + "[gpt35]\n", + "Context: In the few-shot relational triple extraction (FS-RTE), one extracts relational triples from plain texts by utilizing only few annotated samples. The entity-then-relation paradigm, which first extracts all entities and then classifies their relations, ignores the entity discrepancy between relations.\n", + "Key Idea: The authors propose a novel task decomposition strategy, Relation-then- Entity, for FS-RTE and a model, RelATE to instantiate this strategy. RelATE builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities.\n", + "Method: The authors experimentally show that RelATE outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).\n", + "Outcome: RelATE outperforms previous models with an absolute improvement of 18.98% and 28.85% F1 scores.\n", + "Future Impact: N/A\n", + "[mistral]\n", + "Context: In few-shot relational triple extraction (FS-RTE), the goal is to extract relational triples from plain texts using only a few annotated samples. Previous work extracts all entities and then classifies their relations, which ignores the entity discrepancy between relations.\n", + "Key Idea: The authors propose a novel task decomposition strategy called Relation-then-Entity (RelATE) for FS-RTE. It first detects relations in a sentence and then extracts the corresponding head/tail entities of the detected relations.\n", + "Method: The authors propose a model called RelATE that builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and uses the annotated samples of the detected relations to extract the corresponding head/tail entities.\n", + "Outcome: The proposed model outperforms previous work by an absolute gain of 18.98% and 28.85% in F1 in two few-shot settings.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: In few-shot relational triple extraction, one seeks to extract relational triples from plain texts by utilizing few annotated samples. Recent work first extracts all entities and then classifies their relations, which ignores the entity discrepancy between relations.\n", + "Key Idea: The author first proposes a novel task decomposition strategy for FS-RTE, which detects relations occurred in a sentence and then extracts the corresponding head/tail entities of the detected relations. To instantiate this strategy, the author also proposes a model that builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities.\n", + "Method: The author assesses the performance of the proposed strategy against previous works.\n", + "Outcome: Experimental results show that the proposed model outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: Recent work in few-shot relational triple extraction (FS-RTE) first extracts all entities and then classifies their relations. Such an entity-then-relation paradigm ignores the entity discrepancy between relations.\n", + "Key Idea: The authors propose a novel task decomposition strategy, Relation-then-Entity, for FS-RTE, which first detects relations occurred in a sentence and then extracts the corresponding head/tail entities of the detected relations. They further propose a model, RelATE, which builds a dual-level attention to aggregate relation-relevant information to detect the relation occurrence and utilizes the annotated samples of the detected relations to extract the corresponding head/tail entities.\n", + "Method: N/A\n", + "Outcome: Experimental results show that our model outperforms previous work by an absolute gain (18.98%, 28.85% in F1 in two few-shot settings).\n", + "Future Impact: N/A\n", + "\n", + "\n", + "ID: ca53b2c4-2912-4515-aae6-938c3f268a60\n", + "[gpt4]\n", + "Context: The problem of detecting epidemic tendency by mining search logs.\n", + "Key Idea: The key idea is an algorithm that selects epidemic related queries/terms based on click-through information and adopts linear regression to model epidemic occurrences and frequencies of epidemic related terms in search logs.\n", + "Method: The effectiveness of the algorithm is tested by evaluating the correlation between the frequencies of epidemic related terms found in search logs and actual epidemic occurrences.\n", + "Outcome: The results demonstrate that the algorithm effectively identifies epidemic related terms that show a high correlation with epidemic occurrences. It also shows that combining different epidemic related terms improves performance compared to using a single term.\n", + "Future Impact: N/A\n", + "[gpt35]\n", + "Context: The authors address the problem of detecting epidemic tendency by mining search logs.\n", + "Key Idea: The authors propose an algorithm to select epidemic related queries/terms by using click-through information. They adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms in search logs.\n", + "Method: The authors use click-through information to select epidemic-related terms, followed by linear regression to model epidemic occurrences and the frequencies of epidemic-related terms in search logs. They compare the performance of combining different epidemic-related terms against using a single epidemic-related term.\n", + "Outcome: The proposed algorithm is effective in identifying ERTs that show a high correlation with epidemic occurrences. The authors show that combining different ERTs performs better than using a single ERT.\n", + "Future Impact: N/A\n", + "[mistral]\n", + "Context: The problem of detecting epidemic tendency by mining search logs, and the need for an algorithm that can select epidemic related queries/terms.\n", + "Key Idea: The authors propose an algorithm that uses click-through information to select epidemic related queries/terms, and a method to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search logs using linear regression.\n", + "Method: The authors perform experiments to test the effectiveness of the algorithm in finding ERTs which obtain a high correlation value with epidemic occurrences.\n", + "Outcome: The proposed algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. The proposed method performs better when combining different ERTs than using single ERT.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: N/A\n", + "Key Idea: The authors propose an algorithm based on click-through information to select epidemic related queries/terms, and adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search log.\n", + "Method: N/A\n", + "Outcome: The results show the proposed algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. The proposed method performs better when combining different ERTs than using single ERT.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: The problem of detecting epidemic tendency is important.\n", + "Key Idea: Authors propose an algorithm based on click-through information to select epidemic related queries/terms. Authors adopt linear regression to model epidemic occurrences and frequencies of epidemic related terms (ERTs) in search logs.\n", + "Method: Authors design experiments to find epidemic related terms (ERTs).\n", + "Outcome: Experimental results show the proposed algorithm is effective in finding ERTs which obtain a high correlation value with epidemic occurrences. Authors also find the proposed method performs better when combining different ERTs than using single ERT.\n", + "Future Impact: N/A\n", + "\n", + "\n", + "ID: 5e1f387c-d883-4d1f-8397-e4a533a3387b\n", + "[gpt4]\n", + "Context: Hyperspectral imaging is crucial for various applications but is hindered by degradation issues and the scarcity of accurate ground-truth signals, making the training of deep neural networks for restoration tasks challenging.\n", + "Key Idea: The paper introduces a hybrid approach based on sparse coding principles that combines the interpretability of classical techniques with the ability to train model parameters end-to-end without large datasets.\n", + "Method: The efficacy of the proposed method is demonstrated through various denoising benchmarks.\n", + "Outcome: The proposed method is computationally efficient and significantly outperforms state-of-the-art approaches in hyperspectral image restoration.\n", + "Future Impact: N/A\n", + "[gpt35]\n", + "Context: Hyperspectral imaging suffers from various sources of degradations, and lack of accurate ground-truth 'clean' hyperspectral signals makes restoration tasks challenging. Traditional neural networks struggle to train correctly due to limited available data.\n", + "Key Idea: The authors propose a new approach to hyperspectral image restoration based on sparse coding principles, merging classical techniques with deep learning to allow training of parameters without significant data requirements.\n", + "Method: The proposed spectral-spatial sparse coding model is applied to various denoising benchmarks and is shown to be computationally efficient while significantly outperforming the current state of the art.\n", + "Outcome: The proposed spectral-spatial sparse coding model is shown to be highly effective in improving denoising benchmarks.\n", + "Future Impact: The hybrid approach based on sparse coding principles proposed in this paper is a promising avenue for future research in hyperspectral imaging restoration and may be useful for other applications with limited training data.\n", + "[mistral]\n", + "Context: Hyperspectral imaging is used in various fields but it is affected by different sources of degradation, and the lack of accurate ground-truth makes restoration tasks challenging. Training deep neural networks for restoration is difficult in contrast to traditional RGB imaging problems.\n", + "Key Idea: The authors propose a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data.\n", + "Method: The authors perform experiments on various denoising benchmarks to show the computational efficiency and the performance of the proposed method.\n", + "Outcome: The proposed method is computationally efficient and significantly outperforms the state of the art on various denoising benchmarks.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: Hyperspectral imaging offers new perspectives for diverse applications, but the spectral diversity of information comes at the expense of various sources of degradation, and the lack of accurate ground-truth \"clean\" hyperspectral signals acquired on the spot makes restoration tasks challenging.\n", + "Key Idea: The authors propose a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data.\n", + "Method: The authors evaluate their proposed method on various denoising benchmarks.\n", + "Outcome: The authors show on various denoising benchmarks that the proposed method is computationally efficient and significantly outperforms the state of the art.\n", + "Future Impact: N/A\n", + "[Human]\n", + "Context: Hyperspectral imaging offers new perspectives for diverse applications. Unfortunately, the spectral diversity of information comes at the expense of various sources of degradation, and the lack of accurate ground-truth \"clean\" hyperspectral signals acquired on the spot makes restoration tasks challenging. In particular, training deep neural networks for restoration is difficult, in contrast to traditional RGB imaging problems.\n", + "Key Idea: Authors advocate instead for a hybrid approach based on sparse coding principles that retain the interpretability of classical techniques encoding domain knowledge with handcrafted image priors, while allowing to train model parameters end-to-end without massive amounts of data.\n", + "Method: Authors conduct experiments on various denoising benchmarks.\n", + "Outcome: Authors show that, on various denoising benchmarks the proposed method is computationally efficient and significantly outperforms the state of the art.\n", + "Future Impact: N/A\n", + "\n", + "\n" + ] + } + ], + "source": [ + "# Load a few examples for manual inspection\n", + "ids = list(common_ids)[:5]\n", + "for idx in ids:\n", + " print(\"ID:\", idx)\n", + " for model in prediction_models:\n", + " found = [p for p in predictions[model] if p[\"id\"] == idx]\n", + " assert len(found) == 1\n", + " print(f\"[{model}]\")\n", + " print(\"Context:\", found[0][\"context\"])\n", + " print(\"Key Idea:\", found[0][\"key_idea\"])\n", + " print(\"Method:\", found[0][\"method\"])\n", + " print(\"Outcome:\", found[0][\"outcome\"])\n", + " print(\"Future Impact:\", found[0][\"future_impact\"])\n", + " human_annotations = [a for a in annotations if a[\"id\"] == idx]\n", + " for a in human_annotations:\n", + " print(\"[Human]\")\n", + " print(\"Context:\", a[\"context\"])\n", + " print(\"Key Idea:\", a[\"key_idea\"])\n", + " print(\"Method:\", a[\"method\"])\n", + " print(\"Outcome:\", a[\"outcome\"])\n", + " print(\"Future Impact:\", a[\"future_impact\"])\n", + "\n", + " print(\"\\n\")" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "llm", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.8.17" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/notebooks/evaluation_results/gpt35_quality.csv b/notebooks/evaluation_results/gpt35_quality.csv new file mode 100644 index 0000000..e16fd9d --- /dev/null +++ b/notebooks/evaluation_results/gpt35_quality.csv @@ -0,0 +1,6 @@ +,context,key_idea,method,outcome,projected_impact +Cosine Embedding,0.934,0.936,0.895,0.928,0.876 +BLEURT,0.597,0.575,0.510,0.608,0.498 +BERTScore-f1,0.934,0.927,0.910,0.934,0.905 +BLEU,0.524,0.439,0.197,0.452,0.170 +ROUGE-1,0.635,0.582,0.445,0.626,0.371 diff --git a/notebooks/evaluation_results/gpt4_quality.csv b/notebooks/evaluation_results/gpt4_quality.csv new file mode 100644 index 0000000..9ef5c02 --- /dev/null +++ b/notebooks/evaluation_results/gpt4_quality.csv @@ -0,0 +1,6 @@ +,context,key_idea,method,outcome,projected_impact +Cosine Embedding,0.940,0.944,0.894,0.931,0.916 +BLEURT,0.607,0.582,0.510,0.603,0.611 +BERTScore-f1,0.934,0.928,0.908,0.933,0.933 +BLEU,0.384,0.375,0.197,0.355,0.282 +ROUGE-1,0.604,0.572,0.450,0.596,0.563 diff --git a/notebooks/evaluation_results/human_agreement.csv b/notebooks/evaluation_results/human_agreement.csv new file mode 100644 index 0000000..a4fe116 --- /dev/null +++ b/notebooks/evaluation_results/human_agreement.csv @@ -0,0 +1,6 @@ +,context,key_idea,method,outcome,projected_impact +Cosine Embedding,0.935,0.944,0.900,0.936,0.941 +BLEURT,0.656,0.618,0.559,0.671,0.742 +BERTScore-f1,0.942,0.938,0.924,0.950,0.955 +BLEU,0.594,0.464,0.357,0.608,0.642 +ROUGE-1,0.703,0.637,0.540,0.737,0.748 diff --git a/notebooks/evaluation_results/mistral_quality.csv b/notebooks/evaluation_results/mistral_quality.csv new file mode 100644 index 0000000..8a581ea --- /dev/null +++ b/notebooks/evaluation_results/mistral_quality.csv @@ -0,0 +1,6 @@ +,context,key_idea,method,outcome,projected_impact +Cosine Embedding,0.944,0.949,0.905,0.933,0.917 +BLEURT,0.645,0.636,0.554,0.674,0.635 +BERTScore-f1,0.946,0.943,0.920,0.948,0.936 +BLEU,0.590,0.556,0.295,0.665,0.384 +ROUGE-1,0.693,0.662,0.509,0.707,0.599 diff --git a/notebooks/generate_table.ipynb b/notebooks/generate_table.ipynb new file mode 100644 index 0000000..242f7b8 --- /dev/null +++ b/notebooks/generate_table.ipynb @@ -0,0 +1,82 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "import pandas as pd\n", + "\n", + "def load_data(file_path):\n", + " # Load CSV file and set the first column as the index\n", + " return pd.read_csv(file_path, index_col=0)\n", + "\n", + "def create_latex_table(human_agreement, gpt35_quality, gpt4_quality, mistral_quality):\n", + " metrics = human_agreement.index.tolist()\n", + " aspects = ['Context', 'Key Idea', 'Method', 'Outcome', 'Projected Impact']\n", + "\n", + " # Start of the LaTeX table\n", + " latex_table = \"\"\"\n", + "\\\\begin{tabular}{lcccccc}\n", + "\\\\toprule\n", + " & Aspects & CS & BS & BLEURT & BLEU & ROUGE-1 \\\\\\\\\n", + "\\\\midrule\n", + "\"\"\"\n", + "\n", + " # Add data rows for Human Agreement\n", + " latex_table += \"\\\\multirow{5}{*}{\\\\shortstack[l]{Human\\\\Agreement}} \"\n", + " for aspect in aspects:\n", + " scores = human_agreement[aspect].values\n", + " latex_table += f\" & {aspect} & {' & '.join(map(str, scores))} \\\\\\\\\\n\"\n", + " if aspect != aspects[-1]:\n", + " latex_table += \" \"\n", + " latex_table += \"\\\\hdashline\\n\"\n", + "\n", + " # Function to add model data rows\n", + " def add_model_data(model_name, model_data):\n", + " latex_table = f\"\\\\multirow{{5}}{{*}}{{\\\\shortstack[l]{{{model_name}\\\\\\Alignment}}} \"\n", + " for aspect in aspects:\n", + " scores = model_data[aspect].values\n", + " latex_table += f\" & {aspect} & {' & '.join(map(str, scores))} \\\\\\\\\\n\"\n", + " if aspect != aspects[-1]:\n", + " latex_table += \" \"\n", + " latex_table += \"\\\\hdashline\\n\"\n", + " return latex_table\n", + "\n", + " # Add data rows for each model\n", + " latex_table += add_model_data(\"GPT-3.5-Human\", gpt35_quality)\n", + " latex_table += add_model_data(\"GPT-4-Human\", gpt4_quality)\n", + " latex_table += add_model_data(\"Mixtral-8x7B-Human\", mistral_quality)\n", + "\n", + " # End of the LaTeX table\n", + " latex_table += \"\\\\bottomrule\\n\\\\end{tabular}\\n\"\n", + "\n", + " return latex_table\n", + "\n", + "# Load data from CSV files\n", + "human_agreement = load_data('/mnt/data/human_agreement.csv')\n", + "gpt35_quality = load_data('/mnt/data/gpt35_quality.csv')\n", + "gpt4_quality = load_data('/mnt/data/gpt4_quality.csv')\n", + "mistral_quality = load_data('/mnt/data/mistral_quality.csv')\n", + "\n", + "# Generate the updated LaTeX table\n", + "new_latex_table = create_latex_table(human_agreement, gpt35_quality, gpt4_quality, mistral_quality)\n", + "print(new_latex_table)\n" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "llm", + "language": "python", + "name": "python3" + }, + "language_info": { + "name": "python", + "version": "3.8.17" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/notebooks/prompt.txt b/notebooks/prompt.txt new file mode 100644 index 0000000..dea1f57 --- /dev/null +++ b/notebooks/prompt.txt @@ -0,0 +1,55 @@ +## Instructions +You are an expert in computer science. +Your task is to summarize the following five aspects of the papers given the +definitions below. + +## Definitions of Aspects +### Context +- The status quo of related literature or reality which motivated this study. +This could normally be a problem, a research question, or a research gap that +has not been successfully addressed by previous work. +- Anything happened before this study. +### Key Idea +- The main intellectual merit of this paper, often in comparison to the +context. This could normally be a novel idea or solution proposed in this paper +that distinguishes it from what’s already done in literature. +- Proposed in this study. +### Method (Validation Methodology) +- The specific experiment or proof that investigates and validates the key idea. +- CS papers often refer "Method" as algorithm or model, but our definition here is **different**. +- Performed in this study. +### Outcome +- The factual statement about the study output. This could be the experiment +results and any other measurable outcome that has occurred. It marks whether +the key hypothesis is testified or not. +- Produced in this study. +### Future Impact +- The impact of the work on the field explicitly anticipated by the authors, +and potential further research explicitly identified by the author that may improve or +extend this study. + +## Notes +- If an aspect is NOT mentioned in the abstract, mark it as “N/A” (not +applicable). DO NOT come up with your own interpretation. +- Each aspect should be summarized in 1-2 sentences in most cases. +- Each aspect should be self-contained and should not contain references + including other aspects (cross-reference). +- Including specific names of proposed models, datasets, etc., in the summary is acceptable. +- If the problem definition is novel (e.g., proposing a new task), classify it as a Key Idea. +- Non-measurable outcomes should be categorized as Future Impact. +- Impacts that have already occurred should be considered as Outcome. +- A new observation that motivates the proposal of a key idea should be classified under Key Idea. +- Future Impact should not account for real impacts, such as the number of +citations a paper has received. + +## Response Format +The response should be a JSON object in the following format: +```json +{ + "Context": "...", + "Key Idea": "...", + "Method": "...", + "Outcome": "...", + "Future Impact": "..." +} +``` diff --git a/notebooks/summarization_baseline.ipynb b/notebooks/summarization_baseline.ipynb new file mode 100644 index 0000000..e9c7fe1 --- /dev/null +++ b/notebooks/summarization_baseline.ipynb @@ -0,0 +1,332 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 1, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "240\n" + ] + } + ], + "source": [ + "from massw import api_gpt as api\n", + "import pandas as pd\n", + "import jsonlines as jl\n", + "import json\n", + "\n", + "with jl.open(\"../data/annotation_0531.jsonl\") as f:\n", + " annotated_instances = list(f)\n", + "print(len(annotated_instances))\n", + "# Process the annotated instances\n", + "df_annotated = pd.DataFrame(annotated_instances)\n", + "df_annotated.set_index(\"id\", inplace=True)\n", + "df_annotated[\"Context\"] = df_annotated[\"label_annotations\"].apply(lambda x: x[\"Multi-aspect Summary\"][\"Context\"])\n", + "df_annotated[\"Key Idea\"] = df_annotated[\"label_annotations\"].apply(lambda x: x[\"Multi-aspect Summary\"][\"Key idea\"])\n", + "df_annotated[\"Method\"] = df_annotated[\"label_annotations\"].apply(lambda x: x[\"Multi-aspect Summary\"][\"Method\"])\n", + "df_annotated[\"Outcome\"] = df_annotated[\"label_annotations\"].apply(lambda x: x[\"Multi-aspect Summary\"][\"Outcome\"])\n", + "df_annotated[\"Future Impact\"] = df_annotated[\"label_annotations\"].apply(lambda x: x[\"Multi-aspect Summary\"][\"Future Impact\"])\n", + "df_annotated.drop(columns=[\"label_annotations\", \"span_annotations\"], inplace=True)\n", + "ids = df_annotated.index.unique()\n", + "texts = {id: df_annotated.reset_index().drop_duplicates(\"id\").set_index(\"id\").loc[id, \"displayed_text\"] for id in ids}\n", + "azure_config = api.AzureConfig()\n", + "batch = api.Batch(tpm=40000, azure=azure_config)\n", + "\n", + "with open(\"prompt.txt\", \"r\") as f:\n", + " system_prompt = f.read()\n", + "\n", + "example_prompt = \"\"\"\n", + "Title: Attention Is All You Need\n", + "Abstract: The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.\n", + "\"\"\"\n", + "\n", + "example_output = \"\"\"{\n", + " \"Context\": \"The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing sequence transduction models connect the encoder and decoder through an attention mechanism.\",\n", + " \"Key Idea\": \"The authors propose a simple network architecture called Transformer based solely on attention mechanisms and dispenses with recurrence and convolutions.\",\n", + " \"Method\": \"The authors perform experiments on the WMT 2014 English-to-German and English-to-French translation task. The authors apply the proposed model to English constituency parsing both with large and limited training data.\",\n", + " \"Outcome\": \"The proposed model achieves a BLEU score of 28.4 on the WMT 2014 English-to-French translation task. The proposed model achieves a BLEU score of 41.8 on the WMT 2014 English-to-German translation task after training for 3.5 days on 8 GPUs.\",\n", + " \"Future Impact\": \"N/A\"\n", + "}\"\"\"\n" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "\n", + "for i in ids:\n", + " await batch.add(\n", + " \"chat.completions.create\",\n", + " model=\"gpt-35-turbo\",\n", + " # response_format={ \"type\": \"json_object\" },\n", + " messages=[\n", + " {\"role\": \"system\", \"content\": system_prompt},\n", + " {\"role\": \"user\", \"content\": example_prompt},\n", + " {\"role\": \"assistant\", \"content\": example_output},\n", + " {\"role\": \"user\", \"content\": texts[i]},\n", + " ],\n", + " metadata={\"id\": i},\n", + " )\n", + "\n", + "results = await batch.run()" + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "9821d5f7-72b0-4841-a54f-d2af4a04ea3a\n", + "{\n", + " \"Context\": \"The paper addresses the problem of inverse rendering where estimating the spherical harmonic illumination coefficients and texture parameters in a specular invariant colour subspace is challenging.\",\n", + " \"Key Idea\": \"The authors propose a novel approach for inverse rendering based on a linear basis approximation of surface texture, which can account for non-Lambertian specular reflectance and complex illumination of the same light source colour.\",\n", + " \"Method\": \"The proposed approach involves fitting a 3D morphable model to a single colour image of faces through the solution of bilinear equations in a specular invariant colour subspace. This approach recovers statistical texture model parameters without relying on computationally expensive analysis-by-synthesis techniques.\",\n", + " \"Outcome\": \"The proposed approach recovers texture model parameters with an accuracy comparable to that of more computationally expensive methods, while requiring only the solution of convex optimization problems.\",\n", + " \"Future Impact\": \"The proposed approach could be extended to other objects besides faces, potentially offering a more efficient and accurate solution to the problem of inverse rendering.\",\n", + "}\n" + ] + } + ], + "source": [ + "# id2result\n", + "id2result = {}\n", + "for _, result in results.iterrows():\n", + " content = result[\"result\"][\"choices\"][0][\"message\"][\"content\"]\n", + " # Remove \"\"\"json and \"\"\" if present\n", + " content = content.replace('\"\"\"json', \"\").replace('\"\"\"', \"\")\n", + " try:\n", + " id2result[result[\"id\"]] = json.loads(content)\n", + " except Exception as e:\n", + " print(result[\"id\"])\n", + " print(content)" + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "metadata": {}, + "outputs": [], + "source": [ + "with jl.open(\"../data/gpt35_0531.jsonl\", \"w\") as f:\n", + " f.write_all([{\"id\": k, **v} for k, v in id2result.items()])" + ] + }, + { + "cell_type": "code", + "execution_count": 17, + "metadata": {}, + "outputs": [], + "source": [ + "# Mistral baselines\n", + "import ssl\n", + "import os\n", + "import urllib.request\n", + "\n", + "\n", + "def allowSelfSignedHttps(allowed):\n", + " # Bypass the server certificate verification on client side\n", + " if allowed and not os.environ.get('PYTHONHTTPSVERIFY', '') and getattr(\n", + " ssl, '_create_unverified_context', None):\n", + " ssl._create_default_https_context = ssl._create_unverified_context\n", + "\n", + "\n", + "# Enabling self-signed certificates if required\n", + "allowSelfSignedHttps(True)\n", + "\n", + "\n", + "def mistral_request(messages):\n", + " data = {\n", + " \"input_data\": {\n", + " \"input_string\": messages,\n", + " \"parameters\": {\n", + " \"temperature\": 0.6,\n", + " \"top_p\": 0.9,\n", + " \"do_sample\": True,\n", + " \"max_new_tokens\": 500,\n", + " \"return_full_text\": True\n", + " }\n", + " }\n", + " }\n", + " url = 'https://xingjian-ml-apqyj.eastus.inference.ml.azure.com/score'\n", + " api_key = '46g9IOYuQYjwCjlHzwEy9lBOJfQHrjZO' # Mistral API key\n", + " headers = {\n", + " 'Content-Type': 'application/json',\n", + " 'Authorization': ('Bearer ' + api_key),\n", + " 'azureml-model-deployment': 'mistralai-mixtral-8x7b-instru-7'\n", + " }\n", + " body = str.encode(json.dumps(data))\n", + "\n", + " req = urllib.request.Request(url, body, headers)\n", + " try:\n", + " response = urllib.request.urlopen(req)\n", + " result = response.read()\n", + " response = json.loads(result)[\"output\"]\n", + " except urllib.error.HTTPError as error:\n", + " print(\"The request failed with status code: \" + str(error.code))\n", + "\n", + " # Print the headers - they include the requert ID and the timestamp, which are useful for debugging the failure\n", + " print(error.info())\n", + " print(error.read().decode(\"utf8\", 'ignore'))\n", + " return response\n", + "\n", + "\n", + "def get_mistral_result(text):\n", + " messages = [\n", + " {\"role\": \"user\", \"content\": system_prompt},\n", + " {\"role\": \"assistant\", \"content\": \"I understand. Please give me futher information.\"},\n", + " {\"role\": \"user\", \"content\": example_prompt},\n", + " {\"role\": \"assistant\", \"content\": example_output},\n", + " {\"role\": \"user\", \"content\": text},\n", + " ]\n", + " return mistral_request(messages)" + ] + }, + { + "cell_type": "code", + "execution_count": 18, + "metadata": {}, + "outputs": [ + { + "data": { + "application/vnd.jupyter.widget-view+json": { + "model_id": "aa07372be3eb44aaab1ac4d135c58645", + "version_major": 2, + "version_minor": 0 + }, + "text/plain": [ + " 0%| | 0/120 [00:00