Probabilistic Abduction for Visual Abstract Reasoning via Learning Rules in Vector-symbolic Architectures

Michael Hersche, Francesco di Stefano, Thomas Hofmann, Abu Sebastian, Abbas Rahimi

3rd MATH-AI Workshop at NeurIPS'23

Requirements

Hardware

You will need a machine with a CUDA-enabled GPU and the Nvidia SDK installed to compile the CUDA kernels. We tested our methods on an NVIDA Tesla V100 GPU with CUDA Version 11.3.1.

Installing Dependencies

The conda software is required for running the code. Generate a new environment with

$ conda create --name learnVRFenv python=3.7
$ conda activate learnVRFenv

We need PyTorch 1.11 and CUDA.

$ (learnVRFenv) conda install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch -c conda-forge
$ (learnVRFenv) pip install -r requirements.txt

As a last requirement, you need to install the neuro-vsa by cloning the code and running

$ (learnVRFenv) pip install -e /path/to/neuro-vsa/ --no-dependencies

Make sure to activate the environment before running any code.

I-RAVEN Dataset

Generate the I-RAVEN dataset with the instructions proveded here. Save it under dataset/.

Prepare Data

Run the rule preprocessing script:

$ (learnVRFenv) python3 src/util/extraction.py --path dataset/

Run Model

You can run the training and evaluation on the I-RAVEN dataset for all constellation by running one of the following scripts:

(learnVRFenv) experiments/learn_VRF_id.sh
(learnVRFenv) experiments/learn_VRF_ood.sh
(learnVRFenv) experiments/MLP_id.sh
(learnVRFenv) experiments/MLP_ood.sh

Citation

If you use the work released here for your research, please cite our paper:

@inproceedings{hersche2023learnVRF,
  title={Probabilistic Abduction for Visual Abstract Reasoning via Learning Rules in Vector-symbolic Architectures},
  author={Hersche, Michael and di Stefano, Francesco and Hofmann, Thomas and Sebastian, Abu and Rahimi, Abbas},
  journal={3rd Workshop on Math-AI (MATH-AI@ NeurIPS)},
  year={2023}
}

License

Please refer to the LICENSE file for the licensing of our code. Our implementation relies on PrAE released under GPL v3.0