GWU_DNSC 6290: Course Outline

Materials for a technical, nuts-and-bolts course about increasing transparency, fairness, robustness, and security in machine learning.

• Lecture 1: Explainable Machine Learning Models
• Lecture 2: Post-hoc Explanation
• Lecture 3: Bias Testing and Remediation
• Lecture 4: Machine Learning Security
• Lecture 5: Machine Learning Model Debugging
• Lecture 6: Responsible Machine Learning Best Practices
• Lecture 7: Risk Mitigation Proposals for Language Models

Corrections or suggestions? Please file a GitHub issue.

---

Lecture 1: Explainable Machine Learning Models

_{^{Source: Simple Explainable Boosting Machine Example}}

Lecture 1 Class Materials

• Introduction:
  • Syllabus
  • Basic Data Manipulation
  • Primer on Technical Malpractice
  • Whiteboard Notation
• Lecture Notes
• Assignment 1:
  • Model evaluation notebook
  • Full evaluations results
• Reading: Machine Learning for High-Risk Applications, Chapter 2 (pp. 33 - 50) and Chapter 6 (pp. 189 - 217)

Lecture 1 Additional Software Tools

• Python:

  • causalml
  • interpret
  • imodels
  • PiML-Toolbox
  • sklearn-expertsys
  • skope-rules
  • tensorflow/lattice

• R:

  • arules
  • elasticnet
  • gam
  • glmnet
  • quantreg
  • rpart
  • RuleFit

• Python, R or other:

  • h2o-3
  • Rudin Group code
  • xgboost

Lecture 1 Additional Software Examples

• Building from Penalized GLM to Monotonic GBM (simple)
• Building from Penalized GLM to Monotonic GBM
• Simple Explainable Boosting Machine Example
• PiML Assignment 1 Example and simple requirements.txt
• Machine Learning for High-risk Applications: Use Cases (Chapter 6)

Lecture 1 Additional Reading

• Introduction and Background:

  • An Introduction to Machine Learning Interpretability
  • Designing Inherently Interpretable Machine Learning Models
  • Psychological Foundations of Explainability and Interpretability in Artificial Intelligence
  • Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead

• Explainable Machine Learning Techniques:

  • Accurate Intelligible Models with Pairwise Interactions
  • Elements of Statistical Learning - Chapters 3,4, and 9
  • Fast Interpretable Greedy-Tree Sums (FIGS)
  • Interpretable Machine Learning - Chapter 5
  • GAMI-Net: An Explainable Neural Network Based on Generalized Additive Models with Structured Interactions
  • Neural Additive Models: Interpretable Machine Learning with Neural Nets
  • A Responsible Machine Learning Workflow with Focus on Interpretable Models, Post-hoc Explanation, and Discrimination Testing
  • This Looks Like That: Deep Learning for Interpretable Image Recognition
  • Unwrapping The Black Box of Deep ReLU Networks: Interpretability, Diagnostics, and Simplification

---

Lecture 2: Post-hoc Explanation

_{^{Source: Global and Local Explanations of a Constrained Model}}

Lecture 2 Class Materials

• Lecture Notes
• Assignment 2
• Reading: Machine Learning for High-Risk Applications, Chapter 2 (pp. 50 - 80) and Chapter 6 (pp. 208 - 230)

Lecture 2 Additional Software Tools

• Python:

  • allennlp
  • alibi
  • anchor
  • DiCE
  • interpret
  • lime
  • shap
  • PiML-Toolbox
  • tf-explain

• R:

  • ALEPlot
  • DALEX
  • ICEbox
  • iml
  • Model Oriented
  • pdp
  • shapFlex
  • vip

• Python, R or other:

  • h2o-3

Lecture 2 Additional Software Examples

• Global and Local Explanations of a Constrained Model
• Building from Penalized GLM to Monotonic GBM
• Monotonic XGBoost models, partial dependence, individual conditional expectation plots, and Shapley explanations
• Decision tree surrogates, LOCO, and ensembles of explanations
• Machine Learning for High-risk Applications: Use Cases (Chapter 6)

Lecture 2 Additional Reading

• Introduction and Background:

  • On the Art and Science of Explainable Machine Learning
  • Proposed Guidelines for the Responsible Use of Explainable Machine Learning

• Post-hoc Explanation Techniques:

  • A Unified Approach to Interpreting Model Predictions
  • Anchors: High-Precision Model-Agnostic Explanations
  • Elements of Statistical Learning - Section 10.13
  • Extracting Tree-Structured Representations of Trained Networks
  • Interpretability via Model Extraction
  • Interpretable Machine Learning - Chapter 6 and Chapter 7
  • Peeking Inside the Black Box: Visualizing Statistical Learning with Plots of Individual Conditional Expectation
  • Towards Better Understanding of Gradient-based Attribution Methods for Deep Neural Networks
  • Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models
  • “Why Should I Trust You?” Explaining the Predictions of Any Classifier

• Problems with Post-hoc Explanation:

  • General Pitfalls of Model-Agnostic Interpretation Methods
  • Limitations of Interpretable Machine Learning Methods
  • When Not to Trust Your Explanations

---

Lecture 3: Bias Testing and Remediation

_{^{Source: Lecture 3 Notes}}

Lecture 3 Class Materials

• Lecture Notes
• Assignment 3
  • Model evaluation notebook
  • Full evaluations results
• Reading: Machine Learning for High-Risk Applications, Chapter 4 and Chapter 10

Lecture 3 Additional Software Tools

• Python:

  • aequitas
  • AIF360
  • Algorithmic Fairness
  • fairlearn
  • fairml
  • solas-ai-disparity
  • tensorflow/fairness-indicators
  • Themis

• R:

  • AIF360
  • fairmodels
  • fairness

Lecture 3 Additional Software Examples

• Increase Fairness in Your Machine Learning Project with Disparate Impact Analysis using Python and H2O
• Testing a Constrained Model for Discrimination and Remediating Discovered Discrimination
• Machine Learning for High-risk Applications: Use Cases (Chapter 10)

Lecture 3 Additional Reading

• Introduction and Background:

  • 50 Years of Test (Un)fairness: Lessons for Machine Learning
  • Fairness and Machine Learning - Introduction
  • NIST SP1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence
  • Fairness Through Awareness

• Discrimination Testing and Remediation Techniques:

  • An Empirical Comparison of Bias Reduction Methods on Real-World Problems in High-Stakes Policy Settings
  • Certifying and Removing Disparate Impact
  • Data Preprocessing Techniques for Classification Without Discrimination
  • Decision Theory for Discrimination-aware Classification
  • Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification Without Disparate Mistreatment
  • Learning Fair Representations
  • Mitigating Unwanted Biases with Adversarial Learning

---

Lecture 4: Machine Learning Security

_{^{Source: Responsible Machine Learning}}

Lecture 4 Class Materials

• Lecture Notes
• Assignment 4
• Reading: Machine Learning for High-Risk Applications, Chapter 5 and Chapter 11

Lecture 4 Additional Software Tools

• adversarial-robustness-toolbox
• counterfit
• cleverhans
• foolbox
• ml_privacy_meter
• NIST de-identification tools
• robustness
• tensorflow/privacy

Lecture 4 Additional Software Examples

• Attacking a Machine Learning Model
• Machine Learning for High-risk Applications: Use Cases (Chapter 11)

Lecture 4 Additional Reading

• Introduction and Background:

  • A Marauder’s Map of Security and Privacy in Machine Learning
  • BIML Interactive Machine Learning Risk Framework
  • FTC's "Start with Security" guidelines
  • Mitre Adversarial Threat Matrix
  • NIST Computer Security Resource Center
  • The Security of Machine Learning
  • Proposals for model vulnerability and security

• Machine Learning Attacks and Countermeasures:

  • Membership Inference Attacks Against Machine Learning Models
  • Stealing Machine Learning Models via Prediction APIs
  • Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures
  • Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers
  • Robust ML
  • Sponge Examples: Energy-latency Attacks on Neural Networks

• Examples of Real-world Attacks:

  • Fraudsters Cloned Company Director’s Voice In $35 Million Heist, Police Find
  • ISIS 'still evading detection on Facebook', report says
  • Researchers bypass airport and payment facial recognition systems using masks
  • Slight Street Sign Modifications Can Completely Fool Machine Learning Algorithms
  • These students figured out their tests were graded by AI — and the easy way to cheat

---

Lecture 5: Machine Learning Model Debugging

_{^{Source: Real-World Strategies for Model Debugging}}

Lecture 5 Class Materials

• Lecture Notes
• Assignment 5
• Reading: Machine Learning for High-Risk Applications, Chapter 3 and Chapter 8

Lecture 5 Additional Software Tools

• Python:

  • mlextend
  • PiML
  • SALib
  • themis-ml

• R:

  • DALEX
  • drifter

• Other:

  • manifold
  • What-If Tool

Lecture 5 Additional Software Examples

• Advanced residual analysis example
• Advanced sensitivity analysis example
• Basic sensitivity and residual analysis example
• Debugging a Machine Learning Model
• Machine Learning for High-risk Applications: Use Cases (Chapter 8)

Lecture 5 Additional Reading

• Introduction and Background:

  • AI Incident Database
  • Debugging Machine Learning Models
  • Overview of Debugging ML Models
  • Real-World Strategies for Model Debugging
  • Safe and Reliable Machine Learning
  • Why you should care about debugging machine learning models

• Debugging Approaches and Information:

  • A Comprehensive Study on Deep Learning Bug Characteristics
  • DQI: Measuring Data Quality in NLP
  • Identifying and Overcoming Common Data Mining Mistakes
  • PiML User Guide: Diagnostic Suite
  • Predicting Good Probabilities With Supervised Learning
  • Underspecification Presents Challenges for Credibility in Modern Machine Learning

---

Lecture 6: Responsible Machine Learning Best Practices

_{^{A Responsible Machine Learning Workflow Diagram. Source: Information, 11(3) (March 2020).}}

Lecture 6 Class Materials

• Lecture Notes
• Assignment 6 (Final Assessment)
• Reading: Machine Learning for High-Risk Applications, Chapter 1 and Chapter 12

Lecture 6 Additional Software Tools and Examples

• Awesome Machine Learning Interpretability

Lecture 6 Additional Reading

• Introduction and Background:
  • A Responsible Machine Learning Workflow with Focus on Interpretable Models, Post-hoc Explanation, and Discrimination Testing
  • NIST AI Risk Management Framework
  • Interagency Guidance on Model Risk Management (SR 11-7)
  • Responsible AI Practices

---

Lecture 7: Risk Mitigation Proposals for Language Models

_{^{A number of headlines and images relating to language model incidents. Source: Lecture 7 notes.}}

Lecture 7 Class Materials

• Lecture Notes

Lecture 7 Additional Tools

• BBQ Benchmark
• DAIR.AI Prompt Engineering Guide
• PAI’s Responsible Practices for Synthetic Media
• Real Toxicity Prompts
• Truthful QA
• Winogender Schemas

Lecture 7 Additional Reading

• Introduction and Background:
  • Taxonomy of Risks posed by Language Models
  • On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?