eXplainable Machine Learning / Wyjaśnialne Uczenie Maszynowe - 2024

eXplainable Machine Learning course for Machine Learning (MSc) studies at the University of Warsaw.

Winter semester 2023/24 @pbiecek @hbaniecki

Previous year: https://github.com/mim-uw/eXplainableMachineLearning-2023

Meetings

Plan for the winter semester 2023/2024. MIM_UW classes are on Fridays.

• 2023-10-06 -- Introduction, Slides, Audio. Extra reading: Lipton 2017, Rudin 2019
• 2023-10-13 -- Fairness, Slides, Audio, Extra reading: Fairness and ML, Cirillo 2020
• 2023-10-20 -- LIME and friends, Slides, Audio+Video, Extra reading: Why Should I Trust You?, LORE
• 2023-10-27 -- SHAP and friends, Slides, Audio, Extra reading: General Pitfalls of Model-Agnostic Interpretation Methods for Machine Learning Models, Shapley Flow: A Graph-based Approach to Interpreting Model Predictions
• 2023-10-30 -- PROJECT: First checkpoint
• 2023-11-10 -- PDP and friends, Slides, Audio, Extra reading: AlePlot, Peeking Inside the Black Box
• 2023-11-17 -- VIP / MCR Slides, Audio, Extra reading: The Two Cultures, More models, Rashomon quartet
• 2023-11-24 -- Explanations specific to neural networks & Evaluation of explanations
• 2023-12-01 -- Guest lecture
• 2023-12-08 -- PROJECT: Second checkpoint
• 2023-12-15 -- Adversarial XAI
• 2024-01-12 -- Student presentations
• 2024-01-19 -- Guest lecture
• 2024-01-26 -- PROJECT: Final presentation

How to get a good grade

The final grade is based on activity in four areas:

• mandatory: Project (0-35)
• mandatory: Exam (0-35)
• optional: Homeworks (0-24)
• optional: Presentation (0-10)

In total you can get from 0 to 100 points. 51 points are needed to pass this course.

Grades:

• 51-60: (3) dst
• 61-70: (3.5) dst+
• 71-80: (4) db
• 81-90: (4.5) db+
• 91-100: (5) bdb

Homeworks (0-24 points)

• Homework 1 for 0-4 points. Deadline: 2023-10-12 - graded by HBA
• Homework 2 for 0-4 points. Deadline: 2023-10-19 - graded by PBI
• Homework 3 for 0-4 points. Deadline: 2023-10-26 - graded by PBI
• Homework 4 for 0-4 points. Deadline: 2023-11-09 - graded by HBA
• Homework 5 for 0-4 points. Deadline: 2023-11-16 - graded by HBA
• Homework 6 for 0-4 points. Deadline: 2023-11-23 - graded by PBI

Project (0-35 points)

This year's project involves conducting a vulnerability analysis of a predictive models using XAI tools. This analysis should be carried out for a selected model and the results should be summarised in a short RedTeaming report.

Key points:

• Projects can be done in groups of 1, 2 or 3 students
• One model can be analysed by multiple groups (but the discovered vulnerabilities must not be repeated)
• The harder the project, the easier it is to obtain a higher grade.

Important dates:

• 2023-10-30 – First checkpoint: Students chose the model, create a plan of work (to be discussed at the classes). Deliverables: 3 min presentation based on one slide. (0-5 points)
• 2023-12-08 – Second checkpoint: Provide initial experimental results. At least one vulnerability should have been found by now. (0-10 points)
• 2023-01-26 - Final checkpoint: Presentation of all identified vulnerabilities. (0-20 points)

Models:

RedTeaming analysis should be carried out for a selected model. Depending on the difficulty of the model, you may receive more or less points

• the analysis concerns your own model (e.g. from homework). Point conversion rate: x0.8
• the analysis applies to a model from Hugging Face (one of the models available there for any modality). Points conversion factor: x1
• the analysis applies to one of the popular foundation models (e.g. TabPFN https://arxiv.org/abs/2207.01848, Segment Anything https://arxiv.org/abs/2304.02643, Llama 2 https://arxiv.org/abs/2307.09288, RETFound https://www.nature.com/articles/s41586-023-06555-x). Points conversion factor: x1.25

RedTeam Report:

Examples of directions to look for vulnerability (creativity will be appreciated)

• bias / fairness. Does the model discriminate against a protected attribute? (e.g. https://arxiv.org/abs/2105.02317)
• using XAI to find instance level artifacts (e.g. Clever Hans https://doi.org/10.1016/j.inffus.2021.07.015, wolf/snow https://arxiv.org/abs/1602.04938), large residuals and explanations why predictions were wrong
• unintended memorisation (e.g. https://arxiv.org/abs/1802.08232)
• drift in performance gap between datasets (e.g. another image or text dataset)
• model malfunction due to an unintendent use
• wrong behaviour on out-of-distribution samples
• SATML 2024 competition concerning CNN Interpretability based on a NeurIPS 2023 paper Red Teaming Deep Neural Networks with Feature Synthesis Tools

The final report will be a short (up to 4 pages) paper in the JMLR template. See an example.

Literature

We recommend to dive deep into the following books and explore their references on a particular topic of interest:

• Explanatory Model Analysis. Explore, Explain and Examine Predictive Models by Przemysław Biecek, Tomasz Burzykowski
• Fairness and Machine Learning: Limitations and Opportunities by Solon Barocas, Moritz Hardt, Arvind Narayanan
• Interpretable Machine Learning. A Guide for Making Black Box Models Explainable by Christoph Molnar