Foundations of Interpretable AI

Mon 20 Oct 8 a.m. — noon (HST). Room TBD

Abstract

Interpretability has emerged as a key challenge for the widespread adoption of deep learning, especially in domains where AI decisions can profoundly affect human lives (e.g., healthcare, finance). This tutorial will provide an overview of two main approaches to interpretability along with a discussion of their respective strenghts and limitations:

Post-hoc explainability: methods that explain existing models.
Explainable-by-design: methods that build inherently interpretable models.

Schedule

08:00–08:15	Opening remarks: Need for Interpretable AI
8:15–9:00	Post-hoc explainability and explainable-by-design approachs to model interpretability Aditya Chattopadhyay Slides
9:00–9:15	Coffee Break
9:15–10:15	Shapley values based post-hoc explainability methods Jeremias Sulam Slides
10:15–10:30	Coffee Break
10:30–11:30	Information Pursuit: a framework for explainable-by-deisgn ML René Vidal Slides
11:30–noon	Q&A Session

Organizers

Aditya Chattopadhyay — AWS AI Labs
Jeremias Sulam — Johns Hopkins University
René Vidal — University of Pennsylvania