Local Explanations for Deep Learning Models - Fall 2023

This course studies local explanations for deep learning models and ways to evaluate them. It covers attribution, feature importance, data influence, contrastive explanations, application-grounded evaluation, and human trust.

Syllabus

Course overview and logistics
Data Preparation & Transformer
Pretrain-Finetune; Uncertainty Estimation (Part 1)
Uncertainty (Part 2); Prompting; Chain-of-Thoughts
Prompting (Part 2); Evaluation of Free-Text Explanations
Gradient-based Input Attribution
Evaluation of Input Attribution
Select-then-Predict
Pairwise Feature Importance (Effective Attention)
Data Influence
Contrastive Explanations
Application-Grounded Evaluations of Explanations
Human Trust in AI
Challenges in Fostering (Dis)Trust in AI
Explainability as a Dialog
An Overview of Local Explainability Methods & Their Evaluation