An Overview of Local Explainability Methods and Their Evaluation in AI

Learn about various local explainability methods and their evaluation in this comprehensive lecture from UofU Data Science. Explore key concepts starting with the challenges of opting out and understanding confidence scores in AI systems. Dive into different explanation approaches including plain English explanations with free-text and chain-of-thoughts, input attribution methods using gradient-based and select-then-predict techniques, and feature interactions through effective attention. Examine concept-based explanations using TCAV, data influence through influence functions, and contrastive explanations via contrastive editing. Understand explainability as a dialog and explore the taxonomy of explanation evaluation, including simulatability. Address the scarcity of application-grounded evaluations in NLP and conclude with a discussion on building trust in AI systems. The lecture provides a thorough examination of how to make AI systems more interpretable and trustworthy through various explainability methods.