Overview
Learn about gradient-based input attribution methods in machine learning through this 37-minute lecture that explores key evaluation metrics and techniques. Begin with a foundational recap of gradient-based highlighting approaches, then dive into plausibility assessment of attribution methods. Examine comprehensive evaluation frameworks including sufficiency metrics, and conclude with advanced concepts like erasure techniques, Remove and Retrain (ROAR), and Recursive-ROAR methodologies for robust model interpretation.
Syllabus
Recap: Gradient-based highlights/input attribution so-far
Plausibility
Comprehensiveness & Sufficiency
Erasure, ROAR, Recursive-ROAR
Taught by
UofU Data Science