Professor Hima Lakkaraju discusses the many future research directions for building explainable AI including better algorithms for post hoc explanations, theoretical analysis of interpretable models and explanation methods, and empirical evaluation of the utility of model explanations.

View the full playlist: https://www.youtube.com/playlist?list=PLoROMvodv4rPh6wa6PGcHH6vMG9sEIPxL

#machinelearning

0:00 Introduction
1:34 Challenges with LIME: Consistency
2:02 Challenges with LIME: Scalability
2:50 Explanations with Guarantees: BayesLIME and BayesSHAP
7:48 Theoretical Analysis of the Behavior of Explanations/Models
11:09 Characterizing Similarities and Differences
12:42 Model Understanding Beyond Classification
14:47 Intersections with Model Robustness
16:34 Intersections with Model Fairness
18:47 Intersections with Differential Privacy
23:13 New Interfaces, Tools, Benchmarks for Model Understanding