Synthesizing Pareto-optimal Interpretations of Black Box ML Models

This lecture by Supratik Chakraborty from IIT Bombay explores the challenge of synthesizing explainable interpretations of black-box machine learning models. Learn about the inherent tension between explainability metrics (such as decision tree depth and node count) and interpretation accuracy. Discover how Pareto-optimal interpretations can be systematically synthesized using MaxSAT solving techniques, providing PAC-style guarantees on the results. The presentation demonstrates that optimizing for multiple Pareto-optimal interpretations reveals valuable solutions that might be missed when using a single scalar objective function combining accuracy and explainability. The talk presents joint research with Hazem Torfah, Shetal Shah, Sanjit Seshia, and S. Akshay as part of the Theoretical Aspects of Trustworthy AI series at the Simons Institute.