Tomáš Kroupa: Coalitional Games for Explainable Machine Learning

Wednesday  September 14, 2022 at 10am. (room A711 and online, please ask the link to the organizers)

Tomáš Kroupa (Czech Technical University in Prague)

Title: Coalitional Games for Explainable Machine Learning

Absract:

Black-box machine learning models rarely have a built-in mechanism to explain and interpret their output. This deficiency has led to a recent interest in explainability methods for ML and AI. From the game-theoretic point of view, it is notable that the most useful interpretation methods are based on the Shapley and Banzhaf values from the coalitional games. In the talk, I will review the existing approaches to defining such games and computing their value. I will also mention the problems arising from the applications of such methods to generative models containing complex interactions among features.