A Data Management Approach to Explainable AI

Abstract

In recent years, there has been a growing interest in developing methods to explain individual predictions made by machine learning models. This has led to the development of various notions of explanation and scores to justify a model’s classification. However, instead of struggling with the increasing number of such notions, one can turn to an old tradition in databases and develop a declarative query language for interpretability tasks, which would allow users to specify and test their own explainability queries. Not surprisingly, logic is a suitable declarative language for this task, as it has a well-understood syntax and semantics, and there are many tools available to study its expressiveness and the complexity of the query evaluation problem. In this talk, we will discuss some recent work on developing such a logic for model interpretability.