Auditing and Debugging Deep Learning Models via Flip Points: Individual-Level and Group-Level Analysis

Abstract Deep learning models have been criticized for their lack of easy interpretation, which undermines confidence in use important applications. Nevertheless, they are consistently utilized many applications, consequential to humans’ lives, usually because better performance. Therefore, there is a great need computational methods that can explain, audit, and debug such models. Here, we flip points accomplish these goals deep classifiers used social A trained classifier mathematical function maps inputs classes. By way training, the partitions its domain assigns class each partitions. Partitions defined by decision boundaries expected be geometrically complex. This complexity what makes powerful classifiers. Flip on those and, therefore, key understanding changing functional behavior We advanced numerical optimization techniques state-of-the-art linear algebra, as rank determination reduced-order compute analyze them. The resulting insight into model clearly explain model’s output individual level, via an explanation report understandable non-experts. also develop procedure understand audit towards groups people. show examining certain subspaces reveal hidden biases not easily detectable. synthetic data alter improve behaviors. demonstrate our investigating several standard datasets applications machine learning. identify features most responsible particular classifications misclassifications. Finally, discuss implications auditing public policy domain.