Responsible AI: Measuring accuracy and bias in Computer Vision Systems

Pietro Perona

California Institute of Technology, USA

Abstract

Introduction to the topic of Responsible AI in the context of Computer Vision systems. Definitions of bias and fairness. Review of the literature on bias in computer vision systems. Introducing the main challenge - establishing a causal link between variables of interest and algorithmic accuracy. Discussion of whether current techniques, based on observational studies, are adequate for establishing causality. Proposal of an experimental method, based on generative image models, for for measuring bias. Conclusion and directions for future work.