Colloquium: Trustworthy Machine Learning by Kush Varshney, IBM Research

Title:
Results in Trustworthy Machine Learning That Go Against Conventional Wisdom

Abstract:
Trustworthy machine learning, including topics such as fairness, explainability, robustness, and transparency, has started becoming an established field of study in recent years. Some nuggets of conventional wisdom have come to be established along with the establishment of the field. Examples include: 1) under-sampling of dark-skinned people in training datasets yields systematic disadvantages against them, 2) there is a tradeoff between fairness and accuracy, and 3) there is a tradeoff between explainability and accuracy. In this talk, I will present results (some empirical and some theoretical) that highlight the shortcomings of such broad general statements.