This talk presents information-theoretic results on three facets of trustworthy machine learning: fairness, arbitrariness, and privacy. We highlight recent findings at the interface of information theory and differential privacy, including new methods for privacy accounting and the design of privacy mechanisms. We then discuss a post-processing technique, “FairProjection,” designed to ensure group fairness in prediction and classification. We present converse results based on Blackwell’s “comparison of experiments” theorem that captures the limits of group fairness assurance in classification. These results show that existing techniques (including FairProjection) can approach the optimal Pareto frontier between accuracy and group fairness in specific settings. We conclude by reviewing the concept of predictive multiplicity in machine learning. Predictive multiplicity arises when different classifiers achieve similar average performance for a specific learning task yet produce conflicting predictions for individual samples. We present recent metrics and findings on multiplicity and their implications for applications such as content moderation.

Location: SEC LL2.221 or Zoom