Talk: Memorization and Privacy in Learning: Fundamental Limits and New Algorithms

Live stream: TBD

Abstract: Algorithms for machine learning and statistical inference, by design, extract relevant information from their training data. What other information do they extract, and why? In this talk, we will examine these questions. We will first see new algorithms for fundamental tasks such as mean estimation and least squares that satisfy differential privacy, the gold-standard framework for limiting privacy leakage. We will then discuss why, in some settings, strong memorization of individual training examples is necessary for learning, and what this means for privacy and beyond.

Bio: Gavin Brown is a postdoctoral scholar with Sewoong Oh at the University of Washington. He received his PhD in computer science from Boston University, where he was advised by Adam Smith. He works to understand the capabilities and limits of machine learning, with a focus on topics related to privacy and information. He received the Boston University Computer Science Department’s Research Excellence award and, at COLT 2023, the Best Student Paper award.