Seminar • Artificial Intelligence • Generalization, Memorization, and Privacy in Trustworthy Machine Learning

Please note: This seminar will take place in DC 1304.

Mahdi Haghifam, Distinguished Postdoctoral Researcher
Khoury College of Computer Sciences, Northeastern University

Machine learning is transforming numerous aspects of modern society, and its expanding use in high-stakes applications calls for responsible development. In this talk, I will present my research on the foundations and methodologies for building trustworthy ML, centered on three interconnected challenges: generalization, memorization, and privacy.

First, I will show how information-theoretic tools can be used to analyze generalization across different learning setups. Next, I will describe my work on the fundamental limits of memorization in certain high-dimensional convex settings, showing a precise trade-off between memorization and accuracy. Finally, I will propose adaptive and efficient optimization algorithms under differential privacy—a well-established framework designed to protect sensitive data and limit memorization risk—that adapt to the properties of the dataset, resulting in smaller errors.

My results highlight how these three pillars interact, and I will conclude by outlining my plans for future research.

Bio: Mahdi Haghifam is a Distinguished Postdoctoral Researcher at the Khoury College of Computer Sciences, Northeastern University, where he is hosted by Jonathan Ullman. He received his PhD from the University of Toronto and the Vector Institute, advised by Daniel M. Roy.

Mahdi’s research focuses on the foundations and algorithms of trustworthy machine learning, particularly in the areas of privacy, generalization, and memorization. During his PhD, he worked as a research intern at Google Brain and ServiceNow Research. His contributions have been recognized with a Best Paper Award at ICML 2024 and multiple research excellence awards from the University of Toronto.