ML@GT Seminar Series | Understanding Last Layer Retraining Methods for Fair Classification: Theory and Algorithms

Featuring Lalitha Sankar, Arizona State University

Abstract: Last-layer retraining (LLR) methods have emerged as an efficient framework for ensuring fairness and robustness in deep models. In this talk, we present an overview of existing methods and provide theoretical guarantees for several prominent methods. Under the threat of label noise, either in the class or domain annotations, we show that these naive methods fail. To address these issues, we present a new robust LLR method in the framework of two-stage corrections and demonstrate that it achieves SOTA performance under domain label noise with minimal data overhead. Finally, we demonstrate that class label noise causes catastrophic failures even with robust two-stage methods, and propose a drop-in label correction which outperforms existing methods with very low computational and data cost.

Bio: Lalitha Sankar is a Professor in the School of Electrical, Computer and Energy Engineering at Arizona State University. She joined ASU as an assistant professor in fall of 2012, and was an associate professor from 2018-2023. She received  a bachelor's degree from the Indian Institute of Technology, Bombay, a master's degree from the University of Maryland, and a doctorate from Rutgers University in 2007.  Following her doctorate, Sankar was a recipient of a three-year Science and Technology Teaching Postdoctoral Fellowship from the Council on Science and Technology at Princeton University, following which she was an associate research scholar at Princeton. Prior to her doctoral studies, she was a senior member of technical staff at AT&T Shannon Laboratories.

Sankar's research interests are at the intersection of information and data sciences including a background in signal processing, learning theory, and control theory with applications to the design of machine learning algorithms with algorithmic fairness, privacy, and robustness guarantees. Her research also applies such methods to complex networks including the electric power grid and healthcare systems. 

For her doctoral work, she received the 2007-2008 Electrical Engineering Academic Achievement Award from Rutgers University. She received the IEEE Globecom 2011 Best Paper Award for her work on privacy of side-information in multi-user data systems. She was awarded the National Science Foundation CAREER award in 2014 for her project on privacy-guaranteed distributed interactions in critical infrastructure networks such as the Smart Grid. She has led an NSF Institute on Data-intensive Research in Science and Engineering (I-DIRSE), is a recipient of an NSF SCALE MoDL (Mathematics of Deep Learning) grant, and a Google AI for Social Good grant. Sankar was a distinguished lecturer for the IEEE Information Theory Society from 2020-2022. She serves as an Associate Editor for the IEEE Transactions on Information Forensics and Security, IEEE Information Theory Transactions, and was an AE for the IEEE BITS Magazine until August 2024.