Events Calendar

CS Colloquium: Sanghamitra Dutta (Carnegie Mellon University) - Reliable Machine Learning for High-Stakes Applications: Approaches Using Information Theory

Speaker: Sanghamitra Dutta , Carnegie Mellon University

Talk Title: Reliable Machine Learning for High-Stakes Applications: Approaches Using Information Theory

Series: CS Colloquium

Abstract: How do we make machine learning (ML) algorithms not only ethical, but also intelligible, explainable, and reliable? This is particularly important today as ML enters high-stakes applications such as hiring and education, often adversely affecting people's lives with respect to gender, race, etc. Identifying bias/disparity in a model's decision is often insufficient. We really need to dig deeper and bring in an understanding of anti-discrimination laws. For instance, Title VII of the US Civil Rights Act includes a subtle and important aspect that has implications for the ML models being used today: Disparities in hiring that can be explained by a business necessity are exempt. E.g., disparity arising due to code-writing skills may be deemed exempt for a software engineering job, but the disparity due to an aptitude test may not be (e.g. Griggs v. Duke Power '71). This leads us to a question that bridges the fields of fairness, explainability, and law: How can we identify and explain the sources of disparity in ML models, e.g., did the disparity arise due to the critical business necessities or not? In this talk, I propose the first systematic measure of "non-exempt disparity," i.e., the illegal bias which cannot be explained by business necessities. To arrive at a measure for the non-exempt disparity, I adopt a rigorous axiomatic approach that brings together concepts in information theory, in particular, an emerging body of work called Partial Information Decomposition, with causal inference tools. This quantification allows one to audit a firm's hiring practices, to check if they are compliant with the law. This may also allow one to correct the disparity by better explaining the source of the bias, also providing insights into accuracy-bias tradeoffs.

My research bridges reliability in learning with reliability in computing, which has led to an emerging interdisciplinary area called "coded computing". Towards the end of this talk, I will also provide an overview of some of my results on coded reliable computing that addresses long-standing computational challenges in large-scale distributed machine learning (namely, stragglers, faults, failures) using tools from coding theory, optimization, and queueing.

This lecture satisfies requirements for CSCI 591: Research Colloquium


Biography: Sanghamitra Dutta (B. Tech. IIT Kharagpur) is a Ph.D. candidate at Carnegie Mellon University, USA. Her research interests revolve around machine learning, information theory, and statistics. She is currently focused on addressing the emerging reliability issues in machine learning concerning fairness, explainability, and law with recent publications at AAAI'20, ICML'20 (also featured in New Scientist and CMU Engineering News). In her prior work, she has also examined problems in reliable computing, proposing novel algorithmic solutions for large-scale distributed machine learning in the presence of faults and failures, using tools from coding theory (an emerging area called "coded computing"). Her results on coded computing address problems that have been open for several decades and have received substantial attention from across communities (published at IEEE Transactions on Information Theory'19,'20, NeurIPS'16, AISTATS'18, IEEE BigData'18, ICML Workshop Spotlight'19, ISIT'17,'18, Proceedings of IEEE'20 along with two pending patents). She is a recipient of the 2020 Cylab Presidential Fellowship, 2019 K&L Gates Presidential Fellowship, 2019 Axel Berny Presidential Graduate Fellowship, 2017 Tan Endowed Graduate Fellowship, 2016 Prabhu and Poonam Goel Graduate Fellowship, the 2015 Best Undergraduate Project Award at IIT Kharagpur, and the 2014 HONDA Young Engineer and Scientist Award. She has also pursued research internships at IBM Research and Dataminr.

Host: Bistra Dilkina

Audiences: By invitation only.

Contact: Assistant to CS chair