Foundations of Reliable and Lawful Machine Learning: Approaches Using Information Theory
- Sanghamitra Dutta (University of Maryland)
Abstract
How do we ensure that the machine learning algorithms in high-stakes applications, such as hiring, lending, admissions, etc., are fair, explainable, and lawful? Towards addressing this urgent question, this talk will provide some strategies that are deep-rooted in information theory, causality, and statistics. I will discuss a question that bridges the fields of fairness, explainability, and law: how do we check if the disparity in a model is purely due to critical occupational necessities or not? We propose a systematic measure of the legally non-exempt disparity, that brings together a body of work in information theory called Partial Information Decomposition and connects it with causality. I will also briefly talk about some of our other research interests in related topics, such as quantifying accuracy-fairness tradeoffs using Chernoff Information, and also robust counterfactual explanations with theoretical guarantees.