Taming distribution shifts in statistical learning

University Of Chicago, Chicago IL

Investigators

Abstract

In recent years, machine learning has achieved remarkable success in various benchmark tasks, including image classification, question answering, and speech recognition. However, these models tend to exhibit reduced performance when applied to test data that differs from the data they were trained on. This scenario, also known as distribution shift, is a common issue in fields such as healthcare, criminal justice, and robotics. For example, a classification model trained on medical images from one hospital may not perform well on images collected from another hospital due to differences in medical equipment, scanning protocols, and subject populations. This project aims to address the challenges of distribution shift in learning from data by developing principled methods and theories. Additionally, it seeks to train the next generation of data scientists who can handle learning under distribution shifts. The project will focus on three important problems involving distribution shifts: nonparametric regression under covariate shifts, off-policy evaluation, and offline reinforcement learning. To advance our understanding of these challenging problems, the project will (1) investigate more appropriate assumptions on distribution shifts that allow for efficient knowledge transfer between different distributions while also being practical; (2) characterize the information-theoretic limitations of learning under distribution shifts, determining when learning is possible and when it is not; and (3) develop efficient and adaptive learning methods with optimal finite-sample guarantees that adjust to the unknown amount of distribution shift. Throughout this project, new tools will be developed in high-dimensional statistics, information theory, and sequential decision making, and numerous opportunities will be provided to enhance interdisciplinary graduate and undergraduate research training. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.