Revisiting statistical principles for risk minimization

Abstract

Minimizing a loss over a set of data points is ubiquitous in data science. A complete understanding of which loss to minimize, however, is lacking. This talk focuses on an extremely simple model: estimating an unknown mean corrupted by Gaussian noise. An intuitive guess is that minimizing square loss (equivalently, maximizing the likelihood) over the parameter space would be the best estimator possible. Counterintuitively, this is suboptimal in general. Embarrassingly, the reasons for why this occurs are not fully understood! I will explain some reasons why this can happen, what can be done about it, and some related results around this theme.

Based on several papers, joint with Liviu Aoloritei (UC Berkeley), Cong Ma (U. Chicago), Michael I. Jordan (UC Berkeley/INRIA), Annie Ulichney (UC Berkeley), and Nikita Zhivotovskiy (UC Berkeley).

Bio

Reese Pathak is an NSF Mathematical Sciences Postdoctoral Research Fellow (MSPRF) in the Department of Statistics at UC Berkeley, with sponsoring scientist Nikita Zhivotovskiy and an Assistant Professor at Cornell University, in the School of Operations Research and Information Engineering (ORIE).

He obtained his PhD in 2025 from UC Berkeley, in the Department of Electrical Engineering and Computer Sciences, with Mike Jordan and Martin Wainwright. His Ph.D. thesis was awarded the IMS Lawrence D. Brown Award. He was previously an undergraduate at Stanford.