[Colloquium] Xuchen Gong MS PresentationTBD

[via Colloquium]

This is an announcement of Xuchen Gong's MS Presentation
===============================================
Candidate: Xuchen Gong

Date: Monday, December 1, 2026

Time: 10:00 am to 11:00 am
Location: JCL 354
Zoom Link: https://uchicago.zoom.us/j/93809988423?pwd=lUNlE44vj2XJyqH6P9e3rYSfg85Vtr.1
                     Meeting ID: 938 0998 8423
                     Passcode: 302700

Title:  Zeroth-Order Optimization for Private and Sharpness-Aware Learning

Abstract: Zeroth-order (ZO) optimization relying on function evaluations can be an appealing alternative to first-order training when gradients are expensive to obtain or non-existent. However, existing ZO approaches typically give poor estimates of the gradients especially in high-dimensional settings, resulting in degraded model utilities. In addition, classic ZO methods optimize a smoothed version of the original objective, as opposed to the exact function.
 
In this talk, we first explore leveraging public information to guide ZO estimates, particularly in private settings where ZO is inherently suitable to privatize. Our approach reduces variance in gradient estimation and consistently outperforms private first-order baselines in multiple tasks and training scenarios. Second, observing the fact that ZO implicitly encourages flatter minima by solving the smoothed function, we propose new ZO algorithms that minimize a family of sharpness-aware objectives. We study precise characterization of sharpness and show that our methods lead to better generalization.
 
In all, this talk demonstrates that with appropriate guidance and objective design, zeroth-order optimization could be a scalable and privacy-preserving alternative to gradient-based learning in modern machine learning applications.

Advisors: Tian Li
Committee Members Aloni Cohen, Lorenzo Orecchia and Tian Li