ECE Seminar Series: Yingdan Shi

The College of Engineering, Applied Science, and Health presents its ECE seminar series featuring Yingdan Shi, a Ph.D. candidate in computer engineering at Illinois Tech, who will present “Toward Reliable Machine Unlearning: Optimization, Evaluation, and Long-Term Stability.” This seminar will take place on Monday, October 26, from 12:45–1:45 p.m. in the Siegal Hall Auditorium (room 118).

Abstract

Machine unlearning removes the influence of specific data from a trained model without costly retraining. But is approximate unlearning reliable? This talk asks three questions: Can we find a better trade-off between forgetting quality and model utility by searching nonlinear paths in parameter space? Does low accuracy on the forgotten data mean it is truly forgotten, or can the information still be recovered through the model's uncertainty? Does forgetting still work when unlearning requests keep arriving, or does the model suffer “plasticity collapse?” Together, these answers show that reliable unlearning must be evaluated beyond a single model, a single accuracy number, and a single request.

Biography

Yingdan Shi is a Ph.D. candidate in computer engineering at Illinois Tech advised by Assistant Professor Ren Wang. Shi’s research focuses on trustworthy machine learning, with an emphasis on making machine unlearning reliable: how to forget effectively, how to verify that forgetting is genuine, and how to keep forgetting stable when unlearning requests keep arriving. The work spans machine unlearning, RL for LLM agents, and AI safety, and has been published at ICML and ECCV.