Empowering Machine Unlearning through Model Sparsity and Beyond

Abstract: In this talk, I will introduce the concept of Machine Unlearning (MU), a technique designed to remove undesirable data influence from learned discriminative or generative models, ensuring compliance with data regulations. To bridge the gap between exact and approximate unlearning, I will present a novel approach to MU from a model-based perspective, focusing on model sparsity and weight saliency. Through theoretical analysis and practical experiments, I will demonstrate the significant improvements in multi-criteria unlearning that can be achieved by incorporating model-level characteristics, all while maintaining efficiency. Furthermore, I will highlight the practical impact of MU in tackling challenges such as AI robustness and safety.

Bio: Dr. Sijia Liu is currently an Assistant Professor at the Department of Computer Science and Engineering, Michigan State University, and an Affiliated Professor at IBM Research. His research expertise lies in the fields of machine learning, optimization, and signal processing, with a specific focus on trustworthy and scalable ML. He received the Best Paper Runner-Up Award at the Conference on Uncertainty in Artificial Intelligence (UAI) in 2022 and the Best Student Paper Award at the 42nd IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) in 2017. He also received the NSF CAREER Award in 2024.