Knowledge Editing in Multi-modal Foundation Models by Dr. Kaixiong Zhou

Abstract: While FMs have demonstrated remarkable capabilities in storing knowledge and performing reasoning across modalities, including language, vision, and structured data. their stored knowledge is often static and prone to rapid obsolescence due to an evolving world. Moreover, their reliance on imperfect internal representations can lead to hallucinations, a problem that has already had a measurable negative impact. Recently, model editing has emerged as a powerful direction for precisely refining specific knowledge facts in foundation models without full retraining. In this talk, we will introduce the model editing techniques applied on visual, language and graph deep learning models, via directly modify model parameters to adjust specific factual associations or maintain an external repository of updated knowledge that is dynamically injected during inference. 

Bio: Kaixiong Zhou is an Assistant Professor in the Department of Electrical and Computer Engineering at North Carolina State University. Prior to joining NCSU, he was a postdoctoral researcher in the Institute for Medical Engineering and Science at the Massachusetts Institute of Technology. He received his Ph.D. degree in Computer Science from Rice University in 2023. He received the B.S. degree from Sun Yat-Sen University (SYSU) in 2015 and received the M.E. degree from University of Science and Technology of China (USTC) in 2018.

His research interests lie in developing efficient, trustworthy, and use-inspired machine learning algorithms in the fields of graph representation learning, language models, and AI for science. His research contributes to tackling challenges in various applications including synthetic biology, drug discovery, and network analysis. 

All are welcome!