Keynote Speakers

Zhiping Lin
Nanyang Technological University (NTU), Singapore

Zhiping Lin received B.Eng degree in control and automation from South China Institute of Technology, Guangzhou, China, and Ph.D. degree in information engineering from the University of Cambridge, UK, in 1987. From 1987 to 1999, he worked at Calgary University, Canada, Shantou University, China, and DSO National Laboratories, Singapore. Since 1999, he has been an associate professor at the School of EEE, Nanyang Technological University (NTU), Singapore, where he is now serving as Programme Director of Master of Sciences in Signal Processing and Machine Learning, the School of EEE, NTU. Dr. Lin was the Editor-in-Chief of Multidimensional Systems and Signal Processing from 2011 to 2015, and a subject editor of the Journal of the Franklin Institute from 2015 to 2019. He also served as an associate editor for several other international journals, including IEEE Trans. on Circuits and Systems II. He was the coauthor of the 2007 Young Author Best Paper Award from the IEEE Signal Processing Society, and a Distinguished Lecturer of the IEEE Circuits and Systems (CAS) Society during 2007-2008. He served as Chair of IEEE CAS Singapore Chapter during 2007-2008, and 2019. His research interests include multidimensional systems, statistical signal processing and image/video processing, robotics, and machine learning. He has published more than 200 journal papers and over 200 conference papers.

Title: Analytic continual learning with a fast and non-forgetting closed-form solution

Abstract: Continual learning allows models to continuously obtain knowledge in different phases. This learning agenda resembles the human learning process and contributes to the accumulation of machine intelligence. However, conventional continual learning methods suffer from the issues of 1) catastrophic forgetting, where models rapidly lose previously learned knowledge when acquiring new information, and 2) privacy invasion, where storing data for replaying during continual learning is not allowed in some scenarios. In this talk, we introduce a new branch of continual learning, namely analytic continual learning. We first introduce an analytic class-incremental learning (ACIL) algorithm that maintains absolute memorization of past knowledge without privacy invasion. The ACIL extends the recursive least-squares to the class-incremental learning and achieves equivalent results as jointly retraining. Subsequent, we introduce the implementation in the few-shot scenario with Gaussian kernel embedded analytic learning (GKEAL) to address the data-insufficiency in few-shot settings. Finally, we introduce the dual-stream analytic learning (DS-AL), which improves the fitting ability of analytic continual learning.