Some Recent Advances in Incremental Learning

Professor Xiaopeng Hong
Harbin Institute of Technology, China

This talk focuses on the problem of "catastrophic forgetting" faced by deep neural networks when learning new tasks and knowledge, and aims to explore how to retain old knowledge to achieve continual knowledge accumulation. The talk will briefly introduce the mainstream incremental learning approaches, including replay-based forgetting suppression mechanisms, pre-trained model based incremental learning approaches with prompting, and new trends in the era of big models.

Biosketch

Xiaopeng Hong received his Ph.D. degree in computer application technology from Harbin Institute of Technology, P. R. China, in 2010. He is a professor at Harbin Institute of Technology (HIT), P. R. China. He had been a distinguished research fellow at Xi'an Jiaotong University, P. R. China, and an adjunct professor at the University of Oulu, Finland. Xiaopeng had been a PI of over 10 projects such as the National Key R&D Program Projects, PRC, and Infotech Oulu Postdoctoral funding project. He has authored over 80 articles in journals and conferences such as IEEE T-PAMI, CVPR, ICCV, and AAAI. His studies about subtle facial movement analysis was reported by International media like MIT Technology Review. He was the co-author of a 'top paper award' paper in ACM Multimedia 2023 and also the 2020 'IEEE Finland Section best student conference paper'. His current research interests include incremental learning, visual surveillance, and micro-expression analysis.