主题：Continual Learning: Theory and Algorithms

主讲嘉宾：Bing Liu，Distinguished professor at the University of Illinois Chicago

时间：2021年12月16日，10:00-11:30

地点：腾讯会议592859316

主持人：陈小军，副教授

报告人简介：

Bing Liu is a distinguished professor at the University of Illinois Chicago. He received his Ph.D. in AI from University of Edinburgh. His current research interests include continual/lifelong learning, lifelong learning dialogue systems, sentiment analysis, machine learning and natural language processing. He has published extensively in prestigious conferences and journals and authored four books: one about lifelong machine learning, two about sentiment analysis, and one about Web mining. Three of his papers have received Test-of-Time awards, and another one received Test-of-Time honorable mention. He served as the Chair of ACM SIGKDD from 2013-2017 and is the winner of 2018 ACM SIGKDD Innovation Award. He is a Fellow of AAAI, ACM, and IEEE.

报告摘要：

Continual/lifelong learning learns a sequence of tasks incrementally. There are two popular settings, class incremental learning (CIL) and task incremental learning (TIL). A major challenge of continual learning is catastrophic forgetting (CF). While several techniques are available to overcome CF for TIL, CIL remains to be highly challenging due to an additional difficulty of inter-task class separation. In this talk, I will present a theoretical study on how to solve the CIL problem. The key result is that the necessary and sufficient conditions for good CIL performances are within-task prediction and task-id prediction or out-of-distribution detection. Based on this theoretical result, new CIL methods are designed, which outperform strong baselines in both CIL and TIL settings by a large margin. Finally, I will extend this study to the open world, which involves out-of-distribution samples and require learning on the fly or on the job.