CarM: Hierarchical Episodic Memory for Continual Learning
TimeThursday, July 14th4:10pm - 4:30pm PDT
Location3002, Level 3
Event Type
Research Manuscript
AI/ML Design: System and Platform
DescriptionEpisodic memory(EM) is one of methods in continual learning(CL) that manage catastrophic forgetting in the presence of a stream of non-i.i.d. data to learn. However, the effectiveness of EM is highly limited by the fixed memory capacity. Prior CL methods discard samples overflowed EM and never retrieve them back for subsequent training, incurring information loss. Our system, SwapCL, addresses this problem by leveraging a large storage and swapping samples between RAM and storage, instead of discarding a majority of them like existing EM does, and dynamically make use of such abundant information to improve training accuracy and avoid catastrophic forgetting.