Steering Prototype with Prompt-tuning for Rehearsal-free Continual Learning

Zhuowei LiLong ZhaoZizhao ZhangHan ZhangDi LiuTing LiuDimitris N. Metaxas

Zhuowei LiLong ZhaoZizhao Zhang ...+3 Dimitris N. Metaxas

Mar 2023

摘要

Prototype, as a representation of class embeddings, has been explored toreduce memory footprint or mitigate forgetting for continual learningscenarios. However, prototype-based methods still suffer from abruptperformance deterioration due to semantic drift and prototype interference. Inthis study, we propose Contrastive Prototypical Prompt (CPP) and show thattask-specific prompt-tuning, when optimized over a contrastive learningobjective, can effectively address both obstacles and significantly improve thepotency of prototypes. Our experiments demonstrate that CPP excels in fourchallenging class-incremental learning benchmarks, resulting in 4% to 6%absolute improvements over state-of-the-art methods. Moreover, CPP does notrequire a rehearsal buffer and it largely bridges the performance gap betweencontinual learning and offline joint-learning, showcasing a promising designscheme for continual learning systems under a Transformer architecture.

展开全部

图表提取