Representative-exploring Replay for Online Class-Incremental Continual Learning

Libang Zhao, Alysa Ziying Tan, Siwei Feng, Han Yu, Tao Deng, Yuanlu Chen, Feng Wu, Mengyuan Song

Abstract: Online Class-Incremental Continual Learning (CICL) methods often rely on data replay, where a fixed-size memory buffer stores a small subset of previous data to mitigate catastrophic forgetting. However, as incremental tasks progress, the fixed buffer size reduces the number of exemplars stored per class, weakening their representativeness. We argue that preserving classification centers is more effective than maintaining decision boundaries with limited exemplars. This highlights the dual challenges of selecting high-quality exemplars for storage and addressing the imbalance between old and new classes. To tackle these issues, we propose the Representative Exploration Replay (RER) framework. Our approach evaluates exemplar representativeness using a novel metric based on the model’s classification performance, ensuring that more representative exemplars are prioritized for storage. Additionally, it mitigates class imbalance through mutual information gradient masking and knowledge distillation. Comprehensive experiments on three datasets demonstrate that RER