BiaMix Contrastive Learning and Memory Similarity Distillation in Class‐Incremental Learning

ABSTRACT Class‐incremental learning studies the problem of continually learning new classes from data streams. But networks suffer from catastrophic forgetting problems, forgetting past knowledge when acquiring new knowledge. Among different approaches, replay methods have shown exceptional promise for this challenge. But performance still baffles from two aspects: (i) data in imbalanced distribution and (ii) networks with semantic inconsistency. First, due to limited memory buffer, there exists imbalance between old and new classes. Direct optimisation would lead feature space skewed towards new classes, resulting in performance degradation on old classes. Second, existing methods normally leverage previous network to regularise the present network. However, the previous network is not trained on new classes, which means that these two networks are semantic inconsistent, leading to misleading guidance information. To address these two problems, we propose BCSD (BiaMix contrastive learning and memory similarity distillation). For imbalanced distribution, we design Biased MixUp, where mixed samples are in high weight from old classes and low weight from new classes. Thus, network learns to push decision boundaries towards new classes. We further leverage label information to construct contrastive learning in order to ensure discriminability. Meanwhile, for semantic inconsistency, we distill knowledge from the previous network by capturing the similarity of new classes in current tasks to old classes from the memory buffer and transfer that knowledge to the present network. Empirical results on various datasets demonstrate its effectiveness and efficiency.