Prototypes as Anchors: Tackling Unseen Noise for online continual learning

In the context of online class-incremental continual learning (CIL), adapting to label noise becomes paramount for model success in evolving domains. While some continual learning (CL) methods have begun to address noisy data streams, most assume that the noise strictly belongs to closed-set noise-i.e., they follow the assumption that noise in the current task originates classes within the same task. This assumption is clearly unrealistic in real-world scenarios. In this paper, we first formulate and analyze the concepts of closed-set and open-set noise, showing that both types can introduce unseen classes for the current training classifier. Then, to effectively handle noisy labels and unknown classes, we present an innovative replay-based method Prototypes as Anchors (PAA), which learns representative and discriminative prototypes for each class, and conducts a similarity-based denoising schema in the representation space to distinguish and eliminate the negative impact of unseen classes. By implementing a dual-classifier architecture, PAA conducts consistency checks between the classifiers to ensure robustness. Extensive experimental results on diverse datasets demonstrate a significant improvement in model performance and robustness compared to existing approaches, offering a promising avenue for continual learning in dynamic, real-world environments.