Multi-view prototype balance and temporary proxy constraint for exemplar-free class-incremental learning

Abstract

Exemplar-free class-incremental learning recognizes both old and new classes without saving old class exemplars because of storage limitations and privacy constraints. To address the forgetting of knowledge caused by the absence of old training data, we present a novel method that consists of two modules, multi-view prototype balance and temporary proxy constraints, which are based on feature retention and representation optimization. Specifically, multi-view prototype balance first extends the prototypes to maintain the general state of the class and then balances these prototypes combining knowledge distillation and prototype compensation to ensure the stability and plasticity of the model. To alleviate the feature overlap, the proposed temporary proxy constraint sets the temporary proxies to lightly compress the feature distribution during each mini-batch of training. Extensive experiments on five datasets with different settings demonstrate the superiority of our method against the state-of-the-art exemplar-free class-incremental learning methods.