Proxy-anchor and EVT-driven continual learning method for Generalized Category Discovery

Abstract

Continual generalized category discovery aims to continuously learn novel categories in incoming data while avoiding catastrophic forgetting of previously learned ones. Its key component is the model’s ability to separate novel samples, where Extreme Value Theory (EVT) has been effectively employed. We propose a method that integrates EVT with proxy anchors to define boundaries using a probability of inclusion function, enabling rejection of unknown samples. Additionally, we introduce a novel EVT-based loss function to enhance representations, achieving superior performance compared to other deep-metric learning methods. The derived probability functions help separate novel categories from known categories. However, category discovery within novel samples can overestimate the number of new categories. To address this, we propose an EVT-based model reduction approach to discard redundant proxies. We also incorporate experience replay and knowledge distillation during continual learning to prevent catastrophic forgetting. Experimental results show our proposed approach outperforms state-of-the-art methods.