Learning Semantic Anchors for Continual Generalized Category Discovery

Yuxin Fan, Junbiao Cui, Changhao Liu, Xingwang Zhao, Jiye Liang

Keywords: Continual Learning, Generalized Category Discovery

Abstract: Continual Generalized Category Discovery (C-GCD) aims to address the dual challenges of continual learning and generalized category discovery in open environments. This task requires the model to incrementally recognize new classes while resisting catastrophic forgetting of old classes. Existing C-GCD methods do not effectively balance the stability-plasticity dilemma, primarily due to ineffective semantic utilization and knowledge preservation, leading to poor recognition of new classes and catastrophic forgetting of old classes. To address these issues, we propose a novel C-GCD method that leverages textual descriptions and visual features to construct and optimize semantic anchors, and freeze image and text encoders to preserve general pre-trained knowledge. Extensive experiments on several datasets demonstrate that our method significantly outperforms existing C-GCD methods, effectively balancing the stability-plasticity dilemma to achieve enhanced new classes recognition and mitigated forgetting of old classes. Code is provided in the supplementary materials.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 3373