Proxy-Based Classification Boundary Alignment for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) challenges intelligent models to continually learn new classes from limited samples while maintaining performance on previous tasks. Mixup-based methods address this scenario by creating virtual samples, enabling the model to learn explicit or implicit virtual classification boundaries and improve its forward compatibility, thereby reserving space for new classes. However, existing methods typically focus on virtual class construction and lack a crucial comparison of explicit and implicit virtual classification boundaries, hindering the further exploitation of the potential of mixup-based strategies for the FSCIL task. To this end, this paper analyzes virtual classification boundaries and reveals that implicit virtual classification boundaries are more conducive to model forward compatibility but inferior in base model convergence compared with explicit ones. Subsequently, we construct implicit virtual classification boundaries based on inter-class correlations to improve their adaptability to new classes, and propose a novel proxy-based classification boundary alignment method to ensure the stability of real classification boundaries. This method enhances clustering by focusing on sample-to-proxy correlations and improves the model’s ability to discriminate among old and new classes in incremental sessions by integrating projection layers. Moreover, we propose to randomly mix the augmented labels generated by self-supervised learning (SSL) in a linear manner, which increases the virtual classification boundaries to incorporate new classes better. Experimental results on three benchmark datasets, CIFAR100, miniImageNet, and CUB200, demonstrate that our proposed method effectively enhances the model’s forward compatibility while maintaining stability and outperforms the state-of-the-art works.