Enhancing Few-Shot Class Incremental Learning with Local Expert-Guided Class Separation

Continuous learning is a hallmark of biological intelligence, enabling lifelong adaptation. Inspired by this capability, few-shot class incremental learning (FSCIL) seeks to endow models with a similar facility: the ability to continually learn new classes from limited data while retaining previously acquired knowledge. However, existing methods face significant challenges, such as catastrophic forgetting and overfitting to new, limited-data tasks. In this work, we adopt the common paradigm of finetuning the model during the base session and employing a prototypical classifier. Our approach focuses on enhancing the inter-class separation in the feature space. Strong separation is critical as it creates a more distinct and spacious representation for accommodating new classes in subsequent incremental sessions, thereby mitigating catastrophic forgetting. We propose a two-stage, local expert-guided learning framework to progressively enhance inter-class separation. In the first stage, we apply prefix-tuning to help the model acquire domain-specific knowledge and achieve an initial class separation. In the second stage, we introduce local experts, each focusing on a specific region in the feature space. To obtain discriminative expert features that complement the original ones, we insert additional prompts and propose two prompt fusion modules based on the intermediate patch-level features from the backbone. Furthermore, to effectively train the experts, we propose the Selective Anchor-based Supervised Contrastive loss, Prior-Guided Class Separation loss, and Matching Score Reverse Prediction loss. Experimental results on three datasets demonstrate that our method outperforms existing state-of-the-art approaches.