Beyond discriminative features: Invariant Representation Learning for Few-Shot Class-Incremental Learning

Document Type

Article

Publication Date

12-2026

Department

Department of Computer Science

Abstract

Few-shot Class-Incremental Learning (FSCIL) enables models to continuously learn new classes with a small number of samples while preventing catastrophic forgetting of old classes. Current methods focus solely on the essential features of the category, making them too sensitive to changes in non-discriminative features, such as background and illumination, which severely limits generalization during the incremental stage. To address this, we propose the Invariant Representation Learning (IRL) framework, which focuses on learning robust representations against non-discriminative disturbances. The framework is primarily driven by a novel Invariant Representation Enhancer (IRE) and maintained by a Semantic Feature Alignment (SFA) strategy. In the base session, IRE separates visual features into discriminative and non-discriminative components through structural feature decoupling. It then generates perturbation samples via feature recombination and uses contrastive learning to guide the model in eliminating irrelevant feature changes, constructing a robust representation invariant to interference factors. In the incremental phase, SFA maintains the stability of old class features across sessions by constraining cross-session distribution consistency, alleviating semantic drift. Extensive experiments on three benchmark datasets — CUB200, CIFAR100, and mini-ImageNet — show that our method outperforms existing state-of-the-art techniques. The code is available at https://github.com/JichengYuan81/IRL.

Publication Title

Pattern Recognition

Link to Full Text

Share

COinS