Class Incremental Learning (CIL) enables models to learn new classes in images while retaining knowledge of previously learned classes. However, conventional CIL assumes abundant data for new classes, which is often unrealistic in real-world applications such as autonomous driving and medical diagnosis. Few-Shot Class Incremental Learning (FSCIL) addresses this issue by enabling the learning of new classes with minimal data. Nevertheless, it faces challenges like catastrophic forgetting, feature drift, and class imbalance, which can hinder effective adaptation. To tackle these challenges, we propose a method that integrates a feature-decoupled network architecture with a Parameter-Adaptive Freezing (PAF) training strategy, aiming to improve feature stability and adaptability to new classes. Inspired by Equiangular Tight Frame (ETF) theory, we decouple the model into three components: a Feature Extractor, a Feature Distributor, and an ETF Classifier. This structure integrates adaptive virtual class space allocation to ensure optimal embeddings for new classes. Moreover, PAF selectively freezes critical parameters based on the Fisher Information Matrix, which helps minimize forgetting of old classes while simultaneously improving generalization for new classes. Extensive experiments across CIFAR-100, miniImageNet, and CUB-200-2011 demonstrate that our proposed method outperforms state-of-the-art FSCIL methods in both classification accuracy and stability.
Yao et al. (Fri,) studied this question.