Adaptive Decision Boundary for Few-Shot Class-Incremental Learning

TL;DR

This work tackles FSCIL by introducing Adaptive Decision Boundary Strategy (ADBS), a plug-and-play module that assigns and adaptively updates a unique decision boundary per class to reserve space for incoming novel classes and reduce base–novel conflicts. It combines Adaptive Decision Boundaries with an Inter-class Constraint loss to simultaneously refine class prototypes and boundaries, improving inter-class separation. Empirical results on CIFAR100, miniImageNet, and CUB200 show consistent gains when ADBS is integrated with various FSCIL baselines, achieving state-of-the-art performance in many settings. The approach provides a practical, architecture-agnostic enhancement for FSCIL with clear theoretical motivation supported by supplementary proofs and extensive analyses.

Abstract

Few-Shot Class-Incremental Learning (FSCIL) aims to continuously learn new classes from a limited set of training samples without forgetting knowledge of previously learned classes. Conventional FSCIL methods typically build a robust feature extractor during the base training session with abundant training samples and subsequently freeze this extractor, only fine-tuning the classifier in subsequent incremental phases. However, current strategies primarily focus on preventing catastrophic forgetting, considering only the relationship between novel and base classes, without paying attention to the specific decision spaces of each class. To address this challenge, we propose a plug-and-play Adaptive Decision Boundary Strategy (ADBS), which is compatible with most FSCIL methods. Specifically, we assign a specific decision boundary to each class and adaptively adjust these boundaries during training to optimally refine the decision spaces for the classes in each session. Furthermore, to amplify the distinctiveness between classes, we employ a novel inter-class constraint loss that optimizes the decision boundaries and prototypes for each class. Extensive experiments on three benchmarks, namely CIFAR100, miniImageNet, and CUB200, demonstrate that incorporating our ADBS method with existing FSCIL techniques significantly improves performance, achieving overall state-of-the-art results.

Figures (8)

• Figure 1: Illustration of FSCIL classification with (a) Fixed and (b) Our proposed adaptive decision boundaries. A fixed decision boundary strategy often struggles to reserve adequate space for new classes at each incremental stage, resulting in space conflicts between old and new classes. In contrast, our proposed adaptive decision boundary strategy can effectively alleviate this issue by adjusting the decision boundaries of both old and new classes.
• Figure 1: Influence of hyper-parameter $alpha$ on average accuracy of our ADBS integrated with the FSCIL baseline on CUB200 dataset.
• Figure 2: The overall pipeline of our Adaptive Decision Boundary Strategy (ADBS). In the base session, we compress the decision boundaries of the base classes to reserve feature space for the upcoming new classes, as depicted in (a)-(c). Subsequently, in the incremental sessions, while maintaining the boundaries of the base classes, we dynamically adjust the boundaries of the new classes to optimize classification performance and compress the boundaries of current classes to allocate feature space for forthcoming new classes, as shown in (d)-(f). Furthermore, we impose Inter-class Constraint (IC) to enhance class distinguishability in each session.
• Figure 2: Influence of incremental shots on each session's accuracy for 5-Way FSCIL task on miniImageNet dataset.
• Figure 3: Comparison with different baseline methods on CIFAR100, CUB200, and miniImageNet. The dashed line represents the accuracy of the baseline method, while the solid line depicts the classification performance with our proposed ADBS.

Theorems & Definitions (4)

• Proposition 1
• Proof 1
• Proposition 1
• Proof 1