Alleviating Regional Shortcuts for Few-Shot Class-Incremental Learning

2026-07-24

AI Summary

This work addresses the problem in few-shot class-incremental learning (FSCIL) where models misclassify novel-class samples due to overreliance on discriminative regions learned from base classes. The study is the first to identify and formally name this issue as “regional shortcut.” Through a compositional learning perspective, the authors analyze how models spatially reuse features for novel classes and propose a dual-primitive set modeling mechanism that separately learns generic and discriminative primitives to disentangle representations. Integrating spatial pattern analysis with theoretical validation, the proposed method achieves significant performance gains over current state-of-the-art approaches on standard FSCIL benchmarks, while simultaneously improving both classification accuracy and model interpretability.

Abstract

Few-shot class-incremental learning (FSCIL) aims to incrementally learn novel classes with only a few samples while avoiding forgetting base classes. However, current methods show a tendency to misclassify novel-class samples into base classes, which we find to be caused by the excessive focus on base-class-discriminative regions on novel-class samples. In this work, we aim to explore the underlying mechanism for an interpretation and solution. We first provide a compositional view to analyze the transferred and reused spatial patterns on novel-class samples. Then, through extensive experiments and theoretical analysis, we identify both empirically and theoretically that a shortcut exists in the model's base-class training, which naturally forms the excessive focus on only the most discriminative regions (primitives), which we term as the regional shortcut. Finally, based on this interpretation, to address this problem, we propose a compositional-learning-based method to learn two primitive sets (a common set and a discriminative set), which alleviates the regional shortcut by constraining the model to learn and utilize the common primitive set for base- and novel-class recognition. Extensive experiments on standard FSCIL benchmarks demonstrate the effectiveness of our approach, yielding consistent improvements over existing state-of-the-art methods in both accuracy and interpretability.

Problem

• few-shot class-incremental learning
• regional shortcut
• catastrophic forgetting
• discriminative regions
• class confusion

Innovation

• regional shortcut
• compositional learning
• few-shot class-incremental learning
• primitive sets
• catastrophic forgetting