Delve into Base-Novel Confusion: Redundancy Exploration for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) aims to acquire knowledge fromnovel classes with limited samples while retaining information about baseclasses. Existing methods address catastrophic forgetting and overfitting byfreezing the feature extractor during novel-class learning. However, thesemethods usually tend to cause the confusion between base and novel classes,i.e., classifying novel-class samples into base classes. In this paper, wedelve into this phenomenon to study its cause and solution. We first interpretthe confusion as the collision between the novel-class and the base-classregion in the feature space. Then, we find the collision is caused by thelabel-irrelevant redundancies within the base-class feature and pixel space.Through qualitative and quantitative experiments, we identify this redundancyas the shortcut in the base-class training, which can be decoupled to alleviatethe collision. Based on this analysis, to alleviate the collision between baseand novel classes, we propose a method for FSCIL named Redundancy Decouplingand Integration (RDI). RDI first decouples redundancies from base-class spaceto shrink the intra-base-class feature space. Then, it integrates theredundancies as a dummy class to enlarge the inter-base-class feature space.This process effectively compresses the base-class feature space, creatingbuffer space for novel classes and alleviating the model's confusion betweenthe base and novel classes. Extensive experiments across benchmark datasets,including CIFAR-100, miniImageNet, and CUB-200-2011 demonstrate that our methodachieves state-of-the-art performance.