Bias Mitigating Few-Shot Class-Incremental Learning

TL;DR

This work tackles the bias that arises in Few-Shot Class-Incremental Learning (FSCIL) when base-class information dominates the feature space and classifiers during incremental sessions. It reframes FSCIL bias as a model-bias problem and introduces the SSS framework: Mapping ability stimulation to diversify feature mappings, separately dual-feature classification to preserve transferable and discriminative features, and self-optimizing classifiers including calibration with unlabeled data and BGMM-based refinements. Empirically, SSS yields state-of-the-art Overall acc, Inc. acc, and reduced Base/Inc and CInc./PInc. imbalances across miniImageNet, CIFAR100, and CUB200, with ablations confirming the contribution of each component. The approach enables more balanced, robust recognition for base and future incremental classes, advancing FSCIL toward open-world applicability.

Abstract

Few-shot class-incremental learning (FSCIL) aims at recognizing novel classes continually with limited novel class samples. A mainstream baseline for FSCIL is first to train the whole model in the base session, then freeze the feature extractor in the incremental sessions. Despite achieving high overall accuracy, most methods exhibit notably low accuracy for incremental classes. Some recent methods somewhat alleviate the accuracy imbalance between base and incremental classes by fine-tuning the feature extractor in the incremental sessions, but they further cause the accuracy imbalance between past and current incremental classes. In this paper, we study the causes of such classification accuracy imbalance for FSCIL, and abstract them into a unified model bias problem. Based on the analyses, we propose a novel method to mitigate model bias of the FSCIL problem during training and inference processes, which includes mapping ability stimulation, separately dual-feature classification, and self-optimizing classifiers. Extensive experiments on three widely-used FSCIL benchmark datasets show that our method significantly mitigates the model bias problem and achieves state-of-the-art performance.

Figures (9)

• Figure 1: Visualization of the feature space with t-SNE on miniImageNet test set. The feature extractor is trained using only the base class samples. The samples of $15$ incremental (Inc.) classes are scatteredly mapped to the base class positions.
• Figure 2: Classification accuracy imbalance on miniImageNet. There is a serious accuracy imbalance between base and incremental classes in passive ways (e.g., CECDBLP:conf/cvpr/ZhangSLZPX21, C-FSCILDBLP:conf/cvpr/HerscheKCBSR22). Although active ways (e.g. NC-FSCILDBLP:conf/iclr/YangYLLTT23, BidistDBLP:conf/cvpr/Zhao0XC0NF23) somewhat alleviate the above imbalance, they cause a new accuracy imbalance between past and current incremental classes. Our method effectively alleviates the two types of accuracy imbalance and achieves the highest overall accuracy ($\mathrm{Overall\ acc.}_{\mathrm{Avg.}}$).
• Figure 3: The component to stimulate mapping ability, including intra-class transform and inter-class fusion, to expand overall mappable space and compress the feature space occupied by base classes for future incremental classes.
• Figure 4: The components to self-optimizing classifiers, i.e., resisting base class classifiers using the incremental classes and continuously calibrating existing classifiers based on encountered samples.
• Figure 5: Comparison with SOTAs on CIFAR100 and CUB200 datasets in terms of the overall accuracy.

Theorems & Definitions (3)

• Definition 2.1
• Definition 2.2
• Definition 2.3