Strategic Base Representation Learning via Feature Augmentations for Few-Shot Class Incremental Learning

TL;DR

This work tackles FSCIL by preventing forgetting while enabling seamless inclusion of new classes from few examples. It introduces Feature Augmentation driven Contrastive Learning (FACL), which expands the base feature space using feature augmentation, generates proxy classes to reserve space for future classes, and applies a MoCo-based self-supervised contrastive loss to strengthen base-class separation. The approach relies on three core components: feature augmentation to create proxy-like features, proxy-class generation to enlarge the label space, and self-supervised contrastive learning integrated with traditional cross-entropy loss. Empirical results on CIFAR100, miniImageNet, and CUB200 show state-of-the-art performance with notable improvements, validating the method’s ability to prepare the base session for robust incremental learning in realistic data-scarce settings.

Abstract

Few-shot class incremental learning implies the model to learn new classes while retaining knowledge of previously learned classes with a small number of training instances. Existing frameworks typically freeze the parameters of the previously learned classes during the incorporation of new classes. However, this approach often results in suboptimal class separation of previously learned classes, leading to overlap between old and new classes. Consequently, the performance of old classes degrades on new classes. To address these challenges, we propose a novel feature augmentation driven contrastive learning framework designed to enhance the separation of previously learned classes to accommodate new classes. Our approach involves augmenting feature vectors and assigning proxy labels to these vectors. This strategy expands the feature space, ensuring seamless integration of new classes within the expanded space. Additionally, we employ a self-supervised contrastive loss to improve the separation between previous classes. We validate our framework through experiments on three FSCIL benchmark datasets: CIFAR100, miniImageNet, and CUB200. The results demonstrate that our Feature Augmentation driven Contrastive Learning framework significantly outperforms other approaches, achieving state-of-the-art performance.

Figures (5)

• Figure 1: The illustration of the motivation for our framework. Our FACL framework is designed to more effectively learn the base session features and facilitate future class adaptation compared to existing approaches.
• Figure 2: The illustration of the proposed FACL framework. FACL consists of three main components: (a) feature augmentation, where the augmented feature vectors are used to perform the feature augmentation, (b) proxy label space, where proxy labels are created for the augmented feature vectors, and (c) self-supervised contrastive loss, where contrastive loss is applied to the different transformed versions $\mathcal{F}_a$ and $\mathcal{F}_b$, which are created using $\mathcal{F}_{\text{comb}}$ as defined in Eq. \ref{['eq:combined_features']} and built using the original image feature vectors and the augmented feature vectors. (d) Integrated Feature Classifier, where at inference time, the original label logits and the proxy label logits are integrated to perform the evaluation.
• Figure 3: Comparison of CE and FACL Confusion Matrices on the miniImageNet.
• Figure 4: Comparison of t-SNE plots between the existing framework and our proposed FACL model on the miniImageNet dataset (best view in color).
• Figure 5: Accuracy values obtained during the final session for various ($\delta$) values on the CIFAR100 dataset.