Towards Non-Exemplar Semi-Supervised Class-Incremental Learning

TL;DR

This work proposes a non-exemplar semi-supervised CIL framework with contrastive learning and semi-supervised incremental prototype classifier (Semi-IPC), enabling the model to incrementally learn from partially labeled new data while maintaining the knowledge of old classes.

Abstract

Deep neural networks perform remarkably well in close-world scenarios. However, novel classes emerged continually in real applications, making it necessary to learn incrementally. Class-incremental learning (CIL) aims to gradually recognize new classes while maintaining the discriminability of old ones. Existing CIL methods have two limitations: a heavy reliance on preserving old data for forgetting mitigation and the need for vast labeled data for knowledge adaptation. To overcome these issues, we propose a non-exemplar semi-supervised CIL framework with contrastive learning and semi-supervised incremental prototype classifier (Semi-IPC). On the one hand, contrastive learning helps the model learn rich representations, easing the trade-off between learning representations of new classes and forgetting that of old classes. On the other hand, Semi-IPC learns a prototype for each class with unsupervised regularization, enabling the model to incrementally learn from partially labeled new data while maintaining the knowledge of old classes. Experiments on benchmark datasets demonstrate the strong performance of our method: without storing any old samples and only using less than 1% of labels, Semi-IPC outperforms advanced exemplar-based methods. We hope our work offers new insights for future CIL research. The code will be made publicly available.

Figures (13)

• Figure 1: Previous state-of-the-art CIL methods store and replay part of old data to alleviate the forgetting of previous knowledge, and require fully labeled new data to acquire new knowledge at each incremental stage $t$. In this paper, we consider a more challenging and practical non-exemplar semi-supervised CIL scenario where the old data is unavailable and only a few new examples are labeled (other unlabeled new data is accessible).
• Figure 2: Illustration of our two-stage framework for Semi-CIL. The feature extractor is learned via contrastive learning and then frozen in incremental learning process. A unified prototype classifier is updated incrementally with both supervised and unsupervised loss. Our method is non-exemplar based, simple but effective.
• Figure 3: The normalized singular values of the feature matrix show descending curves of the top 12 eigenvalues. In supervised learning, they drop to 0 at $N-1$. SSL has a higher PC-ID with a slower curve decline.
• Figure 4: Contrastive learning results in smaller intra-distance on novel classes, and high feature space uniformity.
• Figure 5: Without saving old data, prototype resample generates pseudo-feature instances around old class prototypes, maintaining decision boundaries between old and new classes.