Rethinking Class-Incremental Learning from a Dynamic Imbalanced Learning Perspective

TL;DR

This paper addresses catastrophic forgetting in class-incremental learning by analyzing data imbalance issues and proposing a novel contrastive learning method with dynamic margin adjustment to maintain balanced feature representations.

Contribution

The paper introduces Uniform Prototype Contrastive Learning (UPCL), a new approach that uses uniform prototypes and dynamic margin adjustment to mitigate data imbalance and improve continual learning performance.

Findings

Achieves state-of-the-art results on CIFAR100, ImageNet100, and TinyImageNet.

Effectively maintains balanced and compact feature distributions during incremental learning.

Addresses the increasing imbalance ratio problem in class-incremental learning.

Abstract

Deep neural networks suffer from catastrophic forgetting when continually learning new concepts. In this paper, we analyze this problem from a data imbalance point of view. We argue that the imbalance between old task and new task data contributes to forgetting of the old tasks. Moreover, the increasing imbalance ratio during incremental learning further aggravates the problem. To address the dynamic imbalance issue, we propose Uniform Prototype Contrastive Learning (UPCL), where uniform and compact features are learned. Specifically, we generate a set of non-learnable uniform prototypes before each task starts. Then we assign these uniform prototypes to each class and guide the feature learning through prototype contrastive learning. We also dynamically adjust the relative margin between old and new classes so that the feature distribution will be maintained balanced and compact. Finally, we demonstrate through extensive experiments that the proposed method achieves state-of-the-art performance on several benchmark datasets including CIFAR100, ImageNet100 and TinyImageNet.