Rethinking Class-Incremental Learning from a Dynamic Imbalanced Learning Perspective

Leyuan Wang,Liuyu Xiang,Yunlong Wang,Huijia Wu,Zhaofeng He

2024-05-24

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.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

### Problems Addressed by the Paper

The paper primarily explores the issue of dynamic data imbalance in Class-Incremental Learning (CIL) and proposes a novel method to address this problem.

#### Core Issues
1. **Catastrophic Forgetting**: Deep neural networks tend to forget knowledge from old tasks when continuously learning new concepts.
2. **Data Imbalance**: The imbalance between old task data and new task data leads to a decline in the performance of old tasks.
3. **Dynamic Imbalance**: As new classes are added during the incremental learning process, the imbalance ratio gradually worsens, further harming the performance of CIL.

#### Solution
The paper proposes **Uniform Prototype Contrastive Learning (UPCL)**:
- **Uniform Prototype Generation**: A set of non-learnable uniformly distributed prototypes is generated before the start of each task.
- **Class-Prototype Assignment**: Classes are assigned to prototypes using a modified Hungarian algorithm.
- **Prototype Contrastive Learning with Dynamic Margin**: Dynamic margins are used to adjust the relative distances between features of new and old classes, achieving balanced and compact feature distributions.

#### Experimental Results
- On multiple benchmark datasets (such as CIFAR100, ImageNet100, and TinyImageNet), the UPCL method significantly outperforms other existing methods.
- UPCL shows consistent performance improvements under different task settings, especially excelling in experiments with more tasks and more classes.

### Summary
The paper aims to solve the problem of dynamic data imbalance in class-incremental learning. By proposing the UPCL method, it effectively retains knowledge from old tasks and improves overall classification performance.