MixER: Mixup-based Experience Replay for Online Class-Incremental Learning

Abstract

Continual learning in the online class-incremental setting aims to learn new classes continuously from a consistent data stream while retaining the knowledge of old classes to prevent catastrophic forgetting. Traditional replay-based methods store and use old-class data to achieve this. However, they often overlook the representation shift caused by the incoming data streams, which leads to suboptimal classification accuracy. In this study, we propose a solution for mitigating representation shifts by incorporating asymmetric mixup training into the replay method. Our approach is based on the concept that mixup-based training enhances the stability of model predictions and gradient norms between training samples. Our method differs from typical mixup augmentation, which is uniformly applied to all data. Instead, it selectively targets the old data stored in the memory buffer, deliberately excluding the classes from the newly incoming data. This approach enables the model to learn new data while preserving the representation of the old data. Moreover, our experiments demonstrate the effectiveness of the proposed method, which not only enhances the performance of replay-based methods but can also be seamlessly integrated as an additional compatible module into various replay-based techniques. Evaluation on the CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets demonstrates that our approach surpasses existing replay-based methods. It addresses the limitations of conventional replay techniques and offers a potential solution for continual learning scenarios. Our source code is publicly available at https://github.com/laymond1/MixER. Authors