Non-exemplar Class-incremental Learning via Dual Augmentation and Dual Distillation
Ke Song, Quan Xia, Zhaoyong Qiu · 2023
Non-exemplar class incremental learning is the challenge of recognizing new and old classes without storing samples of old classes. The incremental learner is usually faced with how to strike a balance between transferability and stability. On the one hand, if the model overfits the current task data, the performance will collapse when encountering and solving unseen tasks. On the other hand, if the model fails to maintain the knowledge already learned, the feature space will be covered by the new features during the incremental learning process, which leads to catastrophic forgetting. In this paper, we propose a new DADD method to balance transferability and stability better. Firstly, we propose a new method named Rotation class augmentation (Rot-classAug) to provide additional classes to avoid overfitting at the current stage and to improve transferability. Secondly, we propose a new Noise-semantic augmentation (N-semanAug) that contains rich old-class information, which significantly alleviates catastrophic forgetting in continuous learning. Experiments on benchmark datasets show that our method has achieved superior performance, outperforming the state-of-the-art methods by a margin of 5%, 3% and 4%, respectively. Code is available at https://github.com/Ke1Song/DADD