Hybrid rotation self-supervision and feature space normalization for class incremental learning

Abstract

Class incremental learning has made great progress in solving the problem of catastrophic forgetting through knowledge distillation method and sample playback method. However, the existing class incremental learning methods still face the problems of limited feature representation and lack of normalized feature space, which makes them perform poorly in long-term incremental tasks. To address the above problems in class incremental learning, we propose a non-exemplar based method named Hybrid Rotation with Feature Space Normalization (HRFSN). Firstly, a novel self-supervised method called Hybrid Rotation Self-supervision (HRS) is designed to overcome the problem of limited features. HRS uses random positive samples to perform rotation prediction tasks, and makes the feature extractor learn more rich feature expression ability through complex rotation prediction tasks. Secondly, to make the learned features more generalized, Feature Space Normalization (FSN) is introduced to constrain the feature value to a normal distribution, which is well matched with HRS. Experimental results on benchmark datasets such as CIFAR-100 and Tiny-Imagenet show that our approach significantly outperforms mainstream incremental learning methods and achieves comparable performance compared to the state-of-the-art methods.