Class-Incremental Learning for Generative Classifiers based on Class Enhancement

Abstract

Abstract Incremental learning of new classes by neural network models would lead to catastrophic forgetting of old classes. To address this problem, this paper proposes a class-incremental learning method for generative classifiers based on class enhancement. First, in order to increase the gap between old and new classes, we train an independent conditional variational auto-encoder for each class that arrives at each stage and reserve the trained weights to record the information of that class. Second, for complex natural image datasets, we incorporate a feature extractor to transform pixel replay into feature replay, making the retained information more representative. Finally, we use importance sampling and the Bayesian criterion for classification to obtain reliable classification results. The experimental results on the MNIST and CIFAR-10 datasets show that the proposed method can improve the classification accuracy of images and effectively reduce the impact of catastrophic forgetting by using batch learning for class-incremental learning. Furthermore, for the CORe50 and OpenLORIS-Object datasets, the proposed method can well adapt to the changes of the real-time environment by using online learning for continuous target recognition, showing its robustness.