Brain-inspired Class Incremental Learning

Wei Wang, Zhiying Zhang, Jielong Guo

Abstract: Most existing class incremental learning methods employ data storage or extended network structures, but they cannot effectively alleviate the catastrophic forgetting problem due to memory resource limitation. To solve this issue, a brain-inspired generative replay model is proposed in this work. First, VAE-ACGAN is used to simulate the memory self-organizing system to improve the quality of the generated pseudo-samples. Then, a shared parameter model and a private parameter model are used to protect the extracted features. Experimental results on MNIST, Permuted MNIST and CIFAR-10 show that the proposed method can lead to significantly higher classification accuracies than other class incremental learning methods. Furthermore, the backward transfer and the forward transfer of positive values demonstrate that the model can achieve a trade-off between task stability and plasticity, effectively preventing catastrophic forgetting effectively.