An Incremental Learning Algorithm Based on Spiking Neural Network and Memory Model

Deep neural network has achieved good results in image classification, even surpassing humans. But training a neural network requires massive amounts of data and energy consumption. When add new tasks, catastrophic forgetting occurs and the neural network has to be trained from scratch. To solve this problem, this paper designs a memory model and proposes an incremental learning algorithm based on spiking neural network (SNN) and the memory model, which separate the image eigenvalue extraction from classification. The spiking neural network realizes the eigenvalue extraction. The memory model stores the eigenvalue information of the sample and realizes classification by comparing the eigenvalue with the eigenvalue information in the memory model. While keeping the original eigenvalue information of the memory model unchanged, incremental learning is realized by adding the eigenvalue information of the new object to the memory model. Experimental results show that the algorithm can flexibly add new object classifications while ensuring that the original classification performance remains unchanged.