Rapid Image Classification is Realized by Spiking Neural Network based on Attention Mechanism and Parallel Neurons

Haicheng Qu, Wanzhong Wang

Abstract: As a third-generation neural network, the Spiking Neural Network (SNN) has the advantages of high efficiency and low power consumption. However, in traditional SNNs, the serial nature of spiking neurons prevents them from fully utilizing their high efficiency properties. At the same time, traditional network structures are prone to the problems of vanishing/exploding gradients and low computational precision. Therefore, this paper proposed a SNNs model for image classification based on attention mechanism and parallel neurons. The model achieves improvements in three aspects: First, using parallel spiking neurons instead of traditional spiking neurons to improve computational efficiency; Second, integrating Spike-Element-Wise (SEW) residual network to effectively avoid the problems of vanishing/exploding gradients and improve program robustness; Third, introducing multidimensional attention mechanism to bring the accuracy of the SNNs close to that of traditional neural networks. This paper was tested on the CIFAR-10 and Fashion-MNIST datasets, with accuracies of 95.11% and 95.65%, respectively, showing higher testing accuracy compared to traditional SNNs.