Spiking-Mix: A Modified Regularization Method Designed for Spiking Neural Network

The advent of data augmentation attracts a lot of attention on how to handle unstructured data. Data augmentation makes it easier to mine the essential data to extract more useful information. And it benefits the neural network's capacity for generalization. Numerous studies contradict our perspective on data augmentation and a part of them are mature enough. However, relative studies on spiking neural networks still have the potential to grow and develop due to the properties of SNN. To achieve deeper neural networks and modify the data augmentation, the primary challenge is how to adjust data variances within a normal range. CutMix is a member of data augmentation. It adopts the idea of regional dropout and patches cropped images on original images. Owing to the rules of neurons, SNN has a single activation value. So lost information is an important issue that we must address. This shows that regional dropout with rules of neurons may sharpen the problem of lost information. So we propose a method to improve the performance of SNN through CutMix. The accuracy of SNN on CIFAR-10 is 1.39% and 0.95% higher than vanilla ANNs(with ZCA pre-processing) and ANNs(with moderate data augmentation and mean/std normalization). And the accuracy of SNN on Fashion-MNIST is 0.9% higher than ANN. The work of the paper is of great significance to developing deeper SNN.