Fractional Gradient Descent Method for Spiking Neural Networks

A fractional-order gradient descent method applicable to spiking neural networks(SNNs) is proposed for the problem of difficulty in training SNNs using stochastic gradient descent method. The method is an improvement of the location of the gradient backpropagation calculation in the training of SNNs and the output form of the last layer in the structure, respectively, and the convergence is proved using the fractional-order difference expression under the Grünwald-Letnikov definition. The test results on both pure SNN and CNN2SNN models simultaneously show that the method can effectively reduce the number of sample cycle inputs and thus the training time without reducing the accuracy.