An approximate gradient descent algorithm for Spiking Neural Network

Spiking Neural Network (SNN) is the third generation of neural networks, which transmits information through the impulse train, but the discrete impulse train has the property of non-differentiable, so it is difficult to apply the gradient descent algorithm to the spiking neural network. In this paper, the approximate derivative of the impulse activity is introduced to simulate the impulse activity, and then the spiking neural network based on the gradient descent algorithm is realized. On this basis, the influence of different approximate derivatives on the training accuracy of the spiking neural network is explored, and the iterative formula of LIF (Leaky Integrate and Fired) neurons is optimized and simplified. The results show that when the approximate derivative is introduced, our neural network has lower consumption, better performance, and the accuracy of the moment function model neural network is higher. We take the MNIST data set as the input of the spiking neural network, convert it into the impulse sequence information by the frequency coding method based on the impulse counting, and transmit it through the simplified LIF neuron model. On the basis of the error back propagation rules, the synaptic weight and error deviation of the neural network are constantly updated. The results show that the proposed algorithm is of higher accuracy and faster speed.