A low-cost and high-speed hardware implementation of spiking neural network

Spiking neural network (SNN) is a neuromorphic system based on the information process and store procedure of biological neurons. In this paper, a low-cost and high-speed implementation for a spiking neural network based on FPGA is proposed. The LIF (Leaky-Integrate–Fire) neuron model and tempotron supervised learning rules are used to construct the SNN which can be applied to the classification of pictures. A combined circuit instead of lookup table implementation method is proposed to realize the complex computing of kernel function in LIF neuron model. In addition, this work replaces the multiplication operations in the weights training with the arithmetic shift, which can speed up the training efficiency and reduce the consumption of computing resources. Experimental results based on Vertix-7 FPGA shows that the classification accuracy is approximately 96% and the average time for classifying a sample is 0.576us at the maximum frequency 178 MHz which achieves approximately 908,578 speedup compared with the software implementation on Matlab.