Spiking neural network model building method

Abstract

The invention provides a spiking neural network model building method for seeking a lighter and more efficient machine vision solution for the practical constraint problems of high resources occupancy, more complex operation and the like in an existing spiking neural network (SNN) image classification model. The conversion from a grayscale image to a pulse sequence is completed through convolution normalization and first pulse time coding, network self-classification is realized in combination with a classic pulse time dependent plasticity (STDP) algorithm and a lateral suppression algorithm, and occurrence of overfitting is effectively suppressed by introducing an adaptive threshold. Experimental results on an MNIST data set show that compared with a traditional SNN classification model, the complexity of a weight updating algorithm is reduced from O (n2) to O (1), and the image recognition accuracy can still be stably kept at about 96%. The provided model is beneficial to the bottom layer transplantation of a software algorithm to a hardware platform, and can provide a reference for the implementation of an edge calculation scheme of a high-efficiency low-power-consumption small intelligent hardware terminal.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a method for constructing a spiking neural network model. Background technique [0002] In recent years, the Spike-Time-Dependent-Plasticity (Spike-Time-Dependent-Plasticity, STDP) algorithm has gradually become one of the mainstream learning algorithms of the SNN model (Spike Neural Network) with its profound physiological foundation and efficient regulatory mechanism, and has been successfully applied On hardware terminals such as Field Programmable Logic Gate Arrays and ASICs. In the existing related fields, there are input feature extraction through sorting coding method and simplified STDP algorithm, and then the output classification is completed with the help of Support Vector Machine (SVM), and finally obtained on the Face / Moto subset of the Caltech dataset. The accuracy rate of 99.1% is achieved; in the prior art, on the basis of the existing STDP algorithm, t...

AI Technical Summary

Problems solved by technology

[0004] The present invention aims at the practical constraints such as high resource occupation and complex operation existing in the existing spiking neural network (SNN) image classification model. In order to seek a more lightweight and efficient machine vision classification solution, this application proposes a spiking neural network The model construction method, the provided model is conducive to the transplantation of software algorithms to the bottom layer of the hardware platform, and can provide a reference for the realization of edge computing solutions for small intelligent hardware terminals with high efficiency and low power consumption

Images