Research And Applications Based On Brain-inspired Spiking Neural Network

In recent years,with the rapid development of artificial intelligence,the models of deep neural network have become complex,and the demands for computing and storage have become higher.The present situation of development has caused huge difficulties for the implementation on the hardware platforms with limited resources.Inspired by the neuromorphic computing,the spiking neural network(SNN)can achieve higher computational efficiency with lower power consumption because it simulates the dynamic discharge characteristics of biological neurons and adopts the event-driven processing mode.However,due to the lack of efficient learning algorithms and network structures,SNN has low performance in practical application.In addition,unreasonable coding method will also affect the efficiency and stability of SNN learning.Aiming at the above two problems,this paper mainly studies the spiking encoding methods and brain-inspired learning algorithm of SNN.Firstly,this paper studies the spiking encoding method of SNN.Because of the randomness of the frequency coding method,the accuracy of the information carried by a single spike is low.Therefore,this paper focuses on the research and improvement of the temporal coding methods.Because the Time-to-First-spike coding method is too sparse,a new time encoding method--differential delay coding method is designed based on the difference of image pixels in this paper.Firstly,the Time-to-First-spike coding method converts each pixel to the precise timing of spikes.Secondly,the Euclidean distance is used to calculate the difference between the central pixel and other pixel points in the 8neighborhood,which is integrated into the first-spike time of each pixel point as the delay time.Then the integrated time is mapped to the spike sequence of the central pixel one by one.Differential delay coding method can improve the stability of the network by turning pixel information into precise spike train.Secondly,in terms of the learning algorithm,a SNN based on the Spike-Timing Dependent Plasticity(STDP)learning rules is constructed according to the learning principle and connection structure of neural receptive field in the visual pathway.Since STDP is a local unsupervised learning algorithm,the learning effect is poor when it is used alone.So,this paper introduces a lateral inhibition mechanism to form a competitive relationship between neurons in the learning layer.In this paper,two methods are used to construct the inhibition structure of this paper.One is based on the WTA algorithm whose inhibition strength remains unchanged,and the other is based on the SOM algorithm whose inhibition strength increases with distance.Through this lateral inhibition structure,the synaptic weights in the learning layer can be updated according to the Triplet-STDP algorithm.Finally,the proposed algorithms are tested and analyzed in detail on standard MNIST dataset.The experimental results prove that the proposed algorithms improve the learning effect and the classification performance of the network.