Design And Application Of Spiking Neural Networks Based On LIF Neurons

In recent years,with the rapid development of neuroscience and technology,brainlike computing has gradually become a hot research direction in the field of artificial intelligence.Spiking neural networks,which originated from the theoretical basis of neuroscience,have become one of the mainstream frameworks in the field of brain-like computing.Spiking neural networks,as the most bionic artificial neural networks at present,contain neural nodes with temporal dynamics characteristics,use spikes as coding information,and build spiking neurons based on biological synaptic structures.Compared with traditional artificial neural networks,spiking neural networks have more biologically interpretable.At present,most researches on spiking neural networks are limited to single-layer or shallow-layer network structures,which can only be used to solve some simple logical classification problems.Although spiking neurons can effectively simulate biological neurons,their learning ability is not as good as the current deep model due to the lack of efficient learning algorithms.In practical classification tasks,traditional deep neural networks can only interpret the communication between data to solve specific tasks by imitating the spatial hierarchy of biological brains and lack generalization ability.Since traditional artificial neural networks do not consider the real simulation of biological neuron membrane potential changes and spike discharge processes,they lack biological interpretability.And leaky integrate and fire(LIF)neurons mimic biological neuronal dynamics well and are more biologically compatible.In order to make the artificial neural networks both biological rationality and high efficiency in pattern recognition tasks,this paper designs two spiking neural network structures based on LIF neurons to improve network performance and learning efficiency,and conducts experimental verification in image classification tasks.The main contributions and innovations of this article are as follows:(1)To overcome the limitations of spiking neural networks in image classification tasks such as low classification accuracy and large delay intervals.This paper studies and analyzes the advantages of unsupervised learning algorithms for spike timing dependent plasticity(STDP),proposes a method that combines STDP learning algorithm with competitive learning,and designs a two-layer spiking neural network structure.The network layer is divided into an input layer and a learning layer,and in the learning layer,a different number of learning neurons are set up for small-batch sample training.The STDP learning algorithm is used to update the weight,STDP offset is introduced to enhance the temporal causal of the spike,and MNIST image dataset is classified and validated.The results show that the classification accuracy of the network structure proposed in this paper reaches 83.179%,which is biologically reasonable and performs well in unsupervised learning classification.(2)To solve the problem that spiking neural networks are difficult to achieve deeper network applications.Based on the advantage of spiking neurons,this paper embeds a spiking mechanism in the deep model.Besides,combining the deep learning and spiking neural networks,two deep spiking neural network structures are designed,which are named the two-layer network structure FC-DSNN and convolutional FC-DSNN,respectively.Poisson coding is used to encode image pixel values into spiking trains,and the Adam optimization algorithm is selected for network training.Classification verification is performed in MNIST and Fashion MNIST image datasets.Experiments show that the performance of these two network structures in image classification tasks has improved,with accuracy rates of 97.67% and 97.09%,respectively.Moreover,the training time is shorter,and the efficiency of image classification is higher.