A Research Of Deep Convolutional Spiking Neural Network Model

Spiking neural network(SNN)is known as the 3rd generation neural network,which mimics the mechanism that neurons in the brain transmit spikes through adaptive synaptic connections for information transformation.Compared with traditional neural networks that require high-energy-consuming graphics cards for training,spikes are sparse in time and space thus spiking neural network can be implemented by creating low-power hard-ware.Spiking neural network has been used in image recognition,obj ect detection,speech recognition and other fields.How to make spiking neural networks achieve performance comparable to that of traditional neural networks is a hot research topic.When using spik-ing neural networks,input needs to be encoded as a spike train,and the spike train output by the network needs to be decoded.Although many learning algorithms for spiking neu-ral networks have been proposed,few attention has been paid attention to the coding and decoding schemes of spiking neural networks.Inspired by traditional deep convolutional neural network,this thesis proposed the design of deep convolutional spiking neural network,including:1.Learnable convolutional encoding layer:traditional spiking encoding method in-cludes rate encoding such as Poisson encoding and time encoding such as delay encoding.Compared with the determinism of the traditional encoding method,con-volutional encoding layer proposed by this thesis is learnable and flexible.Different convolution kernels in encoding layer can be used for different datasets.Compared with traditional encoding method,the convolutional encoding layer will not suffer from performance loss when simulation time is short.2.Multi-scale encoding:Because the existing deep spiking neural network learning algorithms are limited by the dynamics of spiking neurons,they can only achieve the depth of dozens of layers at most,and can not easily go as deep as the traditional deep neural network which can have hundreds even thousands of layers.Therefore,in the exploration of spiking neural network model architecture,it may be more effective to consider widening the network than blindly deepening the network.On the basis of the above learnable convolutional encoding layer,this thesis designs multi-scale spiking encoding with different convolutional encoding layer.This can widens the network structure,extracts more abundant features from the input image and feeds them into the network.Multi-scale features can be fused through the feature fusion layer to improve the performance of the network.3.Temporal-wighted rate decoding:Rate decoding or temporal decoding are com-monly used in spiking neural networks.However,they take into account either the rate information or the temporal information of the spike train.Combining the char-acteristics of rate decoding and temporal decoding,this thesis proposes temporal-wighted rate decoding,which makes use of both the rate and temporal information in the spike train.