Analysis Of Neural Information Feed-forward Propagation Mechanism Based On Spiking Neural Network

Spiking neural network(SNN)simulate neural spiking activity and use neural spiking time or spiking frequency to process information,it performs great on tasks such as classification,regression by training.The neural information is more explicit in spiking neural network compared to that in normal feed-forward network.SNN of feed-forward structure could be mapped with neural information feed-forward propagation path and used as neural information conduction model.In this paper,we construct a spiking neural network based on rank order coding and train the network to classify image by spikeprop algorithm.This paper analyzes the effects of neural network scale,neural noise on efficiency and robustness of neural information transmission.Firstly,this paper analyzes the influence of the depth,width and neuron connection probability of spiking neural network on the transmission efficiency of neural information.Results show that deeper SNN can extract information in higher dimension and improve the accuracy of classification tasks.The width of network barely influence the neural information transmission that related to the image classification task.Reducing the network connection probability slightly is beneficial to the information conduction in network and improve the accuracy of image classification task.Secondly,simulation results demonstrate that neural noise tends to force the neuron to spike in advance,and the repetitiveness of neuron spike decreases with the increasing of noise intensity,which impairs the information conduction.Simulation experiments prove that spiking neural network of rank order coding classify images correctly under background noise and synaptic noise of certain intensify,which verifies the robustness of rank coding mechanism.The simulation experiment also proves that adding noise during training can improve the generalization of the network.Finally,this paper constructs a spiking neural network with multi-compartment neurons.By adding dendritic compartments to neuron model,it provide feedback path for error signal to update neural weights during the training of SNN.Spikeprop algorithm of spiking neural network with multi-compartment neuron is improved.Results show that the SNN has a higher recognition accuracy by adding dendritic structure,and achieve same performance in a smaller scale.This paper not only analyzes the mechanism of neural information transmission,but also provides a theoretical basis for the new model with more physiological characteristics in the field of artificial intelligence.