Design And Application Of A Hybrid Structured Spiking Neural Network

As a powerful data processing tool,Artificial neural networks are prominent in computer vision,intelligent control and brain-inspired computing by simulating the working process of human brain.With the development of Neuroscience,Spiking neural networks with more biological explanations,which closely combine Neuroscience and Deep Learning,have attracted the attention of many scholars in recent years.Its spike signal is encoded by the event-driven mechanism,and used as the transmission signal in networks to simulate the electrical spike signal transmitted between biological neurons in human brain,this kinds of mechanism is closer to the way to biological nervous system works.Since the spike trains are represented as discrete signals,while the stimuli input into the biological nervous system are generally continuous signals.Our cognition of the brain encoding and processing stimuli is still shallow at present,so how to encode the information carried in the stimulus into discrete spike signals that can be processed by the Spiking neural network is still a challenge.Besides,Convolutional neural networks have a well performance in image recognition,but their biointerpretability is poor.Although Spiking neural networks are closer to the actual working mechanism of brain,their feature extraction ability is weak,so improving their feature extraction ability is the key to the development of Spiking neural networks.Inspired by the asymmetry in physiological mechanism of left and right hemibrain in the process of processing visual information,a hybrid Spiking neural network is proposed in this work,experiment shows that the hybrid structure of the Spiking neural networks has a great improvement in feature extraction ability.The contributions and innovations of this paper are as follows:(1)A First-Fire-Time Temporal Rate coding method is proposed.The advantages and limitations of current mainstream spike information coding methods are analyzed,a spike coding method that combines Temporal coding and Rate coding has been proposed to solve the problems of poor anti-interference in Temporal coding and low coding efficiency of Rate coding.(2)In order to solve the shortcomings of traditional Spiking neural networks in the process of feature extracting,the research and analysis of the current hybrid-structure networks,a pseudo-siamese neural network structure model based on the combination of Spiking neural networks and Convolutional neural networks is designed,which is called Pseudo-Siamese-CSNN(PS-CSNN),combining the advantages of Spiking neural networks and Convolutional neural networks.Experimental results show that the accuracy of PS-CSNN model in number image recognition task reaches 96.8%,and it performs better than other Spiking Neural Networks with similar structures in feature extraction.(3)Aiming at the limitation that the ReSuMe supervised learning algorithm can only be applied to single-layer Spiking neural networks,a guidance mechanism based on ReSu Me learning algorithm is designed in this work,which makes input,output and target output spike trains exist in each Spiking neural network layer of PS-CSNN subnetwork.The application of ReSuMe algorithm on multilayer Spiking neural network is realized.