Research On Image Recognition Based On Evolutionary Spiking Neural Networks

As the 3th generation artificial neural networks,spiking neural networks use exact spiking timing coding mode for information processing,and the process of transmitting and information processing through spiking neurons has higher biological interpretability.The evolutionary spiking neural networks combine the advantages of spiking neural networks and evolutionary algorithms.It has stronger power of computing and better flexibility,especially suitable for the processing of complex spatio-spectro temporal data.Based on the model of genetic regulatory networks,the development and generating procedures of spiking neural networks are more in line with the laws of nature.Image recognition is a grave problem that should be focus on and urgent solved in the field of intelligent computing.The key technology of image recognition is to construct a computing model,which has powerful computing ability and can be applied widely.Firstly,we studied on a kind of genetic regulatory networks In this paper,and then we used it to construct spiking neural networks.The model has three layers' structure includes input layer,control layer and output layer.The nodes between input layer and control layer are connect by forward and full connection mode,same as the nodes between control layer and output layer.While the nodes in control layer connect to each other by recurrent mode.The nodes connections of recurrent genetic regulatory networks model represented by the weighted matrix,which has good interaction.In the model,the activity of output nodes controls cell division,generating spiking neurons and generating connections,finally generating the recurrent spiking neural networks.Secondly,the analysis of scale-free structure and small world structure carried out to research the generation method of the network.Scale-free structure and small-world structure are the most important two features in the analysis of networks structure.Through a series of experiments,we analyzed the influence of the number of control nodes and the size of weights on the structure of developmental spiking neural network.In addition,the influence of developmental scale on the structure was also be analyzed.The experimental results show that,the number of nodes,the size ofweights and the developmental scale have a great influence on the formation of scale-free structure and small-world structure.Finally,we used the evolutionary networks model for solving practical pattern recognition problem: image recognition.The spiking neural networks transfer and process information by spiking trains,so need encode the digital images into spiking trains first,in this procedure,the latency-phase encoding was employed.Then,the basic computational model of image recognition obtained through evolving synaptic weights.At the last,the spiking neural networks were trained once again by supervised learning algorithm to improve the computing ability further.The last experiment results has proved that the model has grave pattern recognition ability.