Deep Spiking Neural Network And Its Application

Artificial neural network is abstract and simulation of the human brain system.The traditional neural network represents an analog variable by spiking firing rates of biological neurons.The spiking neurons encode the information based on the exact time of the neural spiking,and integrate the time,space,phase and other information of the neural firing.Compared with the traditional neural network based on neuron firing rates of spiking,spiking neural network is closer to the brain mechanism,tackle with more information,and have a more powerful computing capability.However,the spiking neural network is still at the preliminary study stage currently,and the existing network is limited to a single layer or shallow simple network topology with single output.The restrict network structures make the spiking neural network can only use to solve some simple logic problems.Thus,the expansion of spiking neural network' topologies and building the deep spiking neural network are of great significance.This article studies the neuron models,learning rules and the network structures of the spiking neural network deeply,and explores the network models and learning methods of the deep spiking neural network.Firstly,a multilayer spiking neural network with several inputs and outputs and a hidden layer is built base on SRM model.The network uses the ReSuMe learning method which combines the back-propagation algorithm and the STDP learning rule of the spiking neurons.The network can solve linear inseparable XOR logic problem.Then,let the same simple character images as the inputs and target outputs,and train a spiking neural network with three layers in which the neural numbers of the input layer are equal to the output layer and the hidden layer is a half.The results show that the network outputs can accurately reproduce the inputs,which verifies the learning capability of the multilayer spiking neural network.But there is also a problem that the low-dimensional features of the hidden layer are indistinguishable.Secondly,a deep spiking neural network is built based on the autoencoder.The model contains multiple coding layers,and each coding layer contains an autoencoder and a spiking coding layer network.For each coding layer,study the preliminary features through the autoencoder.Then,let the input spiking features and the low-dimensional features learnt by autoencoder to be the inputs and target outputs,and train the spiking coding layer network which contains only one input layer and output layer.After that,let the outputs of the current spiking coding layer network to be the input of the next coding layer,and train the autoencoder and the spiking coding layer network.Finally,the deep spiking neural network is built by cascading the spiking coding layer networks learnt layer by layer.To analyze the effectiveness of the proposed model,select the handwritten characters datasets to conduct classification experiments,and the output features of the network are classified by SVM classifier.Experimental results show that the classification accuracy and the feature extraction capability of the deep spiking neural network are close to autoencoder.The network model can extract effective low-dimensional high-level features,has good prospects for development.Finally,considering that the time complexity of the spiking neural model is relatively large and the deep spiking neural network is difficult to process images with large sizes,an improving fast SLIC superpixel algorithm is proposed and the superpixels are used as the inputs of the deep spiking neural network to accelerate the network.The improved fast SLIC consists of global clustering algorithm and gradually refined edge algorithm.The algorithm only calculate the pixels on the boundaries of the superpixels,and the time complexity of the algorithm is reduced from ()to(√),where is the number of the image pixels, is the number of superpixels,and is far less than .Experimental results show that the improved algorithm can achieve comparable boundary recall rate and under-segment error rate,but also has a higher processing efficiency.Compared to pixels of the image after reduction,superpixels as the inputs of the deep spiking neural network achieves higher accuracy in the classification experiments.