Research On Image Recognition Algorithm Of SNN Based On STDP

The brain has amazing computing power and its computing consumption is very low.Inspired by the brain,people have constructed shallow artificial neural network and deep artificial neural network successively.These networks also show excellent performance in pattern recognition.However,although these neural networks' inspiration is from the network of the brain,the nature of these neural networks is far from the actual brain model.In information representation,shallow or deep artificial neural networks use numerical information,while real biological neural networks use spikes(action potential);In neuron model,shallow or deep artificial neural networks use a time independent computing model,while real brain networks use the spiking neuron model,which has a strong dependence on time;In learning algorithm,the basic algorithm of shallow or deep artificial neural network is the back propagation algorithm,while the learning algorithm of the brain has a great relationship with an unsupervised learning mechanism STDP.Because of above differences,the power and computing power of the shallow or deep artificial neural networks are far inferior to those of the brain.So more and more researchers are being interested in the computational nature of the brain.To comprehend the running nature of the brain,people need to pay attention to two points,the first is biological plausibility,the second is a good performance for pattern recognition tasks.Based on this consideration,this paper gives a spiking neural network for image recognition.It not only has strong plausibility in neuron model,synaptic model and learning mechanism,but also shows good performance in image recognition.Unlike the supervised system,the learning algorithm is based on STDP,which is a biological learning mechanism.STDP is derived from neuroscience experiment.It is an unsupervised learning mechanism and has great potential of development.Compared with the existing unsupervised and competitive spiking neural network with strong biological plausibility,the thesis introduce a more accurate method of lateral inhibition,thereby simplifying the structure of competitive spiking neural network with strong biological plausibility,and accomplishing the training and recognition of image just relying on one layer of spiking neurons;having reduced every input image's presenting time,and then improving the running efficiency;having put forward a new recognition method,this method groups competing neurons by category,combining the spike number with receptive field weight matrix linearly for each group of neuron in order to generating the reconstructed image of this group of neurons,and then compared the reconstructed images of different groups with the input image by similarity to determine the category of the input image,using this method the recognition accuracy of the network has been improved.This paper first introduces the background and importance of the research of spiking neural network image recognition algorithm based on STDP;Then introduces several neuron models,including the traditional artificial neuron model and the spiking neuron model;next,introduces two main types of neurons information encoding—rate encoding and pulse encoding;following it introduces a biological learning mechanism—STDP learning mechanism,including pair STDP rule and triplet STDP rule;Then introduces the competitive network structure,and gives the design of a competitive spiking neural network,and introduces the application of the network in the handwritten digit recognition and its performance test;finally summarizes the whole paper,points out the shortcomings and the direction of the following efforts.