Research On The Random Learning Method For One Kind Of Spiking Neural Networks

The spiking neural networks were applied in the fields of image recognition,machine learning and artificial intelligence.The probability for a neuron to produce a spike depends on the membrane potentials of the input neurons.The neural membrane potentials depend on the rules of random learning.Therefore,it is necessary to study spiking neural networks based on the random learning methods.In this paper,Bayesian inference,the Hidden Markov model and other statistical methods are used to study the rules of random learning in both the feedforward networks and the recurrent ones.The results show that spiking neural networks based on the random learning rule are efficient.The first chapter of our paper mainly introduces the background of our research and some preliminary knowledge about spiking neural networks and learning rules.In the second chapter,we mainly discuss the recognition of handwritten letters based on Bayesian inference.It is assumed that the spiking neural networks have the Winner-Take-All rule.By considering the accumulation of the membrane potential of feedforward networks we derive the handwritten letter category from the Bayesian formula.In the third chapter,we mainly discuss the inference of spiking neural networks based on Hidden Markov Model.Firstly,the integrate-and-fire model of a single neuron is generalized to the membrane potential accumulation model of two-level spiking neural networks.Then,the Hidden Markov Model is obtained by using Markov hypothesis,and the Hidden Markov Model and Winner-Take-All rule are applied to the spiking neural networks with two layers of neurons.Finally,the general formula of weight update is obtained by EM algorithm,and the output of the spiking neural network under hidden Markov model is obtained.In the fourth chapter,we discuss the Bayesian population decoding of spiking neural networks.Firstly,we suppose the prior distribution of stimuli is the normal distribution.Then we transform the input spike trains into neural membrane potentials.Finally,we discuss the methods of decoding stimuli in the noiseless case and noisy case respectively.In the fifth chapter,we describe the conclusion of this thesis,and point out the direction of further researches.