Construction Of Spiking Neuron Networks Based On Spike Timing-dependent Plasticity Mechanism Fusion Back-propagation Algorithm

Artificial Neural Network can store and process information by drawing on the structure and working mode of biological brain,which greatly improves the level of Artificial Intelligence technology.A new Spiking Neuron Network compared with the traditional neural network,with a richer biological characteristics and more powerful computing ability,and bring hope for strong artificial intelligence to realizing the simulation,extension and expansion of human intelligence.The error Back-Propagation(BP)algorithm is widely used,because it has the advantages of high fault tolerance,learning effectiveness and strong computing ability.But because of Spiking Neuron Network transmission and processing the spike sequence signal,there are discontinuous and nonlinear characteristics;in the process of BP algorithm derivation,the requirement of neuron model must have analytical expressions and other reasons,the existing Spiking Neuron Network BP algorithm has many problems,such as: multi spike sequence learning difficulty,only use specific neuron and network structure,no biologically interpretable,the variety of biological characteristics and application is difficult to reproduce.To solve the above problems,this paper first uses the Spike Timing-Dependent Plasticity(STDP)mechanism to improve the existing Spiking Neuron Network BP algorithm.The main research work is as follows:(1)Using STDP mechanism to improve BP algorithm.The first based on error function of neuron membrane potential measuring and BP algorithm obtain the weight variation measure;then according to the chain rule,the weights of the variables in the formula of membrane potential weight calculation for the decomposition of the time derivative of the membrane potential and the weights of the time derivative of the reciprocal of the product;and then STDP function will as the derivation of time by weight,get universal and multi spiking sequence learning algorithm BP.BP algorithm with universality and multi spiking sequence learning ability is deduced.(2)Single neuron learning simulation experiment of improved BP algorithm.Take input noise and change of DC excitation two kinds of interference neurons actual output frequency respectively;research and analysis when the no mechanism,improved BP algorithm and STDP algorithm to adjust the weights of each neuron,neuron output frequency changes relative to the target output.The results show that the output spike frequency of single neuron is close to the target output frequency after the improved BP algorithm adjusts the weights,and the accuracy of the input noise is increased by 77.4% compared with the STDP mechanism.(3)Neural network learning simulation experiment based on improved BP algorithm.The construction of three layer connection form of feed-forward neural network model in noise and different intensity of DC excitation of two cases,respectively,research and analysis the discharge of every layer when no algorithm,improved BP algorithm and STDP mechanism adjustment network weight.The results show that the mutual correlation index of each layer of the network controlled by the improved BP algorithm achieves the best value,and presents a stable synchronous discharge.In this paper,the Spiking Neuron Network constructed by STDP fusion BP,expresses the biological characteristics of synaptic plasticity,resonance and synchronous discharge,which have not in the previous BP algorithm training network,and provides a new way to solve the problem of Artificial Intelligence to control system stable operation and accurate classification from the biological point of view.