Research And Application Of Robustness Based On Spiking Neural Network

Spiking neural network?SNN?as the third generation artificial neural network?ANN?is the latest product of neural network research.Compared with the first generation and the second generation neural network,SNN is more in line with the biological neural network working mechanism,which in the expression of information and computing power has a huge upgrade.In the first two generations of neural networks,the average spike rate is used to encode the information.The spiking neural network adds time information to the coding process and adds the exact time of the spike to the information coding and expression.Compared with other neural networks this time coding method makes SNN in the information processing,neuronal model and synaptic learning rules have a greater difference.Therefor the study of SNN's internal mechanism and related learning algorithms has a high theoretical and practical significance.SNN's unique advantages make more and more people to concern about it.In the field of biological and artificial intelligence SNN has fruitful research results,and applied in many areas.SNN working mechanism is not difficult to understand,but the SNN learning algorithm determines the performance of the network.In practical applications,the network has to cope with a wide variety of environments and conditions,and performance efficiency under complex conditions is the key for SNN applications.In order to improve the efficiency and robustness of SNN learning algorithm,the main work of this paper is as follows:1.Based on the study of the Spiking neuron model and the working mechanism of SNN,a simpler linear dynamic threshold functionl_thris proposed by analyzing the way of noise interfers Spiking neurons.After training withl_thr,the membrane potential is much lower than the firing threshold at the non-target output time,and can ensure that the membrane potential is strong enough nearby the desired output time.Thus,the proposedl_thrgreatly improve the robustness of SNN.2.Traditional synaptic delay-based learning algorithms,such as DL-ReSuMe,can only increase the delay in the learning process,but not according to the actual optimization needs to speed up the pulse transmission speed.In this paper,we propose a new learning algorithm based on synaptic delay,EDL-ReSuMe.By studying the rules of neurological synaptic adjustment,in the learning process of EDL-ReSuMe,the delay of synapses can be dynamically adjusted?increase or decrease?to improve the learning performance.Compared with DL-ReSuMe,the proposed method has the advantages of high efficiency and high accuracy.3.In order to improve both the robustness and learning efficiency of SNN learning algorithm,this paper integratesl_thrand EDL-ReSuMe algorithm to form an new method called LEDL-ReSuMe.The algorithm can not only makes the trained SNN more robust,but also improve the learning efficiency and accuracy.4.Finally,we verify the validity of the proposed method in both the simulation data and the practical applications,respectively.The experimental results show that the proposed method has stronger robustness,higher accuracy and better learning efficiency.