Research And Application Of Spiking Neural Network Supervised Learning Algorithm

Artificial neural network is a hot research field in the vigorous development of artificial intelligence science.Spike neural network is an emerging artificial neural network model based on pulse timing coding.This method of expressing information with precise pulse time can explain the information processing mechanism of the biological brain more accurately than traditional frequency coding,and it is easy to implement by hardware.But because the spiking neural network has distinct characteristics in terms of information coding,neuron model,network structure,etc.,it is impossible to directly use traditional neural network algorithms for model training.At present,due to the lack of comprehensive and efficient learning algorithms,the powerful spatiotemporal information processing capabilities of the Spike Neural Network have not been effectively used for the time being,and it cannot be widely promoted and applied.In this paper,through the research and analysis of typical supervised learning algorithms of spiking neural networks,especially the in-depth exploration of the learning mechanism of PBSNLR algorithm based on membrane voltage drive,the more efficient,accurate and robust supervised learning mechanism of spiking neural networks explore.The main research contents and contributions of the thesis are as follows:1)A Multi-parameter PBSNLR algorithm with multi-parameter control weight adjustment is proposed.The PBSNLR algorithm itself is an algorithm based on membrane voltage driving,so by introducing the distance between membrane voltage and threshold as the dynamic parameter of the weight adjustment rule,the only problem caused by the weight adjustment range during the algorithm training can be effectively changed.Experiments have proved that the Multi-parameter PBSNLR algorithm can accurately learn the target pulse signal,and it has a higher learning efficiency compared to the traditional PBSNLR algorithm.2)Propose a more robust and accurate R-Multi-parameter PBSNLR algorithm.Aiming at the problem that negative samples near the target ignition time in the PBSNLR algorithm may be misclassified and the problem that the threshold pull-down in the Multi-parameter PBSNLR algorithm may cause insufficient membrane voltage accumulation,the time division idea in the noise-threshold strategy is used for reference to the noise-threshold strategy.Threshold strategy is improved,and a dynamic threshold strategy that is still applicable in the case of fine step size is proposed,and it is combined with the Multi-parameter PBSNLR algorithm to propose the R-Multiparameter PBSNLR algorithm.Experiments show that the R-Multi-parameter PBSNLR algorithm has higher learning efficiency and accuracy.3)By applying the algorithm to the classic optical character image recognition task and disease data prediction task,the performance of multiple spiking neural network supervised learning algorithms in real application scenarios is compared and analyzed,and the performance of R-Multi-parameter PBSNLR is verified.In the application scenario,it can still maintain faster learning efficiency and better accuracy,and at the same time have a certain anti-noise performance.It is a highly practical supervised learning algorithm.