Improvement And Application Of Spiking Neural Networks Supervised Learning Algorithm

Spiking neural networks are new generation of neural network models inspired by biology.Different from traditional neural networks,they can effectively simulate the working mode of biological neurons.At the same time,they can integrate temporal information into the coding process,so they have stronger information processing ability.The connection between neurons is dynamic and plastic,which can adjust the connection weight adaptively and improve the fault tolerance of spiking neural networks.Spiking supervised learning algorithm with higher biological rationality can make spiking neural networks learn more and deeper decision rules more effectively.Spiking neural networks have more prominent advantages than traditional neural networks in biological interpretability,response speed and energy consumption.Because of the lack of efficient learning algorithm,they have not been popularized in practical application at present.In supervised learning,spiking neural networks can be trained using the back propagation algorithm to improve the accuracy and generalization ability of the model.Due to the discrete and nonlinear mechanism of spiking neural networks themselves,it is difficult to construct efficient supervised learning algorithms of spiking neural networks.Most of the existing algorithms are not suitable for deep or large networks.In addition,the software algorithm obtained from the research of spiking neural networks have not been widely used,which also restricts the development of spiking neural networks to a certain extent.In order to improve the efficiency of the supervised learning algorithm of spiking neural networks and further verify the advantages of the existing algorithm platform,this paper makes relevant research and analysis based on the supervised learning algorithm of spiking neural networks and theirs application.The main contents and research results are as follows:(1)In order to improve the efficiency of supervised learning algorithm of spiking neural networks,the existing classical algorithm Tempotron is optimized and innovated.Combining the advantages of mini batch gradient descent mechanism and Tempotron algorithm,an improved algorithm is proposed and named MB-Tempotron algorithm.In order to verify the efficiency of the algorithm,the MB-Tempotron algorithm is used to conduct a classification experiment on the handwritten digital dataset MNIST.The results show that,compared with the traditional Tempotron algorithm,the classification accuracy of MBTp algorithm based on mini batch gradient descent for weight adjustment is improved by 4.2%,which solves the problem of low accuracy of Tempotron in dataset classification to some extent.(2)In order to evaluate the performance of the improved MB-Tempotron algorithm in more complex scenarios,an experiment of face emotion recognition classification was carried out based on MTFL dataset.The Neu Cube architecture with high performance and low power consumption is studied and analyzed.By comparing the experimental results of Neu Cube with those of other network models,the advantages of Neu Cube architecture for face emotion recognition and classification experiments are illustrated.Compared with Neu Cube,MB-Tempotron learning algorithm has higher classification accuracy,which further shows the feasibility and effectiveness of MB-Tempotron algorithm.