Research On Unsupervised And Supervised Learning Algorithms Of Spiking Neural Network

In recent years,with the continuous deepening of artificial neural network research,it has reached a new height.The third-generation artificial neural network centered on Spiking Neural Networks(SNN)has become a research hotspot in the field of artificial neural networks due to its highly bionic characteristics.Based on Spiking Neural Networks,its unsupervised and supervised learning algorithms are studied in this paper.The details are as follows:Firstly,in view of the problem that the fully connected network structure is easy to overfit after training,combined with the mechanism of ipsilateral inhibition in biology and the idea of partially inactivating neurons,a network structure based on lateral inhibition and probabilistic connection is constructed.Simulation experiments show that this structure can effectively reduce the network’s Degree of overfitting.Aiming at the problem that the frequency coding is too random and the time coding information is too concentrated and easy to be interfered,a time-frequency coding method is proposed,and the effectiveness of the method is verified by simulation experiments.Secondly,in order to solve the problem of low learning efficiency and instability of the traditional unsupervised learning algorithm based on pulse time-dependent plasticity(STDP),combined with the pulse nearest matching mode and the steady-state plasticity in the network,an improved unsupervised learning algorithm based on STDP rules is designed.The simulation experiment results show that the proposed unsupervised learning algorithm has a faster learning rate and higher stability.Thirdly,in view of the slow convergence speed of the Multi＿Re Su Me supervised learning algorithm,combined with the different characteristics of synaptic delay in the biological nervous system,a Multi＿Re Su Me learning algorithm with delay parameters is proposed.At the same time,in view of the problem that the non-Hebbian term in the algorithm is fixed and the algorithm is unstable,an adaptive non-Hebbian term Multi＿Re Su Me learning algorithm is proposed.Combining the advantages of the two algorithms,an improved Multi＿Re Su Me learning algorithm is proposed.Simulation experiments show this The algorithm can effectively accelerate the network learning rate and increase the stability of training.Finally,on the network structure based on lateral inhibition and probabilistic connection,systematic simulation experiments are carried out on unsupervised and supervised learning algorithms.The recognition accuracy of the network is tested on the MNIST data set and Fashion-MNIST data set,and the design is verified The effectiveness of the network and learning algorithm.