| The processing of information by humans mainly depends on billions of neurons forming complex neural networks,and the information is transmitted in the form of pulses.At present,Spike neural networks are a good combination of the characteristics of human biological nervous system and machine learning.The neural network has achieved some results in practical applications such as image processing,reinforcement learning,and speech processing,but at this stage it is only at an initial research stage,and many theoretical framework algorithms are not clear enough,so more and more people are startingStudy its theory and application.In the study of impulsive neural networks,unsupervised learning has always been an important part.Based on the STDP(Spike-Timing-Dependent Plasticity)rule,impulsive neural networks have clear biological meaning and can handle relatively complex image classification.Problem,in order to improve the efficiency and accuracy of the algorithm,this paper has made improvements to its network structure and classification layer algorithm.The main contents of this article are as follows:Firstly,the biological significance and basic theoretical knowledge of the impulsive neural network are described.On the basis of its biological significance,the conventional impulsive neuron model is introduced,and the coding mechanism and learning algorithm of the impulsive neuron are described.Network structure.Secondly,the synaptic plasticity rules are introduced in detail,the development process and different STDP rule variants are studied,and then the commonly used fully connected network structure based on STDP rules is improved,and a sparse probability connection method is proposed.Experimental verification shows that the structure can effectively reduce the network training time.Then a pulse neural network structure based on Poisson coding rules is constructed,and dynamic bias is added to the Poisson coding mechanism,and the traditional classification algorithm is also improved.A voting competition mechanism is proposed.The neuron performance categories compete for voting,optimizing the performance of network organizations with the same number of neurons on image classification.Finally,the above research is applied and verified on the MNIST data set,which shows that the network structure based on sparse probability connection can indeed effectively reduce the training time of the network.Secondly,further experiments and research on the competitive learning mechanism are conducted,indicating that the competition The number of times will be very good when it is four times,and the accuracy rate of verification on the MNIST data set has reached 98.1%,which is an average increase in accuracy compared to the pulse neural network of the same network size without the voting competition mechanism.About 6%.Moreover,when the number of neurons is small,the training time is not increased and better results are obtained.Further experiments are carried out on the Fashion-MNIST dataset to verify that the algorithm proposed in this paper has achieved certain results. |
