Research And Application Of Spiking Convolutional Neural Network Model

Spiking neural network(SNN),as the third-generation artificial neural network,transmits information by receiving and transmitting spikes.It is biologically more realistic than traditional artificial neural network and can better mimic the working principle of biological neurons.Therefore,it has attracted much attention in the field of brain-like computing.Existing spiking neural networks mainly combines unsupervised learning and spike time dependent plasticity(STDP)rules in signal processing,image classification,and speech recognition.Compared with traditional artificial neural networks,SNN has advantages in information expression and computing ability.However,the existing SNN network structure is relatively shallow,and it is difficult to ensure the superiority of the results while improving computing power.Therefore,based on the research of Spiking neuron model,STDP learning rules and sparse coding,combined with convolution,pooling and other operations in the convolutional neural network model and back propagation algorithm,this thesis proposes a sparse independent local network based spiking convolution neural network(Spiking-CNN)model and the multi-layer Spiking-CNN model based on back propagation STDP(BP-STDP)learning algorithm.The main contents of this thesis include:(1)The current SNN training mainly uses STDP learning rules and offline training conversion.The training algorithm based on STDP rules is relatively simple,and the conversion to SNN after offline training has no advantage over traditional artificial neural networks.Therefore,this thesis proposes a Spiking-CNN model based on the sparse independent local network.Firstly,the Spiking convolution kernel is trained by the sparse independent local network,then the features of input data are extracted by Spiking convolution pool operation.After that,these features are further selected by the feature discovery layer of the probability LIF spiking neuron model,and the fully connected layer finally conducts the classification.Experiments show that the proposed model has better classification results and robustness against additive noise when dealing with image classification problems than existing single-layer SNN models.(2)In order to overcome the problem that SNN with complex network structure is challenging to train,a multi-layer Spiking-CNN model based on BP-STDP algorithm is proposed.Firstly,the STDP representation learning algorithm is used to perform multi-layer spiking convolution kernel training to obtain multiple convolution kernels,and then the full connection layer training is performed by the BP-STDP algorithm after the convolution pooling operation.In order to prevent the spiking neurons from entering the "fake death" state in the full connection layer,a weight normalization method is proposed to normalize the training weights,and the classification is finally implemented in the fully connected layer by the BP-STDP algorithm.The experimental results show that the proposed multi-layer Spiking-CNN model can achieve good performance on image classification problems.Compared with existing multi-layer SNN models,the new model has better classification results and stronger robustness against additive noises.