A Research Of Spiking Based Deep Network Model And Its Application

The spiking neural network(SNN)is called the third generation neural network,which is the product of the intersection of various theoretical systems such as neuroscience and computational science.It aims to simulate the information exchange,learning and storage modes of biological neurons.SNN has powerful computing storage capacity and ultra-low power consumption,and has unique advantages in timing special diagnostics extraction.Deep learning neural network has great deep feature extraction ability,and it has excellent performance in various tasks such as image processing,speech recognition and natural language processing.The spiking deep neural network is a hotspot of the current research,which not only retains the SNN biological pulse characteristics,extracts time series information,but also strengthens the deep feature extraction ability,increasing efficiency and accuracy.Inspired by the convolutional neural network in deep neural networks,this thesis analyzes its frame structure,calculation method,learning method,etc.Then designed a multi-layer spiking convolutional neural network(MSCNN).The main work includes :(1)This thesis proposes a neuron model:the sliding window statistical model.The spiking neurons highly mimic the discharge of biological neurons,and the calculation of neuronal cell membrane voltage is very complicated.By detecting the change process of neuron excitation after spike reception,the sliding window statistical model is designed to convert the calculation problem into statistical problem,and the calculation time of membrane voltage is reduced under the premise of retaining biological characteristics.Compared with the traditional neuron model,the sliding window statistical model has higher efficiency,can effectively prevent over-fitting,and increases the robustness of the model.(2)This thesis proposes a coding method:the contrast encoding method.The information transfer unit of the SNN is a spike,so all inputs need to be encoded from a value to a spike train.The current general spike coding method does not consider the characteristics of the image itself.The specific value of the pixel is not the most important.The difference between the pixel and the surrounding pixel is the key to image recognition.Contrast coding encodes the contrast between image pixels into a spike train,and replaces the edge extraction filter that is common in CNN,which not only reduces the number of layers of the network,but also reduces the computational cost.At the same time,the effect of feature extraction is great.(3)This thesis proposes a spiking deep neural network model: MSCNN,which is implemented and used for image classification.Firstly,inspired by CNN,the network structure is designed.Secondly,according to the characteristics of spiking neurons,the calculation method of convolutional layer and pooled layer feedforward network is designed.Then the convolution kernel learning algorithm is designed to ensure the convergence of convolutional layer.Finally,the result of image classification is obtained by the fully connected layer and the classifier.The sliding window statistical neuron model and contrast coding proposed in this paper are completely independent of MSCNN and can be applied to any SNN.The focus of this thesis is to design and complete the MCSNN model and apply it to image classification tasks.After theoretical analysis and experimental comparison,the effectiveness,robustness and noise immunity of MSCNN are demonstrated.