Research And Application Of Deep Spiking Neural Network Model

Deep learning and spiking neural network are hot topics in the field of artificial intelligence.Deep learning has demonstrated its powerful computing capabilities in multiple areas.Its mainstream model abstracts data through hierarchical representations.As the third generation neural network,spiking neural network's unique neuron model makes its influence in the field of brain-like computing increasing.Previous studies have shown that spike timing dependent plasticity(STDP)can be used in spiking neural networks to extract simple visual features.However,the networks used in these studies are relatively shallow and usually only one layer can be trained.Another study shows that deep learning uses rate-based backpropagation neural networks to increase recognition robustness by increasing the number of network layers.Therefore,this dissertation carries out a lot of theoretical and practical research on spiking neural network and deep learning.Based on the analysis of the existing deep learning model,this thesis integrates the special neuron model of spiking neural network,synapse learning rules and the processing of external input information mechanisms,constructed a STDP-based spiking deep neural network(SDNN)and proposed a pulse-time neural coding,which mainly includes the following contents:(1)Studied and proposed a dynamic region generation algorithm based on Spiking neural network,and as a coding mechanism of SDNN.The algorithm divides the picture into small areas in advance and uses the ignition time of different areas as the network coding value.Compared with spiking's traditional encoding method,the algorithm is closely integrated with biological mechanisms.The ignition mechanism comes from the way that neurons transmit information.The diversity of network morphology comes from multiple receptive fields in the visual cortex,providing a new way to generate superpixel regions.This algorithm provides a new perspective to extract features and classify images for spiking neural network.(2)Studied and proposed STDP-based Pulsed Deep Neural Network(SDNN).In order to adapt to the requirement of spiking neural network,the convolutional neural network(CNN)architecture was adjusted and STDP learning rules were adopted so that the neurons gradually learned the characteristics of the input images.The validity of the new model was verified by experiments in the Caltech 101 dataset and the MNIST dataset,respectively.SDNN is more efficient and has better classification effect than traditional spiking neural network model.This shows that the coding method and learning mechanism(STDP rules and competitive learning)effectively distinguish the representation of objects.This combination is the key to understanding the primate vision system learning methods,achieving low energy consumption and improving processing capabilities,and it is also significant for artificial vision system.