Model Optimization Of Spiking Neural Network And Its Application In Image Recognition

As one of the three cutting-edge technologies in the 21 st century,artificial intelligence has a far-reaching impact on social economy and people's way of life,and is leading the trend of development of the times.Neuroscience has been playing an important role in the development of artificial intelligence.Spiking neural network,called the third generation artificial neural network,is the product of the combination of neuroscience and artificial intelligence.By simulating the functions of neuron and synapse in biological neural system and information processing methods,Spiking neural network uses time coding to express and transmit information,which has rich biological characteristics.Spiking neural network has unique advantages in dealing with tasks related to spike trains,which has aroused concern and attention of a large number of scholars.In this thesis,the theoretical research and applied extension of Spiking neural network are conducted from two aspects of computational model and image recognition.(1)The external current in Spiking neuron model is optimized in this thesis.Firstly,the appropriate synapse model is selected and synaptic integration is used to generate the synaptic external current of neurons.Then,the effects of synaptic external current on the dynamics and performance of Spiking neural network are studied by designing related experiments and compared with the traditionally used noise external current.The results show that synaptic external current can significantly increase the complexity of network's activity,and can effectively enhance the reconstruction and classification capabilities of network.(2)In this thesis,a new image recognition method based on the bio-visual cortex processing mechanism is proposed.The image is extracted in the space-time orientation by synaptic integration mechanism and the primary visual cortex network.Firstly,the image features are separated by a preprocessing process similar to the bio-visual cortex processing mechanism,and then a simple Spiking supervised learning algorithm is adopted to train the output layer connections of the network.The results show that the proposed method achieves 96% accuracy on the MNIST dataset and is better than many other Spiking image recognition methods.The combination of preprocessing and supervised learning not only simplifies the training process of the network,but also reduces the complexity of the training algorithm,and greatly reduces the number of training samples and enables the network to have a higher recognition accuracy.The high-precision image recognition under small sample data is realized,which provides a basis for the wide application of the spiking neural network model in the field of intelligent computing,and has important theoretical and applied research value.