Research On Spiking Neural Network Algorithm For Image Classification

The spiking neural network is the most studied brain-inspired artificial neural network at present.Compared with the traditional artificial neural network,the spiking neural network is the accurate modeling of biological neurons,the information processing mechanism is modeling the synaptic spike and only participates in the calculation when the spike is fired.In theory,it has low power consumption and strong nonlinearity processing capacity.Spiking neural network is usually used for spatio-temporal pattern recognition,which is processing data with temporal and spatial characteristics.At present,the spiking neural network has poor classification accuracy or needs to combine with traditional classification methods.According to the image processing algorithm of the visual cortex,we establish a spiking neural network model with complex structure and biological feasibility to improve the performance in image recognition.Firstly,electrophysiological studies show that the mammalian visual cortex V2 has directional selectivity to visual stimuli and improve the robustness in image recognition.On this basis,we build spiking neural networks for handwritten digit classification.After got the firing frequency of V2,we designed a classification method based on the spiking neural network and compared it with the traditional convolution neural network.Secondly,in the visual cortex,neurons in the MSTd region can efficiently identify the spiking patterns from neurons in the MT region,the process is similar to sparse coding in non-negative matrix decomposition.In this thesis,we use the synaptic plasticity and the evolutionary algorithm to adjust the weights of spiking neural networks automatically,realized local features extraction and image reconstruction,Finally,we verify the classification of MT optic flow by support vector machine.Finally,through the above two visual cortex algorithms,we established spiking neural networks for image classification,reducing the number of samples and classification time,and improved the effect of spiking neural networks in image classification.