Research And Application Of Multilayer Spiking Neural Network Learning Algorithm Based On Hierarchical Feature Extraction

As a new generation of artificial neural networks,Spiking Neural Networks(SNN)has strong biological foundation and computational potential.However,the multi-layer SNN learning algorithm has shortcomings such as information coding redundancy,learning instability and slow convergence.Therefore,this paper uses the improved HMAX model and spike smoothing strategy to optimize neural information coding and learning algorithm of multi-layer SNN.The details are as follows:(1)A neural information coding algorithm based on improved HMAX model and phase encoding is proposed to optimize information feature extraction and spiking sequence generation.The algorithm uses Gabor filter with different angles to simulate the directional sensitivity of the visual receptive field,and the Fast ICA algorithm is used to fuse the information sparse expression into the HMAX model to extract higher-level features and reduce information redundancy.Finally,the algorithms draws on the idea of phase coding,and uses the improved spike sampling method to transform the feature information into a spiking sequence,which provides effective information expression for multi-layer SNN learning.(2)A multi-layer SNN learning algorithm based on Spike-error-smoothing and Normalization Strategy is proposed to improve the learning stability and convergence speed of multi-layer SNN.The algorithm considers the effects of presynaptic and postsynaptic spikes and uses the learning window to smooth the neighborhood spikes of the target spike,thereby providing a more stable synaptic weight adjustment.At the same time,the algorithm uses the idea of batch normalization to normalize the synaptic weights,avoiding the synaptic weight falling into the gradient saturation region and improving the learning convergence speed.(3)Based on the above research results,a multi-layer SNN image recognition model based on hierarchical feature extraction is proposed,which provides a new method for SNN application to pattern recognition.Experiments show that,on the data sets of MNIST and part of Caltech101,the image recognition model has better image recognition ability than the same type models and better robustness to noise.