Research On Feedforward Multi-Spike Neural Network Based On Hierarchical Visual Algorithm

As the third generation of neural networks,Spiking Neural Network(SNN)is a dynamical system which can effectively simulate biological neuron of continuous delivery of information over time.SNN is a core part of the neuromorphic model and has high scientific value.Neuromorphic model can be divided into two parts: neural information coding and learning algorithms.And neural information coding includes feature extraction and spiking sequence generation.Learning algorithm processes spiking sequence generated by the encoding.The main methods of feature extraction are HMAX,PCA,LDA etc.HMAX simulates the four-layer structure of V1 cells in the biological receptive field.It has drawn much attention for its good feature extraction performance.SNN is more biomimetic because of its unique neuron construction and calculation formula of synaptic weight.Its main algorithm are PSD,Tempotron,Re Su Me etc.PSD uses the time difference between the target spiking sequence and the actual spiking sequence to adjust the synaptic weight of spiking neural network.and PSD has strong robustness.This paper studies the SNN based on visual hierarchical system.HMAX and PSD are optimized to propose a neuromorphic model.The work of this article mainly has three aspects:Firstly,We improve HMAX,because it is not consider the direction sensitivity of biological visual cortex cells and the sparseness of information transmission.In line with the sparseness,we simplify the S1 layer using a single Gabor filter window.To simulate biological vision's sensitivity to the vertical direction,we strengthen the feature of filtered orientation by adding a sharpened replica of the filtered image.Experiments show that the improved HMAX is more accurate and also has a good effect on the noise image processing.Secondly,we research for a feedforward and multi-spiking neural network algorithm based on Precise-Synaptic-Adjustment(PSA)to solve the problem that the weight adjustment based on the W-H rule of PSD is hard to guarantee.The solution is the best and the convergence speed is difficult to control.Experiments show that PSA can speed up the algorithm convergence and solve the problem of low learning efficiency,and its robustness is slightly stronger than PSD.Finally,HMAX-PSA is proposed.And it is used to solve the problem that the conventional machine learning method has poor performance in noise image recognition.Experiments show that the neuromorphic model is not only highly biomimetic,in line with the rules of biological cognitive behavior,but also can effectively extract the features of spatio-temporal data to complete the task of intelligent identification.It has strong robustness.