Brain-Inspired Learning Algorithm Based On Spiking Neural Network And Its Application

Biological vision system has powerful ability to process visual information,neuroscience has revealed several key mechanisms about vision system in recent years,providing a powerful biological basis for simulating visual system and brain-inspired learning algorithms.At present,although deep learning has made great breakthroughs,its learning algorithm is quite different from the biological neuron system.Spiking Neural Network(SNN)is considered as the "Third generation artificial neural network",its signal encoding and transmission are more biological reasonable.This essay constructs a hierarchical SNN model and compares the learning effects of unsupervised STDP rule,reward-modulation STDP rule and supervised pre-training conversion algorithm on SNN model.The main work and innovation of this essay are as follows:(1)A hierarchical SNN model based on unsupervised STDP rule is proposed to simulate the feature extraction and recognition in ventral pathway.Visual cortex cells are simulated by spiking convolution and max-pooling,and the competitive of "winner take all" is utilized by referring to the biological lateral inhibition,so that spike neurons can extract specific visual features according to STDP rules.To solve the "stability and plasticity dilemma" caused by STDP rule,the intrinsic homeostasis plasticity are used to optimize the spike rates and synaptic weights in SNN.The classification accuracy on MINIST and ETH-80 datasets reaches 97.58% and 82.5% respectively,which surpasses other unsupervised training SNN models.(2)Referring to the effect of dopamine in brain,two reward mechanisms are proposed which are RM-STDP rule and error adjust membrane potential.Reward signals are made to control the change direction of synaptic weights by STDP globally,so the rewarded(punished)activities in SNN are enhanced(weakened).The weights change in the direction of decreasing SNN output errors,effectively improves the limitation of unsupervised STDP rule,and also improves the classification performance of SNN.The classification accuracy on MINIST and ETH-80 datasets reaches 98.0% and 88.4%,which is significantly higher than that of SNN based on unsupervised STDP rule.(3)Combining the advantages of ANN and SNN,the supervised pre-training parameter conversion in SNN is realized,and the spike rates of neurons is dynamically adjusted by using the homeostasis mechanism.Supervised error-backpropagation algorithm is used to pre-train an ANN model,and then convert the trained parameters which are normalized into a SNN with the same structure,realizing the effective training of deep SNN.To solve the spike rates reduction in converted SNN,using homeostasis mechanism to adjust the spike rates of deep neurons in SNN dynamically.Compensates for the cumulative error of deep neurons and speeds up the simulation convergence speed of the converted SNN.