| In recent years,neural networks have produced important breakthroughs in the direction of object detection,image recognition,semantic segmentation and other applications,but there are certain differences between the real brain mechanism and the computer mechanism of existing neural networks,which required large amount of resources for computation has led to an arithmetic bottleneck and energy consumption bottleneck.Spiking Neural Networks(SNN),inspired by biological plausibility,is the third generation of artificial neural networks,which have advantages of low energy consumption,high computational efficiency,SNNs have wide application prospects in processing largescale,complex temporal information tasks and ultra-low-power edge computing tasks.In addition,the combination with SNN and the store-and-calculate integrated brain-inspired chip provides a new solution to break the von Neumann bottleneck and break through the current AI arithmetic hindrance.However,the high performance in supervised training for SNN is limited due to the complex neuronal dynamics and the non-differentiable of spike activity.This paper based on the study of surrogate gradient function of SNN in back propagation learning algorithm,we explore the image classification algorithm of SNN in dynamic vision direction and deploy it in hardware application,the work in this paper is as follows:(1)Neuromorphic datasets are widely used as training benchmarks by SNN.This work uses dynamic visual sensors to capture gesture data and perform complete data preprocessing including data segmentation,position integration and feature extraction.SNN with two training algorithms are built for dataset classification,new synaptic operation restriction terms are introduced into the loss function to reduce spike firing rate and network power consumption,and the resulting models are deployed on a brain-inspired processor to validate the low latency and low power consumption benefits of SNN.(2)To address the problem of accuracy loss due to gradient errors generated by using surrogate gradient functions during SNN training,this paper models the information flow over the spatio-temporal domain and proposes an progressive gradient learning algorithm that preserves the gradient propagation information and is able to obtain accurate gradients while maintaining the update capability of the network.Finally,experiments are conducted on mainstream static and neuromorphic datasets,and the higher network accuracy proves the effectiveness of the algorithm.(3)This paper proposes an integrated acceleration scheme for the storage and computation of neural networks based on memristors for the SCNN structure designed in this paper.Firstly,the characteristics of the memristor and the mainstream memristor models are analysed,and the advantages of the memristor model based on SNN architecture are further compared with other neural network architectures.Finally,hardware deployment on the memristor neural network is carried out to achieve network computation and acceleration using the features of memristors. |
