A Study Of Spiking Neural Network Algorithm Based On Spike-Timing Coding

The spiking neural network is a computational model constructed from neurons with biological behavior.Compared to the artificial neural network,it shares more common points with the biological nervous system on computing paradigm.Among all spiking neural network algorithms,the ones based on spike-timing coding can make full use of the information encoded by spike trains,while the ones based on rate coding can not.Thus the former ones have stronger computing power.The focus of this thesis is mainly the spiking neural network algorithm based on spike-timing coding.The methods for constructing spiking neural network models are reviewed.The SpikeProp algorithm is analyzed in detail,and it is generalized on recurrent networks.The software acceleration schemes for the algorithm are proposed and compared.The generalized algorithm is tested on the Iris flower dataset and the MNIST dataset.Finally,the hardware acceleration schemes for inference tasks using the spiking neural network are also analyzed.The main innovations of the thesis are as following:1.The key formula of SpikeProp algorithm and its sufficient condition are proved through a formal method.The method of computational graph is introduced in order to intuitively explain the process of error back-propagation in SpikeProp.And it is used to generalize SpikeProp algorithm to Recurrent SpikeProp algorithm on general structures of spiking neural network,assisted with the principle of total differential form invariance.2.The software acceleration schemes based on GPUs,which use GeNN library and Pytorch library are proposed.They are tested and compared in performance.3.A scalable architecture of the mixed-signal spiking neural network chip for inference tasks is proposed.In this architecture,a CAM and an SRAM are used to store the parameters of the spiking neural network,and each single synapse circuit is used for multiple logical synapses.It separates the computation and the parameter storage,thus achieves the better utilization of the circuits.The input and output units,together with the routing scheme based on binary search tree,also make it possible to achieve better scalability.