Research On Efficient Routing Mechanism Of Large-scale Spiking Neural Network

Inspired by the research of brain science and neurocomputing,the spiking neural network,as a computational model that mimics the mechanism of the biological brain,has temporarily emerged in application scenarios such as bionic vision,smell,and memory,and has gradually become the mainstream of brain model research.According to brain science research,the number of neuronal synapses in the biological brain is 3-4 orders of magnitude larger than the number of neurons.The biological real-time processing time per cycle of the biological brain is about 1-10 ms,and the brain's working process is accompanied by a large number of spikes.Signal transmission.Therefore,the biological real-time simulation of the brain model can be abstracted into the communication problem of ultra-large-scale graph calculation and massive micro-spike packets,which brings huge challenges to the traditional von Neumann computer architecture.Among them,how to solve the problem of high fan-in and fan-out of a large number of tiny spike packets under the existing computer architecture is a key issue of the spike neural network simulation platform.To this end,this paper focuses on the scientific problem of designing and implementing the efficient routing mechanism of large-scale spiking neural networks under the three-dimensional closed-loop network topology of the CPU cluster environment.Efficient routing mechanism.Through node load simulation and macaque brain model simulation comparison experiments,it is proved that the spike transmission speed of this method is about 10% higher than that of the traditional method,and the speed of brain model simulation is about an order of magnitude.The main work of the thesis is as follows:(1)Combining the characteristics of the spiking neural network,compare and analyze the existing computer network topologies,and select the torus as the basic topology of the cluster.Analyze the relationship between the degree of the node and the radius of the network topology,the length of the routing table,and the cost of the link,and comprehensively select a six-degree torus as the topology model of the computing node to achieve low communication cost transmission.Based on this model,a set of multicast transmission protocol for small data packets is designed to reduce the number of spikes' router table lookups and the number of spike packets.According to the mapping structure of the imspike neural network in the cluster,a hierarchical routing table and an imspike multi-level routing table lookup mechanism are designed.(2)Use graph theory to establish mathematical models of neurons and cluster nodes for spike neural network engineering,study the relationship between synapses and node links between neurons,and optimize the calculation of the shortest path algorithm for spike transmission between pairs of nodes.It also analyzes the relationship between local path selection and global load balance,and proposes a path planning algorithm that takes into account the load balance between nodes,so that the number of paths supported by each node in the cluster is roughly equal to avoid node load imbalance.(3)Combining the neuron cluster hierarchical mapping logic,spike multicast transmission protocol,spike routing path calculation and planning,and three-level routing table structure proposed in this paper,a routing table generation algorithm combining node hardware resources is proposed to combine the spike routing mechanism.And synapse weight information is distributed and stored on each node.