Design And Implementation Of Partition Methods Of Brain Network Oriented To Computer Cluster

In recent years,brain simulation research combining computer science and brain science has become a new development direction in the field of artificial intelligence.Brain simulation is to better understand the brain and explore paths for the realization of general artificial intelligence.However,the complexity of neural network of the brain makes brain simulation put forward higher requirements for computing resources.High-performance computer cluster built on the basis of traditional CPU/GPU or neuromorphic chips has become the main hardware for brain simulation.How to efficiently simulate the brain on the computer cluster is a problem worth exploring.This thesis mainly studies how to divide the brain network to achieve efficient simulation when using computer cluster to simulate the brain.In this thesis,the brain network partitioning algorithm for computer cluster is designed.The brain network partitioning algorithm can divide the brain network into a corresponding number of sub-networks according to the number of computer nodes.The research on brain network partitioning algorithm includes the following three contents:(1)The thesis analyzes the problem of brain network partitioning and the main process of brain simulation in computer cluster.It is clear that the ultimate goal of brain network partitioning is to shorten simulation time of computer cluster simulating brain network.It can be concluded that the main factors affecting the simulation time are communication overhead and computing overhead.The brain network can be partitioned from two aspects of reducing communication traffic and balancing computational overhead.(2)The reason of partitioning the brain network with neuron population as the basic unit are explained in this thesis.(3)In order to balance the computing overhead of each computer node,the brain network partitioning algorithm based on the number of synapses is designed in this thesis.The brain network is divided based on the number of synapses corresponding to the neuron cluster,and a specified number of subnetworks are obtained.The total number of synapses corresponding to each subnetwork is approximately equal.In order to reduce the communication overhead of the computer cluster,the brain network partitioning algorithm based on the Louvain algorithm is designed in this thesis,the brain network is first transformed into an undirected weighted network with neuron population as a node,and the Louvain algorithm is then used to partition this undirected weighted network.The results of each partitioning are recorded,and an appropriate partitioning result is found according to the given number of computer nodes.At the same time,this thesis compares and analyzes the brain network partitioning algorithms based on synaptic number and the brain network partitioning algorithm based on the Louvain algorithm in terms of synaptic computation,communication and simulation time.In this thesis,a prototype system for brain network partitioning is also implemented.Firstly,the requirements of system are analyzed,and seven functional requirements are identified:user login,user registration,uploading the brain network connection file,customizing brain network connections,partitioning the brain network,viewing all partitioning results,and viewing the partitioning result details and result analysis.Secondly,the prototype system is designed and implemented,and the overall design of the system and the design of each module are described.In particular,the methods provided by the the NEST simulator are used in the customizing the brain network connection function.Finally,the system is tested,which shows that the system can divide the brain network and display the division results.