Large-scale Brain Network Simulation Technology Study On Super Computer

The brain science research areas has become the focus of the advanced countries in recent years.The European Union,the United States,and Japan and other countries have proposed their brain project.The research of human brain mechanism not only will deepen the understanding of the human brain,thus promoting the development of the treatment of brain diseases,but also will have revolutionary effects on information technology,enormous social and economic benefits.After decades of development,the computing peak of the fastest supercomputer TH-2 has reached 55 Mflops.The worldwide countries are developing Eflops-scale supercomputers,facing a number of significant technical challenges,such as system architecture,power consumption,reliability,programming mode.Research of brain science has important inspiration and impetus to solve the above issues.Neural computing chips researched by IBM brought a glimmer of hope to develop a supercomputer referencing functions of brain.Brain is mainly composed of neurons and synapses,and its mechanism of function is complex and diverse.From the point of computational view,the brain functions are realized by activities of neural networks composed of neurons.On the other hand,largescale neural network calculation and the current supercomputer have parallel,distributed,multi-level and other common features,so we can use supercomputer to simulate the neural network to realize the in-depth study of brain function.Brain is consist of billions of neurons,about ten thousand billion synaptic connections.The current simulation scale of neural network is only about 105,but regional connection of brain requires a minimum neural network of 106 neurons.This paper is based on the perspective of informatics study on brain function,primarily research on neurons,neural networks,and large-scale neural network simulations.Combining neural network simulations and the computational advantage of supercomputer TH-1A,this paper implemented the neural network simulation software(NEST)on the TH-1A.This study is based on neurons and neural networks,has conducted the study of neural network simulation and completed the following work:(1)Analyzing of neural network structure,neuron models,synaptic models,neural network models,selecting suitable models for the large scale neuronal network simulation.(2)Deploying of neural network simulation platform NEST on the TH-1A;building a large scale neural network using the NEST on the TH-1A;using analysis tools to study the memory usage and scalability of the NEST when simulating large-scale neuronal networks.(3)Studying the methods to improve the efficiency of the NEST simulation: reducing data structure redundancy of neurons and synapses;improving efficiency of network connections;running neuronal network simulation code efficiently on supercomputer TH-1A,so as to improve the performance of neural network simulations of the NEST,and expand the scale of simulation of neural networks.(4)Building 1mm~3 mammals full-scale simulation model of Cerebral Cortical areas,enable multiple neural network-level interconnection,research the methods of communications between certification levels of neural network.