基于GPU实时仿真小脑模型神经网络优化的研究  

A Real-time Simulation Cerebellar Neural Network Model Based on GPU Optimization Research

作　　者：陈嘉卿 黄嘉嘉 许弢 CHEN Jia-qing;HUANFG Jia-jia;XU Tao(Faculty of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学智能制造学部,广东江门529020

出　　处：《五邑大学学报（自然科学版）》2023年第4期31-37,46,共8页Journal of Wuyi University(Natural Science Edition)

基　　金：五邑大学高层次人才科研启动项目(2019AL020);江门市科技局项目(2020JC01036)。

摘　　要：为了解决实时仿真小脑模型存在仿真规模小、仿真时序波动较大以及不能充分利用硬件性能等问题,本文提出了一种并行处理多种不同计算模块的优化方案.首先引入显卡的流特性优化模型的并行架构,然后为不同计算模块和不同细胞分配相应的线程加速网络计算.最后通过仿真延迟眨眼条件反射试验验证小脑模型在运动学习中的记忆形成,并分析了仿真细胞的电生理结果.本文提出的优化方案对多并行模型同样适用,该模型对小脑控制器的研究和小脑运动学习记忆形成具有理论意义和工程上的应用价值.In order to solve the problems of small simulation scale,large simulation time sequence fluctuation and insufficient utilization of hardware performance,an optimization scheme for parallel processing of various computing modules is proposed in this paper.Firstly,the streaming characteristics of the graphics card are introduced to optimize the parallel architecture of the model,and then the network computation is accelerated by assigning threads to different computational modules and different cells.Finally,the memory formation of cerebellar model in motor learning was verified by simulated delayed blink conditioned reflex test,the electrophysiological results of the simulated cells were also analyzed.The optimization scheme proposed in this paper is also applicable to multi-parallel models,which has theoretical implications and engineering application value for the study of cerebellar controllers and cerebellar motor learning memory formation.

关 键 词：图形处理器 统一计算设备架构 小脑模型 运动记忆 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]