基于改进MapReduce模型的BP神经网络并行化研究  

Research on Parallelization of BP Neural Network based on Modified MapReduce Model

作　　者：李楠 于孟渤 贾珍珍 王一惠 李昕宸 邹淑雪[1] LI Nan;YU Meng-bo;JIA Zhen-zhen;WANG Yi-hui;LI Xin-chen;ZOU Shu-xue(College of Computer Science and Technology,Jilin University,Changchun Jilin 130012,China)

机构地区：[1]吉林大学计算机科学与技术学院,吉林长春130012

出　　处：《通信技术》2018年第4期799-804,共6页Communications Technology

基　　金：2017年吉林大学国家级大学生创新创业训练项目(No.2017A53224)~~

摘　　要：为了提高BP神经网络算法并行化速率,利用神经网络并行化思想,提出了一种基于Hadoop平台的改进Map Reduce编程模型及并行化的实现。采用Map Reduce编程模型,用神经网络训练集的一组样本的键/值替代单一键/值,通过分组标记将同一value值对应的reduce工作方式分散为多个reduce进行工作,实现各个任务节点并行处理大数据,从而减少了处理大规模数据集的运行时间。选用不同大小数据集进行测试,通过与传统的神经网络并行化进行对比,发现改进后的Map Reduce并行编程模型提高了神经网络的并行速率,在处理大数据集时具有一定的优越性。In order to improve the parallelization rate of BP neural network algorithm,using neural network parallelization idea,an improved MapReduce programming model based on Hadoop platform and its parallelization are proposed.By using MapReduce programming model,the keysalues of a set of samples of a neural network training set are used to replace a single keyalue,and the reduce work mode corresponding to the same value is dispersed into multiple reduce tasks by grouping markup to implement each task node to process large data in parallel,thereby reducing time for processing large-scale data sets.Different sizes of data sets are selected for testing and compared with the traditional neural network parallelization.It was found that the improved MapReduce parallel programming model improves the parallel speed of neural networks and has certain advantages when dealing with large data sets.

关 键 词：BP神经网络 MapReduce编程模型 MapReduce改进模型 大数据集 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]