基于MapReduce的卷积神经网络算法研究

A method for convolutional neural networks training based on MapReduce framwork

卷积神经网络(convolutional neural networks,CNN)是深度学习技术应用最成熟的模型之一,利用卷积神经网络来提取特征进行目标识别和分析是当前比较热的研究方向。目前CNN主要以单机串行方式实现,随着大数据时代的到来,串行模式突显出训练时间过长,内存不足等问题。为此,本文提出了一种在分布式处理Hadoop平台上,基于MapReduce框架并行训练CNN的算法MR-TCNN。并通过实验证明,提出的方法与传统单机串行训练方式相比,在大数据上有更快的训练速度。

Convolutional neural networks （CNN） are one of the most mature deep learning models. Howev- er, CNN is generally serially trained by one machine, and serial training has problems such as long time consuming, and difficulties to train due to insufficient memory when facing massive data. For these prob- lems, this paper presents a method for CNN training based on MapReduce framework, MR-TC- NN. Experimental results demonstrate that this method has better training speed, comparing with convention- al training method.