基于改进Hadoop的受限玻尔兹曼机云计算实现

Realization of RBM cloud computing based on improved Hadoop

针对受限玻尔兹曼机(RBM)面对大数据时存在模型训练缓慢的问题,设计了基于Hadoop的RBM云计算实现方法.针对RBM训练方法,改进了Hadoop任务消息通信机制以适应模型迭代周期短的特点;设计了MapReduce框架,包括Map端实现吉布斯采样,Reduce端完成参数更新;依据Hadoop任务组合方式,将RBM的训练应用于深度玻尔兹曼机(DBM)中.通过手写数字识别实验证明,该计算方法在大规模数据条件下能够有效加速RBM训练,且适应于深度学习模型的学习.

To resolve the slow training of Restricted Boltzmann Machine for handling large data the realization of RBM training based on cloud platform Hadoop is designed.In view of the training method of RBM Hadoop tasks message mechanism was improved to suit RBM′s short iteration cycle MapReduce framework was designed including Map function implemented Gibbs sampling and Reduce function completed parameter update based on Hadoop task combinations RBM′s cloud training was used in Deep Boltz？mann Machine′s training.The handwritten numeral recognition experiments show that this cloud training method can accelerate RBM training effective under large？scale data condition and work well in deep learning model training.