随机受限玻尔兹曼机组设计

发布时间：2018-01-08 11:04

本文关键词：随机受限玻尔兹曼机组设计　出处：《上海交通大学学报》2017年10期 　论文类型：期刊论文

【摘要】：为提高受限玻尔兹曼机(Restricted Boltzmann Machine,RBM)数据学习能力和抑制训练的特征同质化问题,提出一种随机受限玻尔兹曼机组(Random-RBM Group,R-RBMG)设计.对观测数据进行随机维度组合,在随机维度组合的基础上构建子RBM群组并实施训练,随后依据神经网络的层数选择模型特征组合方式,针对浅层结构设置为均值组合方式,针对深层模型设置为隐单元叠加方式.理论分析表明,随着组内模型数目的增加,R-RBMG所要学习的训练目标将逐渐接近于标准RBM的训练目标,并且能够有效减少特征同质化带来的影响;实验结果表明,与衰落机制相比,R-RBMG能够有效提高RBM的特征学习能力,应用所组建的浅层结构和深层结构特征,将MNIST(Mixed National Institute of Standards and Technology)数据库实验的分类准确率分别提高了2%和0.4%.
[Abstract]:In order to improve the data learning ability of constrained Boltzmann machine restricted Boltzmann Machine (RBM) and the problem of feature homogeneity of suppression training. A random constrained Boltzmann unit Random-RBM Group R-RBMG design is proposed. The random dimension combination of observation data is carried out. On the basis of the random dimension combination, the sub-#en0# group is constructed and trained. Then, according to the layer number of neural network, the model feature combination mode is selected, and the shallow structure is set to mean combination mode. The theoretical analysis shows that the training target of R-RBMG will be close to the standard RBM training target with the increase of the number of models in the group. And it can effectively reduce the influence brought by the homogenization of characteristics; The experimental results show that compared with the fading mechanism, R-RBMG can effectively improve the feature learning ability of RBM, and apply the constructed shallow and deep structure features. MNIST(Mixed National Institute of Standards and Technology. The classification accuracy of database experiment was improved by 2% and 0.4 respectively.
【作者单位】：海军航空大学基础实验部;海军航空大学信息融合所;
【基金】：国家自然科学基金(61102167) 泰山学者工程专项经费(No.st201511020)资助
【分类号】：TP181
【正文快照】： 目前,以受限玻尔兹曼机(Restricted Boltz-mann Machine,RBM)为基础的深度学习模型已经在数据降维、图像分类、跨媒体检索和时间序列分析等机器学习领域得到了较好应用.与此同时,对RBM进行优化以提高深度模型数据学习能力也一直是深度学习领域的研究热点.RBM在训练过程中存在

【相似文献】