摘要:概率图模型是目前机器学习研究的热点，基于概率图模型构造的生成模型已广泛应用于图像和语音处理等领域.受限玻尔兹曼机（restricted Boltzmann machines，简称RBMs）是一种概率无向图，在建模数据分布方面有重要的研究价值，RBMs既可以结合卷积算子构造深度判别模型，为深度网络提供统计力学的理论支持，也可以结合有向图构建生成模型，提供具有多峰分布的先验信息.主要综述了以RBMs为基础的概率图模型的相关研究.首先介绍了基于RBMs的机器学习模型的基本概念和训练算法，并讨论了基于极大似然估计的各训练算法的联系，比较了各算法的log似然损失；其次，综述了RBMs模型最新的研究进展，包括在目标函数中引入对抗损失和W距离，并构造基于RBMs先验的变分自编码模型（variational autoencoders，简称VAEs）、基于对抗损失的RBMs模型，并讨论了各实值RBMs模型之间的联系和区别；最后，综述了以RBMs为基础的模型在深度学习中的应用，并讨论了神经网络和RBMs模型在研究中存在的问题及未来的研究方向.



Abstract:The Probabilistic graph is a research hotspot in machine learning at present. Generative models based on probabilistic graphs model have been widely used in image generation and speech processing. The restricted Boltzmann machines (RBMs) is a probabilistic undirected graph, which has important research value in modeling data distribution. On the one hand, the RBMs model can be used to construct deep neural network, and on the other hand, it can provide statistical support of deep nets. This paper mainly summarizes the related research of RBMs based probability graph model and their applications in image recognition. Firstly, this paper introduces the basic concepts and training algorithms of RBMs. Secondly, this paper summarizes the applications of RBMs in deep learning; and then, this paper discusses existing problems in research of neural nets and RBMs. Finally, this paper gives a summary and prospect of the research on the RBMs.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5840