和积网络研究综述

Survey of sum-product networks

中文关键词  和积网络  SPNs的学习技术  概率图模型  深度学习

英文关键词  sum-product network  learn techniques of SPNs  probabilistic graphical model: deep learning

中文摘要

      和积网络(SPNs)是一种基于有根有向无环图的深度概率图模型. 除了叶节点外, 其余节点由求和节点或求积节点组成. 和积网络与概率图模型密切相关, 但是, 和积网络计算过程仅涉及简单的网络多项式求和运算和求积运算, 且能够实现精确和近似推理. 与经典的概率图模型相比, 和积网络可以从训练数据中构建易于推理的模型.此外, 和积网络也可以作为类似于神经网络的深度学习模型使用. 本文主要从和积网络的基本原理、理论研究、学习技术、变体模型及各领域具体应用等问题进行详细阐述. 首先, 概述和积网络的基本原理, 包括和积网络理论的研究现状. 其次, 概述了和积网络的几类变体模型, 并总结了和积网络学习技术中的参数学习和结构学习方面的学习算法. 除此之外, 本文还从自然语言处理、语音识别、医学研究等特定应用领域概述了基于和积网络的应用模型.最后, 根据现有的研究基础对和积网络未来的发展趋势及方向进行了展望.

英文摘要

      Sum-product networks (SPNs) are deep probabilistic graphical models based on rooted directed acyclic graphs. Except for leaf nodes, the remaining nodes are composed of summation nodes or quadrature nodes. Sum-product networks are closely related to probabilistic graph models, but the calculation process of sum-product networks only involves simple network polynomial summation and quadrature operations, and can achieve accurate and approximate reasoning. Compared with classical probability graph models, sum-product networks can construct models that are easy to reason from training data. In addition, sum-product networks can also be used as deep learning models similar to neural networks. This paper mainly elaborates on the basic principles, theoretical research, learning techniques, variant models, and specific applications in various fields of sum-product networks. Firstly, the basic principles of sum-product networks are summarized, including the current research status of sum-product network theory. Secondly, several types of variant models of sum-product networks are summarized, and learning algorithms for parameter learning and structure learning in sum-product network learning techniques are summarized. In addition, we have also outlined sum-product network based application models in specific application fields such as natural language processing, speech recognition, and medical research. Finally, based on the existing research foundation, the future development trends and directions of sum-product networks are prospected.