受限玻尔兹曼机及其变体研究综述  

Review of research on restricted Boltzmann machine and its variants

作　　者：汪强龙 高晓光[1] 吴必聪 胡子剑 万开方[1] WANG Qianglong;GAO Xiaoguang;WU Bicong;HU Zijian;WAN Kaifang(School of Electronic Information,Northwestern Polytechnical University,Xi’an 710129,China;Stratégie etéconomie d’Entreprise,Paris 1 Panthéon Sorbonne,Paris 75005,France)

机构地区：[1]西北工业大学电子信息学院,陕西西安710129 [2]巴黎一大管理学院,巴黎75005

出　　处：《系统工程与电子技术》2024年第7期2323-2345,共23页Systems Engineering and Electronics

基　　金：国家自然科学基金(61573285,62003267);陕西省重点研发计划项目(2023-GHZD-33);中央高校基本科研业务费专项资金(G2022KY0602);电磁空间作战与应用重点实验室(2022ZX0090)资助课题。

摘　　要：受限玻尔兹曼机作为学习数据分布和提取内在特征的典型概率图模型,是深度学习领域重要的基础模型。近年来,通过改进受限玻尔兹曼机的模型结构和能量函数得到众多新兴模型,即受限玻尔兹曼机变体,可以进一步提升模型的特征提取性能。研究受限玻尔兹曼机及其变体能够显著促进深度学习领域的发展,实现大数据时代海量信息的快速提取。基于此,对近年来受限玻尔兹曼机及其变体的相关研究进行系统回顾,并创新性地从训练算法改进、模型结构改进、模型深层融合研究和模型相关最新应用4个方面进行全面综述。其中,重点梳理受限玻尔兹曼机训练算法和变体模型的发展史。最后,讨论受限玻尔兹曼机及其变体领域的现存难点与挑战,对主要研究工作进行总结与展望。As a typical probabilistic graphical model for learning data distribution and extracting intrinsic features,the restricted Boltzmann machine(RBM)is an important fundamental model in the field of deep learning.In recent years,numerous emerging models,i.e.,RBM variants,have been obtained by improving the model structure and energy function of RBM,which can further enhance the feature extraction performance of the model.The study of RBM and its variants can significantly contribute to the development of the deep learning field and realize the rapid extraction of massive information in the era of big data.Based on this,the relevant research on RBM and its variants are systematically reviewed in recent years,and the improvement of training algorithm,model structure,deep model fusion research and the latest application are creatively reviewed.In particular,the focus is on sorting out the develop history of training algorithms and variants for RBM.Finally,the existing difficulties and challenges in the field of RBM and its variants are discussed,and the main research work is summarized and prospected.

关 键 词：受限玻尔兹曼机 深度学习 受限玻尔兹曼机变体 概率无向图 特征提取 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]