自编码神经网络理论及应用综述  

Theories and Applications of Auto-Encoder Neural Networks:A Literature Survey

作　　者：袁非牛[1,2] 章琳 史劲亭[1,4] 夏雪 李钢 YUAN Fei-Niu;ZHANG Lin;SHI Jin-Ting;XIA Xue;LI Gang(School of Information Technology, Jiangxi University of Finance and Economics, Nanchang 330032;College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418;School of Mathematics and Computer Science, Jiangxi Science and Technology Normal University, Nanchang 330038;Vocational School of Teachers and Technology, Jiangxi Agricultural University, Nanchang 330045;College of Mathematics and Computational Science, Yichun University, Yichun, Jiangxi 336000)

机构地区：[1]江西财经大学信息管理学院,南昌330032 [2]上海师范大学信息与机电工程学院,上海201418 [3]江西科技师范大学数学与计算机科学学院,南昌330038 [4]江西农业大学职业师范技术学院,南昌330045 [5]宜春学院数学与计算机科学学院,江西宜春336000

出　　处：《计算机学报》2019年第1期203-230,共28页Chinese Journal of Computers

基　　金：国家自然科学基金(61862029);江西省高校科技落地计划(KJLD12066);江西省教育厅科技项目(GJJ170317)资助~~

摘　　要：自编码器是深度学习中的一种非常重要的无监督学习方法,能够从大量无标签的数据中自动学习,得到蕴含在数据中的有效特征.因此,自编码方法近年来受到了广泛的关注,已成功应用于很多领域,例如数据分类、模式识别、异常检测、数据生成等.该文对传统自编码基础理论、自编码方法、改进技术以及应用领域进行了比较全面的综述.首先,该文介绍传统自编码基础理论与实现方法,分析自编码器的一般处理框架.然后,讨论现有各种改进的自编码器,分析这些方法的创新点、所要达成的目的和可能存在的问题.随后,该文介绍自编码器的实际应用领域,分析这些领域的代表性自编码算法,并详细地分析、比较和总结这些方法的特点.最后,总结现有方法存在的问题,并探讨了自编码器的将来发展趋势和可能挑战.An auto-encoder is one of unsupervised learning methods in the deep learning community.It can automatically learn effective and robust features from a large amount of unlabeled data.It has been received a lot of attention in recent years.Therefore,auto-encoders and their variants have successfully been applied in a wide range of fields,such as image classification,pattern recognition,anomaly detection,and data generation.To provide researchers a quick overview of auto-encoders,this paper makes a full survey on the basic theory of traditional auto-encoders,related algorithms,improved techniques and several applications in detail.We first introduce the fundamental theory and common implementations of auto-encoder methods,and then we focus on analyzing the general processing framework of auto-encoder methods.A general auto-encoder method usually contains an encoding structure and a decoding one.The encoding structure is often a contracting path that provides abundant context information for each pixel,while the decoding one is just an expanding path for localization information.Then,we discuss traditional auto-encoder methods and a series of improved methods,and analyze the innovation,motivation and existing problems of these methods.Specifically,we successively introduce some representative methods of improved auto-encoder methods,and we also point out the advantages and drawbacks of these auto-encoder methods.These methods analyzed in this paper mainly include denoising auto-encoders,marginalized denoising auto-encoders,sparse auto-encoders,contractive auto-encoders,saturating auto-encoders,convolutional auto-encoders,transforming auto-encoders,and other variants of auto-encoders.In addition,we specially discuss two kinds of marginalized auto-encoder methods,which are marginalized denoising auto-encoders for domain adaption and marginalized denoising auto-encoders for nonlinear representation,respectively.Afterwards,we introduce the major application fields of these methods based on auto-encoders,and then we also

关 键 词：自编码器 深度学习 无监督学习 特征学习 约束 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]