第3讲 深度学习中的经典网络模型及训练方法  

Classical Network Models and Training Methods in Deep Learning

作　　者：韩伟[1] 张雄伟[2] 张炜[3] 吴从明[4] 吴燕军[3] 

机构地区：[1]解放军理工大学指挥信息系统学院研究生2队 [2]解放军理工大学指挥信息系统学院 [3]中国卫星海上测控部 [4]中国人民解放军73615部队

出　　处：《军事通信技术》2016年第1期90-97,共8页Journal of Military Communications Technology

基　　金：国家自然科学基金资助项目(61402519;61471394);江苏省自然科学基金资助项目(BK20140071;BK20140074;BK2012510)

摘　　要：深度学习依靠其良好的复杂特征提取表达能力,在诸多研究领域都引起了强烈的反响,并越来越受到关注。网络模型结构是深度学习的重要组成内容,模型的训练方法则决定着网络的性能。文章以深度学习中的经典网络模型和训练方法来展开研究,首先简单介绍了深度学习中常用的网络模型,然后针对经典的深度置信网络和堆叠自编码器的基本组成模型及基本模型的训练算法,由基本模型构建的深度网络及其训练方法,模型的常见变型等方面进行了详细研究,最后展望了网络模型及其训练方法的发展方向。Depending on its higher ability of extracting complex features, deep learning has attracted more and more attention and caused strong repercussions in many research areas. Network architecture is an important aspect of deep learning and the training methods affect the performance of the network. In this paper, the classical network models and training methods in deep learning were studied. Firstly, common network models widely used in deep learning were introduced briefly. Then, the basic models and the corresponding training algorithms of both the deep belief networks and the stack auto-encoder, the deep hierarchical networks and their training algorithms constructed by the basic models were presented. Finally, the conclusions and the prospects of the network architectures and the training methods were discussed.

关 键 词：深度学习 网络模型 训练方法 深度置信网络 堆叠自编码器 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]