深度学习研究综述  

Review of Deep Learning

作　　者：尹宝才[1] 王文通[1] 王立春[1] 

机构地区：[1]北京工业大学城市交通学院多媒体与智能软件技术北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2015年第1期48-59,共12页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61390512)

摘　　要：鉴于深度学习在学术界和工业界的重要性,依据数据流向对目前有代表性的深度学习算法进行归纳和总结,综述了不同类型深度网络的结构及特点.首先介绍了深度学习的概念;然后根据深度学习算法的结构特征,概述了前馈深度网络、反馈深度网络和双向深度网络3类主流深度学习算法的网络结构和训练方法;最后介绍了深度学习算法在不同数据处理中的最新应用现状及其发展趋势.可以看到:深度学习在不同应用领域都取得了明显的优势,但仍存在需要进一步探索的问题,如无标记数据的特征学习、网络模型规模与训练速度精度之间的权衡、与其他方法的融合等.Considering deep learning＇s importance in academic research and industry application, this paper reviews methods and applications of deep learning. First, the concept of deep learning is introduced, and the main stream deep learning algorithms are classified into three classes： feed-forward deep networks, feed-back deep networks and bi-directional deep networks according to the architectural characteristics. Second, network architectures and training methods of the three types of deep networks are reviewed. Finally, state-of-the-art applications of mainstream deep learning algorithms is illustrated and trends of deep learning is concluded. Although deep learning algorithms outperform traditional methods in many fields, there are still many issues, such as feature learning on unlabeled data; the balance among network scale, training speed and accuracy;and model fusion.

关 键 词：深度学习 深度神经网络 卷积神经网络 反卷积网络 深度玻尔兹曼机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]