CASIA OpenIR  > 毕业生  > 硕士学位论文
基于深度神经网络的无监督特征学习
Alternative TitleUnsupervised Feature Learning Based on Deep Neural Network
黄培浩
Subtype工学硕士
Thesis Advisor王亮
2015-05-21
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机应用技术
Keyword深度学习 无监督特征学习 聚类分析 数据降维 Deep Learning Unsupervised Feature Learning Clustering Analysis Dimensionality Reduction
Abstract随着移动互联网的发展,从移动设备涌入大量多样化的数据,这些数据的处理和分析给我们带来了机遇和挑战。而深度学习研究也在计算能力的提升以及大数据等因素的推动下,突破了计算能力不足、易过拟合等困难,开始发挥其高度抽象的层级式表达能力,成为处理复杂互联网数据的强有力工具。然而大量无标注数据的增长,带来了数据的标注难题,也带来了对无监督特征学习的巨大需求。 本文的工作以深度神经网络为工具,主要针对大量无标注数据的特征学习问题,研究内容如下: 1. 提出一种面向聚类分析应用的深度嵌入网络,该模型通过保持数据原始空间中的局部结构并增加对数据的表达的组稀疏约束,从而使深度神经网络学习一种适于聚类的表达。克服了传统的聚类方法没有充分挖掘和利用数据中深层特征的缺点。该模型在人脸及物体图像数据集上的实验效果优于传统的聚类方法。 2. 提出一种基于深度广义自编码机的降维方法,该方法通过迭代地探索数据点之间的关系以及数据点的表达,学习到数据内在的低维流形空间中的数据特征。克服了传统降维和子空间学习方法过于简单以及自编码机没有充分利用数据点之间关系的缺点。在流形学习、人脸识别等任务中,深度广义自编码机都要优于传统的降维方法。 以上两个基于深度神经网络的工作,都在无监督特征学习中充分利用了数据之间的关系。这些关系初步揭示了数据的内部结构,而这种结构正是无监督特征学习的关键所在。
Other AbstractWith the rapid development of the mobile Internet, more and more diverse data come from mobile devices. These big data bring us challenges and chances on data processing and analysis. At the same time, with accelerated computing power and big data, Deep Learning has overcome its disadvantages such as overfitting. It has become a powerful tool to handle complex data on the Internet. While so many data are available, most of them are unlabeled and this brings a great need for unsupervised feature learning. This thesis takes the deep neural network as a tool and targets on the problem of unsupervised feature learning. We propose in this thesis two methods to do unsupervised feature learning with deep neural network. The first one is the deep embedding network for clustering, which makes use of local structure preservation and group sparsity to force the network to learn a clustering-oriented representation. The second one is the generalized auto-encoder which aims to learn a feature from the intrinsic manifold space of data by exploring the relation between data points and learning their representation iteratively.
Other Identifier201228014629070
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/7763
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
黄培浩. 基于深度神经网络的无监督特征学习[D]. 中国科学院自动化研究所. 中国科学院大学,2015.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20122801462907(6741KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[黄培浩]'s Articles
Baidu academic
Similar articles in Baidu academic
[黄培浩]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[黄培浩]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.