文献类型：期刊文献

英文题名：A deep learning method for multimodal data

作者：Zhao, Haiyan[1,2]; Li, Guohe[1]; Niu, Wenliang[2]; Liu, Kun[2,3]

第一作者：Zhao, Haiyan;赵海燕

通讯作者：Zhao, Haiyan

机构：[1] College of Geophysics and Information Engineering, China University of Petroleum, Beijing, China; [2] College of Applied Science & Technology, Beijing Union University, Beijing, China; [3] College of Computer Science and Technology, Jilin University, Changchun, China

第一机构：College of Geophysics and Information Engineering, China University of Petroleum, Beijing, China

年份：2015

卷号：11

期号：12

起止页码：4237-4244

外文期刊名：Journal of Computational Information Systems

收录：EI(收录号：20153201107507);Scopus(收录号：2-s2.0-84938314753)

语种：英文

外文关键词：Artificial intelligence - Classification (of information) - Gaussian distribution - Modal analysis

摘要：Deep learning is an active area in machine learning community recently. It has been shown powerful in modeling high-level concepts from raw inputs, which is meaningful for artificial intelligence. However, when the training data samples are generated from multi-modal distribution, the induced concept space would be extremely complicated which would make it difficult for learning and knowledge discovery. We propose a deep learning model for multi-modal data learning in this paper, adopting a restricted boltzmann machine with multiple layers. We propose the method for model training and classification of unseen samples within a discriminative model, which models the predictive distribution of the training data samples. The proposed model is capable of capturing complex concept classes and has good representation ability. We evaluate the proposed model on two benchmark data sets of deep learning, and the results show that our method is effective and superior to currently state-of-the-art multi-modal deep learning methods. ?, 2015, Binary Information Press. All right reserved.