文献类型：期刊文献

英文题名：Double-fold localized multiple matrixized learning machine

作者：Zhu, Changming[1];Wang, Zhe[1,2,3];Gao, Daqi[1];Feng, Xiang[1]

机构：[1]E China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China;[2]Shanghai Key Lab Intelligent Informat Proc, Shanghai 200433, Peoples R China;[3]Soochow Univ, Prov Key Lab Comp Informat Proc Technol, Suzhou 215006, Peoples R China

年份：2015

卷号：295

起止页码：196

外文期刊名：INFORMATION SCIENCES

收录：;EI(收录号：20150600498857);WOS:【SCI-EXPANDED(收录号:WOS:000346543000012)】；

基金：This work was partially supported by Natural Science Foundations of China under Grant Nos. 61272198 and 21176077, Innovation Program of Shanghai Municipal Education Commission under Grant No. 14ZZ054, the Fundamental Research Funds for the Central Universities, and Shanghai Key Laboratory of Intelligent Information Processing of China under Grant No. IIPL-2012-003.

语种：英文

外文关键词：Localized learning; Matrixized learning; Nonlinear learning; Gating model; Pattern representation

摘要：In this paper, we develop an effective multiple-matrixized learning machine named Double-fold Localized Multiple Matrixized Learning Machine (DLMMLM). The characteristic of the proposed DLMMLM is that it possesses double folds of local information from data. The first fold lies in the whole representation space which consists of different matrix representations. It is known that each pattern can be represented by different matrix representations. The matrices have their respective representation information and can play different discriminant roles in the final classification. Therefore from the viewpoint of the whole representation space, each matrix has its own local information. The second fold is that in each matrix representation learning, different pattems represented with the same matrix representation can carry different information. Therefore in the pattern space with the same matrix size, local information of different patterns should be introduced into the classifier design. On the whole, the advantages of the proposed DLMMLM are: (i) establishing a pattern-depended function in the matrixized learning so as to realize different roles of patterns for the first time; (ii) adopting the double-fold local information in both the representation space and the pattern space; (iii) proposing a new nonlinear classifier that is different from the state-of-the-art kernelization one; and (iv) getting a tighter empirical generalization risk bound in terms of the Rademacher complexity and thus achieving a statistically superior classification performance than those classifiers without the introduction of the double-fold local information. (C) 2014 Elsevier Inc. All rights reserved.