A comparative investigation on subspace dimension determination
Publication in refereed journal


Times Cited
Web of Science9WOS source URL (as at 01/04/2020) Click here for the latest count
Altmetrics Information
.

Other information
AbstractIt is well-known that constrained Hebbian self-organization on multiple linear neural units leads to the same k-dimensional subspace spanned by the first k principal components. Not only the batch PCA algorithm has been widely applied in various fields since 1930s, but also a variety of adaptive algorithms have been proposed in the past two decades. However, most studies assume a known dimension k or determine it heuristically. though there exist a number of model selection criteria in the literature of statistics. Recently, criteria have also been obtained under the framework of Bayesian Ying-Yang (BYY) harmony learning. This paper further investigates the BYY criteria in comparison with existing typical criteria, including Akaike's information criterion (AIC), the consistent Akaike's information criterion (LAIC), the Bayesian inference criterion (BIC), and the cross-validation (CV) criterion. This comparative study is made via experiments not only on simulated data sets of different sample sizes, noise variances, data space dimensions, and subspace dimensions, but also on two real data sets from air pollution problem and sport track records, respectively. Experiments have shown that BIC outperforms AIC, CAIC, and CV while the BYY criteria are either comparable with or better than BIC. Therefore, BYY harmony learning is a more preferred tool for subspace dimension determination by further considering that the appropriate subspace dimension k can be automatically determined during implementing BYY harmony learning for the principal subspace while the selection of subspace dimension k by BIC, AIC, CAIC, and CV has to be made at the second stage based on a set of candidate subspaces with different dimensions which have to be obtained at the first stage of learning. (C) 2004 Elsevier Ltd. All rights reserved.
All Author(s) ListHu XL, Xu L
Name of ConferenceWorkshop on New Developments in Self-Organizing Systems
Start Date of Conference11/09/2003
End Date of Conference14/09/2003
Place of ConferenceHibikino
Journal nameNeural Networks
Detailed descriptionthe Joint official Journal of International Neural Network SocietEuropean Neural Network Society and Japanese Neural NetworksSociety ,.
Year2004
Month10
Day1
Volume Number17
Issue Number8-9
PublisherPERGAMON-ELSEVIER SCIENCE LTD
Pages1051 - 1059
ISSN0893-6080
eISSN1879-2782
LanguagesEnglish-United Kingdom
KeywordsBYY harmony learning; data smoothing; model selection; principal component analysis; subspace dimension
Web of Science Subject CategoriesComputer Science; Computer Science, Artificial Intelligence; COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE; NEUROSCIENCES

Last updated on 2020-02-04 at 01:58