Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

On-line EM algorithm for the normalized gaussian network.

M Sato1, S Ishii

  • 1ATR Human Information Processing Research Laboratories, 2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0288, Japan.

Neural Computation
|January 15, 2000
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Topography of cat medullary ventral surface hypoxic acidification.

Journal of applied physiology (Bethesda, Md. : 1985)·1992
Same author

Ion channels on synaptic vesicle membranes studied by planar lipid bilayer method.

Biophysical journal·1992
Same author

Nucleotide sequence of cDNA for porcine heme oxygenase and its expression in Escherichia coli.

Biochemistry international·1992
Same author

Immunohistochemical evidence that rat FSH cells contain beta II-subspecies of protein kinase C.

Endocrinologia japonica·1992
Same author

Postnatal ontogenesis of neurons containing GABAA alpha 1 subunit mRNA in the rat forebrain.

Brain research. Molecular brain research·1992
Same author

Immunocytochemical and in situ hybridization evidence for the coexistence of GABA and tyrosine hydroxylase in the rat locus ceruleus.

The Anatomical record·1992

This study introduces an online EM algorithm for normalized Gaussian networks (NGnets). This method efficiently handles dynamic environments and complex function approximation tasks.

Area of Science:

  • Machine Learning
  • Artificial Intelligence
  • Computational Statistics

Background:

  • Normalized Gaussian Networks (NGnets) offer a powerful approach for function approximation by partitioning input spaces with local linear regression units.
  • Existing batch Expectation-Maximization (EM) algorithms can be computationally intensive, especially in dynamic environments.

Purpose of the Study:

  • To develop an efficient on-line EM algorithm for NGnets adaptable to dynamic environments.
  • To enhance the robustness of NGnets against singular input distributions.
  • To enable dynamic adaptation of network structure through unit manipulation.

Main Methods:

  • Derivation of an on-line EM algorithm from the batch EM algorithm using a discount factor.
  • Introduction of a regularization method to handle singular input distributions.

Related Experiment Videos

  • Implementation of unit manipulation mechanisms (production, deletion, division) for dynamic environment adaptation.

    • Main Results:

    • The proposed on-line EM algorithm is shown to be equivalent to the batch EM algorithm under specific discount factor scheduling.
    • The algorithm functions as a stochastic approximation method for maximum likelihood estimation.
    • Experimental results demonstrate suitability for function approximation in dynamic environments and effective application to robot dynamics problems.

    Conclusions:

    • The developed on-line EM algorithm provides an efficient and adaptive solution for NGnets in dynamic environments.
    • The regularization and unit manipulation mechanisms enhance the algorithm's robustness and flexibility.
    • The approach shows promise for real-world applications like robot dynamics, outperforming existing methods in comparative analyses.