Geometrical singularities in the neuromanifold of multilayer perceptrons

Shun Ichi Amari, Hyeyoung Park, Tomoko Ozoki

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

15 Scopus citations

Abstract

Singularities are ubiquitous in the parameter space of hierarchical models such as multilayer perceptrons. At singularities, the Fisher information matrix degenerates, and the Cramer-Rao paradigm does no more hold, implying that the classical model selection theory such as AIC and MDL cannot be applied. It is important to study the relation between the generalization error and the training error at singularities. The present paper demonstrates a method of analyzing these errors both for the maximum likelihood estimator and the Bayesian predictive distribution in terms of Gaussian random fields, by using simple models.

Original languageEnglish
Title of host publicationAdvances in Neural Information Processing Systems 14 - Proceedings of the 2001 Conference, NIPS 2001
PublisherNeural information processing systems foundation
ISBN (Print)0262042088, 9780262042086
StatePublished - 2002
Event15th Annual Neural Information Processing Systems Conference, NIPS 2001 - Vancouver, BC, Canada
Duration: 3 Dec 20018 Dec 2001

Publication series

NameAdvances in Neural Information Processing Systems
ISSN (Print)1049-5258

Conference

Conference15th Annual Neural Information Processing Systems Conference, NIPS 2001
Country/TerritoryCanada
CityVancouver, BC
Period3/12/018/12/01

Fingerprint

Dive into the research topics of 'Geometrical singularities in the neuromanifold of multilayer perceptrons'. Together they form a unique fingerprint.

Cite this