Adaptive natural gradient learning algorithms for unnormalized statistical models

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

2 Scopus citations

Abstract

The natural gradient is a powerful method to improve the transient dynamics of learning by utilizing the geometric structure of the parameter space. Many natural gradient methods have been developed for maximum likelihood learning, which is based on Kullback-Leibler (KL) divergence and its Fisher metric. However, they require the computation of the normalization constant and are not applicable to statistical models with an analytically intractable normalization constant. In this study, we extend the natural gradient framework to divergences for the unnormalized statistical models: score matching and ratio matching. In addition, we derive novel adaptive natural gradient algorithms that do not require computationally demanding inversion of the metric and show their effectiveness in some numerical experiments. In particular, experimental results in a multi-layer neural network model demonstrate that the proposed method can escape from the plateau phenomena much faster than the conventional stochastic gradient descent method.

Original languageEnglish
Title of host publicationArtificial Neural Networks and Machine Learning - 25th International Conference on Artificial Neural Networks, ICANN 2016, Proceedings
EditorsAlessandro E.P. Villa, Paolo Masulli, Antonio Javier Pons Rivero
PublisherSpringer Verlag
Pages427-434
Number of pages8
ISBN (Print)9783319447773
DOIs
StatePublished - 2016
Externally publishedYes
Event25th International Conference on Artificial Neural Networks, ICANN 2016 - Barcelona, Spain
Duration: 6 Sep 20169 Sep 2016

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume9886 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference25th International Conference on Artificial Neural Networks, ICANN 2016
Country/TerritorySpain
CityBarcelona
Period6/09/169/09/16

Keywords

  • Multi-layer neural network
  • Natural gradient
  • Ratio matching
  • Score matching
  • Unnormalized statistical model

Fingerprint

Dive into the research topics of 'Adaptive natural gradient learning algorithms for unnormalized statistical models'. Together they form a unique fingerprint.

Cite this