Presentation 2012-11-08
New Graphical Model "Firing Process Network" : A Model with Easy Learning
KAZUYA TAKABATAKE, SHOTARO AKAHO,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) We propose a versatile multivariate probabilistic model that can easily learn its structure and parameters from a given dataset. In conventional graphical models, structure-learning or parameter-learning is intractable for large models, since it concerns the whole network. In the proposed model, each node has a manifold respectively, and the model distribution is defined as the limiting distribution of a Markov chain that is iterative m-projections to these manifolds. An upper-bound of a Bregman divergence shows that the model distribution obtained by the proposed learning algorithm is close to the empirical distribution of data. The proposed learning algorithm is performed by a node-by-node manner, since each node is only responsible for its own manifold. Experiments shows the performance of the proposed model.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) graphical model / learning / computational cost / MCMC / information geometry
Paper # IBISML2012-78
Date of Issue

Conference Information
Committee IBISML
Conference Date 2012/10/31(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Information-Based Induction Sciences and Machine Learning (IBISML)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) New Graphical Model "Firing Process Network" : A Model with Easy Learning
Sub Title (in English)
Keyword(1) graphical model
Keyword(2) learning
Keyword(3) computational cost
Keyword(4) MCMC
Keyword(5) information geometry
1st Author's Name KAZUYA TAKABATAKE
1st Author's Affiliation Human Technology Institute, AIST()
2nd Author's Name SHOTARO AKAHO
2nd Author's Affiliation Human Technology Institute, AIST
Date 2012-11-08
Paper # IBISML2012-78
Volume (vol) vol.112
Number (no) 279
Page pp.pp.-
#Pages 8
Date of Issue