Affiliation : Dept.of Computational Intelligence and Systems Science, Tokyo Institute of Technology

Lab : Sumio Watanabe Lab in P&I Lab., Tokyo Institute of Technology.

Research : learning theory

I received a Ph.D degree from Tokyo-Tech(Sep,29,2006).

Doctoral Dissertation: Statistical Learning Theory of Variational Bayes
<ps, pdf>.

This dissertation analyses the variational Bayesian learning algorithms for mixture models.

The approximation accuracy of the variational Bayesian approach is shown by evaluating

the variational stochastic complexity,
also known as variational free energy.

Also the main results show how the hyperparameters of the prior distribution influence

the learning process and have some implications for how to design the learning algorithm.

Our paper:"Stochastic complexity for mixture of exponential families
in generalized variational Bayes"

will appear in Theoretical Computer Science.

I made a presentation in IDEAL2006.

Our paper:"Stochastic Complexities of General Mixture Models in Variational Bayesian Learning"

has been accepted for publication in Neural Networks, <pdf>.

It's my great pleasure to
announce publication of our paper:

"Stochastic Complexities of Gaussian Mixtures in Variational Bayesian
Approximation"

Kazuho Watanabe, Sumio Watanabe

JMLR 7(Apr):625--644, 2006 <pdf>.

Our paper:"Variational Bayesian Stochastic Complexity of Mixture Models"
has been accepted for a poster presentation at

NIPS*2005.

abstract

An oral presentation at ICONIP2005(Taipei, Taiwan, Oct.30-Nov.2). Paper title:"Variational Bayesian Algorithm and Stochastic Complexity for Mixture Models".

An oral presentation at NOLTA2005(Bruges, Belgium, Oct.18-21). Paper title:"On Variational Bayes Algorithms for Exponential Family Mixtures".

Our paper:"Stochastic complexity for mixture of exponential families in variational bayes" has been accepted for presentation at ALT2005(Singapore,Oct.8-11). In this presentation, I showed the asymptotic behavior of the stochastic complexity (the free energy) in the Variational Bayesian learning of mixture of exponential familiy distributions.

The 3rd Mathematical Science Forum in Tokyo Tech. I made a presentation entitled "Stochastic complexity and variational approximation".

Poster presentation at 2005 IEEE Tokyo Student Workshop(2005.02.15).

Oral presentation at IEEE International Conference on Cybernetics and Intelligent Systems(2004.12.01).

Proceeding manuscript<ps,
pdf>:"Lower bounds of stochastic complexities in
variational bayes learning of gaussian mixture models," Proc. of IEEE
CIS04,Singapore, pp.99-104, 2004.

Our paper "Estimation of the Data Region Using Extreme-value
Distributions" was accepted for the 15th international conference on
Algorithmic Learning Theory (ALT2004)(Padova,Italy,Oct.2-5).

I attended the international symposium
ISITA2004(Parma,Italy,Oct.10-13)
and made a presentation there.

IBIS2003(Kyoto,2003.11.11-12): "Estimating the Data Region Using the Asymptotic Distributions of Extreme-value Statistics"<ps, pdf>

IEICE Neurocomputing Technical meeting at Tokyo Institute of Technology.:"Learning the Data Region using Extreme-value Statistics"<ps, pdf>

"Upper Bounds of Bayesian Generalization Errors in Reduced Rank Regression"IEICE Trans., Vol.J86-A,No.3,pp.278-287,2003

The English abstract of this paper is here.

IBIS2002(Fuji-Yoshida,2002.09.08-10)

"Learning Curves of Reduced Rank Regression in Bayesian Estimation"<ps, pdf>

Attending the meeting of IEICE Neurocomputing Technical Group at Tamagawa Univ.(2002.3.18)

"Bayes Generalization Errors of Reduced Rank Approximation"<ps,pdf>.

power point:about learning theory and its application(in Japanese)

seminar notes(ps,pdf ):cited from "Learning Machines and Algorithms"(Kyoritu Shuppan)p.42〜p.46 written by Sumio Watanabe

write to me please!!