Research Activities

Statistical Study of Artificial Neural Networks

I'm studying artificial neural networks from the standpoint of multivariate data analysis. In particular, I'm interested in automatic selection of network structures and architectures using statistical techniques.

Generative Models for Vision

Generative models are stochastic models about the generation of given data from hidden causes. Estimation algorithms for the parameters of the models are regarded as unsupervised learning algorithms for the internal representation of the data. Using image data, such models can learn visual internal representation fit for the statistical characteristic of the images. I am searching for the minimal generative model to emulate the human vision system.

Articles

A. Utsugi (2002) "Independent Components of Natural Images under Variable Compression Rate", Neurocomputing, vol.49, no. 1-4, pp. 175-185.
A. Utsugi (2001) "Ensemble of Independent Factor Analyzers with Application to Natural Image Analysis", Neural Processing Letters, vol.14, no.1, pp. 49-60.

Mixture of Factor Analyzers

The Mixture of Factor Analyzers (MFA) is one of unsupervised learning models and a straightforward fusion of the mixture-of-Gaussian model and the factor-analysis model. MFA represents a hidden structure in data by a set of local linear manifolds. Since such a framework of hidden structures is very flexible and expansive, it is expected to be the core of adaptive cognition systems.

Articles

A. Utsugi and T. Kumagai (2001) "Bayesian Analysis of Mixtures of Factor Analyzers", Neural Computation, vol. 13, no. 5, pp. 993-1002.

Bayesian Self-Organizing Map

Self-Organizing Map (SOM) is a unique neural network model which has the ability to extract topological structures hidden in environmental data. Such discovered structures may be useful for the improvement of systems for prediction, classification and control. In reality, SOM is too flexible and fits to any data distributions whichever topology it postulates. I gave this over-flexible model a stiff property and an ability to control this stiffness by itself on a statistical principle. This results in a convenient data-analysis tool without cumbersome trial-and-error tuning.

Demonstrations using java applets and GIF-animations

Articles

A. Utsugi (2000) "Bayesian sampling and ensemble learning in generative topographic mapping", Neural Processing Letters, vol. 12, no. 3, pp. 277-290.
A. Utsugi (1998) "Density estimation by mixture models with smoothing priors", Neural Computation, vol. 10, no. 8, pp. 2115-2135.
A. Utsugi (1997) "Hyperparameter selection for self-organizing maps", Neural Computation, vol. 9, no. 3, pp. 623-635.
(reprinted in "Self-Organizing Map Formation, Foundations of Neural Computation", eds. K. Obermayer and T. J. Sejnowski (2001), MIT Press, pp.277-289.)
A. Utsugi (1996) "Topology selection for self-organizing maps", Network: Computation in Neural Systems, vol. 7, pp. 727-740.
A. Utsugi (1994) "Lateral interaction in Bayesian self-organizing maps", Trans. IEICE D-II, vol. J77-D-II, pp. 1329-1336. (in Japanese)
A. Utsugi (1993) "A Bayesian model of topology-preserving map learning", Trans. IEICE D-II, vol. J76-D-II, pp. 1232-1239. (in Japanese)

Previous Studies

I studied mathematical models for various application areas including human-computer interface, image sensing and emotional decision making.

Articles

M. Kitajima, A. Utsugi (1993) "Validation of the fuzzy model of attraction emotions", Japanese Journal of Fuzzy Theory and Systems, vol. 5, pp. 55-67.
A. Utsugi, M. Ishikawa (1991) "Construction of Inner Space Representation of Latticed Network Circuits by Learning", Neural Networks, vol. 4, pp. 81-87.
A. Utsugi, M. Ishikawa (1991) "Learning of Linear Associative Mapping by Latticed Network Circuits", Systems and Computers in Japan, vol. 22, pp. 56-65.

Home page