Organization Contact Person Ryohei Nakano Affiliation NTT Communication Science Labs. Title Senior Research Scientist, Supervisor Research Area Adaptive Computation, Neural Networks Last Graduated & Year Graduated University of Tokyo, 1971 Degree Doctor of Engineering Date of Birth 1947 Research Partner Kazumi Saito Affiliation NTT R&D Management Department Title Associate Manager Research Area Neural Networks, Machine Learning Last Graduated & Year Graduated Keio University, 1985 Degree Doctor of Engineering Year of Birth 1963 Project Name Discovery of Nonlinear Rules from Numerical Database Objective In many numeric data of multiple variates, there may exist nonlinear relationship (rules) among variates. Discovering such nonlinear rules is an important research topic. As nonlinear rules, we assume polynomials with real exponents. A problem of finding such polynomials from numeric data can be formalized as neural networks learning. Since the learning is known quite difficult, we employ a very efficient 2nd-order supervised learning algorithm called BPQ. The tradeoff between the model complexity and the fitting accuracy will be solved by employing the MDL (Minimum Description Length) princple. We will investigate the computational characteristics (ex. scalability, noise tolerance) of the law discovery method using neural networks, and moreover will try to extend the method to include qualitative variates (nominal or ordinal ones) as well as quantitative ones. List of Selected Publications K.Saito and R.Nakano: A connectionist approach to numeric law discovery (in Japanese), Trans. of IPSJ, Vol.37, No.9, pp.1708-1716 (1996). K. Saito and R. Nakano: Partial BFGS update and efficient step-length calculation for three-layer neural networks, Neural Computation, Vol.9, No.1, pp.239--257 (1997). K. Saito and R. Nakano: Law discovery using neural networks, Proc. of 15th International Joint Conference on Artificial Intelligence (IJCAI '97), pp.1078--1083 (1997).