Seminar on Computational Learning and Adaptation


  Law Discovery using Neural Networks

Kazumi Saito
NTT Communication Science Laboratories
2-4, Hikaridai, Seika-cho, Soraku-gun
Kyoto 619-0237 Japan

In this talk, I introduce RF6.2, a method for discovering empirical laws from nominal and numeric data in the form of nominally conditioned polynomials. A rule set of this type can be regarded as a numeric function, and we show that such a function can be closely approximated by a single three-layer neural network. RF6.2's basic strategy involves training single neural networks using a regularization method, then extracting a rule set through decomposition of the best trained neural network with respect to the nominal explanatory variables. Experiments using four data sets show that RF6.2 works well in discovering very succinct and interesting laws, even from data containing irrelevant variables and a small amount of noise.


Date: Thurs., Feb. 1

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule