Controlling 
-entropy with a Neural -Feature Detector
Ryotaro Kamimura
Information Science Laboratory,
Tokai University,
1117 Kitakaname Hiratsuka,
Kanagawa 259-12, Japan
Abstract
A neural -feature detector is proposed and used to extract a small number of main or essential features in input patterns. Features can be detected by controlling -entropy for -feature detectors. The -entropy is defined by the difference between Rényi entropy and Shannon entropy. The -entropy controller aims at maximizing information contained in a few important -feature detectors, while information for all other feature detectors is minimized. Thus, the -entropy controller can maximize and minimize information, depending on situations. The neural -feature detector is applied to four problems: an artificial F-H detection problem, recognition of six alphabet characters, the feature detection of 26 alphabet characters, and the inference of consonant cluster formation. Experimental results confirm that by controlling -entropy a small number of principal features can be detected, which can be interpreted intuitively. In addition, it is shown that generalization performance is improved by minimizing -entropy.
