Neural α-feature detector for feature detection and generalization

Abstract

We propose a neural /spl alpha/-feature detector used to extract a small number of main or essential features in input patterns. Features can be detected by controlling /spl alpha/-entropy for /spl alpha/-feature detectors. The /spl alpha/-entropy is defined by the difference between Renyi entropy and Shannon entropy. The /spl alpha/-entropy controller aims to maximize information contained in a few important /spl alpha/-feature detectors, while information for all other feature detectors is minimized. Thus, the /spl alpha/-entropy controller can maximize and simultaneously minimize information. The neural /spl alpha/-feature detector was applied to the inference of consonant cluster formation. Experimental results confirmed that by controlling /spl alpha/-entropy a small number of principal features can be detected, which can intuitively be interpreted. In addition, we could see that generalization performance is improved by minimizing /spl alpha/-entropy.