Shigeru Katagiri, Biing-Hwang Juang, Alain Biem
DISCRIMINATIVE FEATURE EXTRACTION
Abstract:Pattern recognition consists of two main stages: feature extraction
and classification. Needless to say, these two constituent processes should
be designed systematically in a manner consistent with accurate recognition.
However, such consistency has not yet been achieved in pattern
recognition methods up to now. We thus propose in this paper a novel
solution to this important long standing problem. The proposed method
is mainly based on a recent discriminative learning theory, the Minimum
Classification Error formalization and the Generalized Probabilistic Descent
method, and referred to as Discriminative Feature Extraction. A
key idea of Discriminative Feature Extraction is to embed both procedures
of feature extraction and classification in a smooth functional form and
consistently design both stages so as to reduce the number of misclassifications.
An application of the method to speech recognition clearly shows
the great promise of this new approach.