Presentation	2009/12/10 How Should We Confront Pattern Recognition Tasks? : Wandering between Generalized Probabilistic Descent Method and Minimum Classification Error Training Shigeru KATAGIRI,
PDF Download Page	PDF download Page Link
Abstract(in Japanese)	(See Japanese page)
Abstract(in English)	Twenty years ago, we were involved in the development of a versatile discriminative training method for pattern recognition, called the Generalized Probabilistic Descent method (GPD). The most fundamental concept of GPD formalism was to formalize a given task in a functional form that can be optimized mathematically. Along this concept, GPD was applied to several different types of pattern recognition, such as continuous speech recognition, key-word spotting, and discriminative feature extractor design, and it contributed to performance improvement for those recognition tasks to a certain extent. However, in parallel with such a spread, research concern rapidly shifted from the principle of formalizing a given task entirely to a constituent element of GPD, i.e., an employment of smooth classification error count loss. The concern shift to the minimization of this classification error count loss is natural because of its consistency with the ideal Bayes error status. Accordingly, this minimization framework was named Minimum Classification Error training (MCE) and it gradually became more popular than GPD. MCE actually shares the same design goal, i.e., the achievement of the Bayes error status, with other recent discriminative training disciplines such as Support Vector Machine and Boosting. As a result, research on key components of discriminative training, such as loss section and margin control, has been attracting a great deal of interest. This leads us to ask: can this current approach of drilling down the components lead to a fundamental solution to pattern recognition problems? The current research may focus too much on the finer points of the problems. Probably, it is worth revisiting the alternative approach to the formalization of the entire task, which was attempted in the GPD framework. In this talk, through our experiences of wandering between GPD and MCE, we analyze the background of the transition in pattern recognition research interests and make a suggestion about a future direction.
Keyword(in Japanese)	(See Japanese page)
Keyword(in English)
Paper #	PRMU2009-14
Date of Issue

Conference Information
Committee	PRMU
Conference Date	2009/12/10(1days)
Place (in Japanese)	(See Japanese page)
Place (in English)
Topics (in Japanese)	(See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To	Pattern Recognition and Media Understanding (PRMU)
Language	JPN
Title (in Japanese)	(See Japanese page)
Sub Title (in Japanese)	(See Japanese page)
Title (in English)	How Should We Confront Pattern Recognition Tasks? : Wandering between Generalized Probabilistic Descent Method and Minimum Classification Error Training
Sub Title (in English)
Keyword(1)
1st Author's Name	Shigeru KATAGIRI
1st Author's Affiliation	Faculty of Science and Engineering, Doshisha University()
Date	2009/12/10
Paper #	PRMU2009-14
Volume (vol)	vol.109
Number (no)	344
Page	pp.pp.-
#Pages	26
Date of Issue