Toggle Main Menu Toggle Search

Open Access padlockePrints

Generalization issues in multiclass classification - new framework using mixture of experts

Lookup NU author(s): Srinivasan Meenakshi Sundaram, Dr Wai Lok Woo, Emeritus Professor Satnam Dlay

Downloads

Full text for this publication is not currently held within this repository. Alternative links are provided below where available.


Abstract

In this paper we introduce a new framework for expert systems used in real time speech applications. This consists of mixture of experts (MoEs) trained for multiclass classifications problems such as speech. We focus mainly on the generalization issues which are surprisingly ignored in established methods and demonstrate how severe these can be when the framework is drafted as a system. We limit this paper by addressing the issues, presenting the MoE capabilities to overcome and statistical perspective behind the training is briefly presented. Significant leap in the performance is achieved and justified by an impressive 10 % improvement on word recognition rate over the best available frameworks and an impressive 18.082 % over the baseline HMM. Critically the error rate is reduced by 10.61% over other connectionist models and 23.29 % over baseline HMM method. (15 References).


Publication metadata

Author(s): Meenakshisundaram S, Woo WL, Dlay SS

Publication type: Article

Publication status: Published

Journal: WSEAS Transactions on Information Science and Applications

Year: 2004

Volume: 1

Issue: 6

Pages: 1676-1681

Print publication date: 01/12/2004

ISSN (print): 1790-0832

Publisher: World Scientific and Engineering Academy and Society


Share