Kagan Tumer's Publications

Display Publications by [Year] [Type] [Topic]

A Framework for Estimating Performance Improvements in Hybrid Pattern Classifiers. K. Tumer and J. Ghosh. In Proceedings of the World Congress on Neural Networks, pp. III:220–225, INNS Press, San Diego, June 1994.

Abstract

Classification methods often perform significantly below Bayesian limits in complex, high-dimensional classification tasks because of model bias, inadequate training data and noise/variability in the data. When several classifiers are used for a given task, selecting one method over all others discards potentially valuable information. Strategies aimed at suitably combining the results of multiple classifiers are expected to perform better than any single method, and reduce overall bias and noise. An underwater passive sonar data set consisting of over 1000 samples processed to produce different 25-dimensional and 24-dimensional feature vectors is used in this study to examine an evidence combination framework. An analysis of the conditions that the data sets must satisfy, and the conditions under which improvements can be obtained is provided, and the results are presented for hybrid networks using both local and global classifiers.

Download

(unavailable)

BibTeX Entry

@inproceedings{tumer-ghosh_wcnn94,
        author={K. Tumer and J. Ghosh},
        title="A Framework for Estimating Performance Improvements in
                Hybrid Pattern Classifiers",
        booktitle="Proceedings of the World Congress on Neural Networks",
        publisher="{INNS} Press",
        pages ={III:220-225},
        address="San Diego",
	month={June},
	abstract={Classification methods often perform significantly below Bayesian limits in complex, high-dimensional classification tasks because of model bias, inadequate training data and noise/variability in the data. When several classifiers are used for a given task, selecting one method over all others discards potentially valuable information. Strategies aimed at suitably combining the results of multiple classifiers are expected to perform better than any single method, and reduce overall bias and noise. An underwater passive sonar data set consisting of over 1000 samples processed to produce different 25-dimensional and 24-dimensional feature vectors is used in this study to examine an evidence combination framework. An analysis of the conditions that the data sets must satisfy, and the conditions under which improvements can be obtained is provided, and the results are presented for hybrid networks using both local and global classifiers.},
	bib2html_pubtype = {Refereed Conference Papers},
	bib2html_rescat = {Classifier Ensembles},
        year={1994}
}

Generated by bib2html.pl (written by Patrick Riley ) on Wed Apr 01, 2020 17:39:43