CONF misr03a/IDIAP Confusion Matrix Based Entropy Correction in Multi-stream Combination Misra, Hemant Morris, Andrew EXTERNAL http://publications.idiap.ch/attachments/reports/2003/misra_2003_eurospeech.pdf PUBLIC http://publications.idiap.ch/index.php/publications/showcite/morris-rr-02-53 Related documents Proceedings of ISCA European Conference on Speech Communication and Technology (Eurospeech) 2003 Geneve, Switzerland September 2003 IDIAP-RR 2002 53 An MLP classifier outputs a posterior probability for each class. With noisy data, classification becomes less certain, and the entropy of the posteriors distribution tends to increase providing a measure of classification confidence. However, at high noise levels, entropy can give a misleading indication of classification certainty. Very noisy data vectors may be classified systematically into classes which happen to be most noise-like and the resulting confusion matrix shows a dense column for each noise-like class. In this article we show how this pattern of misclassification in the confusion matrix can be used to derive a linear correction to the MLP posteriors estimate. We test the ability of this correction to reduce the problem of misleading confidence estimates and to enhance the performance of entropy based full-combination multi-stream approach. Better word-error-rates are achieved for Numbers95 database at different levels of added noise. The correction performs significantly better at high SNRs. REPORT morris-RR-02-53/IDIAP Confusion matrix based posterior probabilities correction Morris, Andrew Misra, Hemant classifier correction confidence measures confusion matrix MLP classifier EXTERNAL http://publications.idiap.ch/attachments/reports/2002/rr02-53.pdf PUBLIC Idiap-RR-53-2002 2002 IDIAP An MLP classifier outputs a posterior probability for each class. With noisy data classification becomes less certain and the entropy of the posteriors distribution tends to increase, therefore providing a measure of classification confidence. However, at high noise levels entropy can give a misleading indication of classification certainty because very noisy data vectors may be classified systematically into whichever classes happen to be most noise-like. When this happens the resulting confusion matrix shows a dense column for each noise-like class. In this article we show how this pattern of misclassification in the confusion matrix can be used to derive a linear correction to the MLP posteriors estimate. We test the ability of this correction to reduce the problem of misleading confidence estimates and to increase the performance of individual MLP classifiers. Word and frame level classification results are compared with baseline results for the Numbers95 database of free format telephone numbers, in different levels of added noise.