Geoffrey Zweig and Mukund Padmanabhan. Boosting gaussian mixtures in an lvcsr system. In Proceedings of ICASSP 2000.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....of classes is larger than two, e.g. 3] This is the case in most speech applications, where the classes are, for example, vowels [4] or HMM states [5] While multiclass versions of most classification algorithms exist, e.g. 6] SVM) and [2] AdaBoost) they tend to be complex. For example, in [7, 8], a multiclass version of AdaBoost was applied to speech recognition, but the original algorithm had to be simplified due to a high computational cost. A more common approach is to construct the multiclass classifier by combining the outputs of several binary ones [3, 9] Typically, the ....

G. Zweig. Boosting gaussian mixtures in an LVCSR system. In ICASSP, pages 1527--30, 2000.

....the weights of hard to classify examples are increased relative to the easy ones. The outputs of the classifiers are then combined in such a way as to guarantee certain bounds on both training and testing error [12] We report results here using an extension to Adaboost that was presented in [11] and that allows for large speedups in training time. The extension was motivated by the scale of the problem, where we have tens of millions of labeled training pairs, thousands of classes, and hundreds of thousands of gaussians that model the probability density of the classes. A. Results The ....

G. Zweig and M. Padmanabhan, "Boosting gaussian mixtures in an LVCSR system", Proceedings of ICASSP, 2000.

....the weights of hard to classify examples are increased relative to the easy ones. The outputs of the classifiers are then combined in such a way as to guarantee certain bounds on both training and testing error [8] We report results here using an extension to AdaBoost that was presented in [6] and that allows for large speedups in training time. The extension was motivated by the scale of the problem, where we have tens of millions of labeled training pairs, thousands of classes, and and hundreds of thousands of gaussians that model the probability density of the classes. The input to ....

....a weight, C3Xa :W D , that is related to the probability with which can be misrecognized as W . This implies that the complete classifier has to be designed in one step during the next iteration using gradient descent techniques. This process was simplified by the approximation in [6], which allowed the classifier to be designed in two steps. The weights over all classes for a given feature vector were summed up 1 C3X D 5 bdc 1 C X .W D , and each feature vector now was associated with a single weight that is related to the probability of its having been ....

[Article contains additional citation context not shown here]

G. Zweig and M. Padmanabhan, "Boosting gaussian mixtures in an LVCSR system", Proceedings of ICASSP 2000.

....the weights of hard to classify examples are increased relative to the easy ones. The outputs of the classifiers are then combined in such a way as to guarantee certain bounds on both training and testing error [8] We report results here using an extension to AdaBoost that was presented in [6] and that allows for large speedups in training time. The extension was motivated by the scale of the problem, where we have tens of millions of labeled training pairs, thousands of classes, and and hundreds of thousands of gaussians that model the probability density of the classes. The input to ....

....a weight, u Z3av] g , that is related to the probability with which [ can be misrecognized as . This implies that the complete classifier has to be designed in one step during the next iteration using gradient descent techniques. This process was simplified by the approximation in [6], which allowed the classifier to be designed in two steps. The weights over all classes for a given feature vector were summed up d Z3a m wyx uzd Z a ] g , and each feature vector now was associated with a single weight that is related to the probability of its having been misclassified ....

[Article contains additional citation context not shown here]

G. Zweig and M. Padmanabhan, "Boosting gaussian mixtures in an LVCSR system", Proceedings of ICASSP 2000.

....the weights of hard to classify examples are increased relative to the easy ones. The outputs of the classifiers are then combined in such a way as to guarantee certain bounds on both training and testing error [8] We report results here using an extension to AdaBoost that was presented in [6] and that allows for large speedups in training time. The extension was motivated by the scale of the problem, where we have tens of millions of labeled training pairs, thousands of classes, and and hundreds of thousands of gaussians that model the probability density of the classes. The input to ....

....a weight, w 3cx : ia , that is related to the probability with which ] can be misrecognized as . This implies that the complete classifier has to be designed in one step during the next iteration using gradient descent techniques. This process was simplified by the approximation in [6], which allowed the classifier to be designed in two steps. The weights over all classes for a given feature vector were summed up f 3c a o y z w f c ia , and each feature vector now was associated with a single weight that is related to the probability of its having been misclassified ....

[Article contains additional citation context not shown here]

G. Zweig and M. Padmanabhan, "Boosting gaussian mixtures in an LVCSR system", Proceedings of ICASSP 2000.

....the weights of hard to classify examples are increased relative to the easy ones. The outputs of the classifiers are then combined in such a way as to guarantee certain bounds on both training and testing error [8] We report results here using an extension to AdaBoost that was presented in [6] and that allows for large speedups in training time. The extension was motivated by the scale of the problem, where we have tens of millions of labeled training pairs, thousands of classes, and and hundreds of thousands of gaussians that model the probability density of the classes. The input to ....

....is assigned a weight, D t (i; y) that is related to the probability with which x i can be misrecognized as y. This implies that the complete classifier has to be designed in one step during the next iteration using gradient descent techniques. This process was simplified by the approximation in [6], which allowed the classifier to be designed in two steps. The weights over all classes for a given feature vector were summed up D t (i) P y D t (i; y) and each feature vector now was associated with a single weight that is related to the probability of its having been misclassified during ....

[Article contains additional citation context not shown here]

G. Zweig and M. Padmanabhan, "Boosting gaussian mixtures in an LVCSR system", Proceedings of ICASSP 2000.

No context found.

Geoffrey Zweig and Mukund Padmanabhan. Boosting gaussian mixtures in an lvcsr system. In Proceedings of ICASSP 2000.

No context found.

G. Zweig and M. Padmanabhan, "Boosting gaussian mixtures in a lvcsr system," in Proc. ICASSP-00, Istambul, 2000, pp. 1527--1530.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC