Optimizing Boosting with Discriminative Criteria

We describe the use of discriminative criteria to optimize  Boosting based ensembles. Boosting algorithms may create  hundreds of individual classifiers in order to fit the training  data. However, this strategy isn’t feasible and necessary for  complex classification problems, such as real-time continuous  speech recognition, in which only the combination of a few of  acoustic models is practical. How to improve the classification  accuracy for small size of ensemble is the focus of this paper.  Two discriminative criteria that attempt to minimize the true  Bayes error rate are investigated. Improvements are observed  over a variety of datasets including image and speech  recognition, indicating the prospective utility of these two  criteria.