Improving the Generalization Ability of MCE/GPD Learning and its Application to Multistage Building Learning

The Minimum Classification Error (MCE) / Generalized Probabilistic Descent (GPD) learning proposed by Katagiri and Juang in 1992 has attracted a great deal of attention for its high recognition performance and wide range of applications including the case where the length of feature vectors is variable like speech recognition. In this report, we propose a new method to improve the generalization performance of the MCE learning by employing an regularization technique which is widely used to solve ill-posed problems. Feed-forward neural networks are employed to evaluate the performance of the proposed method.