Rapid Discriminative Learning Based on Misclassification Measure

摘要

The current research is based on Minimum Classification Error Learning (MCE/GPD) using Generalized Probabilistic Descent (GPD), which is known as a high-performance discriminative learning method. MCE/GPD is an excellent recognition technique that has been applied to speech recognition because of its high recognition performance and its ability to deal with variable-length vectors. However, like other recognition techniques, it suffers from the problem that recognition performance drops for untrained data (generalization ability problem). There is also the practical fault that training time is lengthy due to the complexity of the algorithm. In the current research, the authors propose a new learning method that improves the generalization ability by introducing regularized learning to avoid ill-posed problems and increases learning speed according to a hierarchical model arrangement, which should solve these two problems. They used a hierarchical neural network for performance evaluation.