A new algorithm for fast discriminative training

Qi Li, Biing-Hwang Juang
2002 IEEE International Conference on Acoustics Speech and Signal Processing  
Currently, almost all discriminative training algorithms for nonlinear classifier design are based on gradient-descent methods, such as backpropagation and generalized probabilistic descent algorithms. Those algorithms are easy to derive and are effective in applications; however, a drawback for the gradient-descent approaches is the slow training speed, which limits their applications to large training problems, such as large vocabulary speech recognition and many other applications with time
more » ... equirements. On the other hand, some training algorithms, such as maximum likelihood estimation (MLE), are fast but they are not in discriminative training; therefore, the performances are not as good as the discriminative one. To address the problem, we proposed a fast discriminative training algorithm in this paper. It is a batch-mode algorithm derived from the objective function of minimal error rates. The significant theoretical advantage is its closed-form solution for parameter estimation during iterations. Thus, training problems can be solved in a few iterations. Experiments show that the proposed algorithm provides better performances than MLE.
doi:10.1109/icassp.2002.5743663 dblp:conf/icassp/LiJ02 fatcat:nb4wf5uqybbejovaqijpeeoxn4