On semi-supervised learning of Gaussian mixture models for phonetic classification
Jui-Ting Huang, Mark Hasegawa‐Johnson · 2009
This paper investigates semi-supervised learning of Gaussian mixture models using an unified objective function taking both labeled and unlabeled data into account. Two methods are compared in this work - the hybrid discriminative/generative method and the purely generative method. They differ in the criterion type on labeled data; the hybrid method uses the class posterior probabilities and the purely generative method uses the data likelihood. We conducted experiments on the TIMIT database and a standard synthetic data set from UCI Machine Learning repository. The results show that the two methods behave similarly in various conditions. For both methods, unlabeled data improve training on models of higher complexity in which the supervised method performs poorly. In addition, there is a trend that more unlabeled data results in more improvement in classification accuracy over the supervised model. We also provided experimental observations on the relative weights of labeled and unlabeled parts of the training objective and suggested a critical value which could be useful for selecting a good weighing factor.