| In this thesis we adopt a framework by combining model-based feature extraction with discriminative classifiers in supervised learning. We discuss the likelihood space as features extracted from an assumed probabilistic model and linear discriminative classifier as the classification method. Different from the conventional and popular practice on generative model-based methods in machine learning, where generative models serve as classifiers and Bayesian decision rule is adopted as the classification criterion, the generative models in our system is statistical feature generator and the classification can be carried out by any effective linear or nonlinear discriminant functions.;To demonstrate the effectiveness of the proposed method, First, we prove the method can achieve the same classification result as traditional Bayes classifier when the model assumption is correct. Then we show how the classification performance gets improved for the proposed method when the model assumption is mis-specified. Also we study the relationships between the classification error performance and sample size as well as model adequacy. Furthermore, we propose Robust Adjust Likelihood function to adjust the model parameters in the likelihood space when the assumed model is mis-specified. We explore the proposed method in semi-supervised learning and SAR image classification problems in practice. |
