Improved Quasi-Supervised Learning by Expectation-Maximization

Abstract

In this paper, a new statistical learning method was developed that implements the quasi-supervised learning method in an expectation-maximization loop. First, automatic strategies were generated that separated the samples drawn from different distributions into respective sample sets using the posterior probabilities computed via quasi-supervised learning based on partially separated samples. An expectation-maximization loop was then constructed by combining this procedure with the posterior probability computation step using the new separated sample sets. In controlled experiments on recognition problems with varying difficulties, the proposed method was observed to consistently outperform the plain quasi-supervised learning method.