Theory and algorithms for semi-supervised learning

点击翻译按钮获取中文摘要

The investigator studies semi-supervised learning from a decision theoretical point of view. The research shows that in the Bayesian framework, unlabeled data should be used to construct a prior for the purpose of improving predictive learning. More generally, the investigator considers the problem of constructing priors and learning predictive structures on hypothesis spaces from unlabeled data. Under this unified framework, the investigator systematically studies theoretical and algorithmic consequences of semi-supervised learning. Statistical machine learning is concerned with building computer systems that can predict unobserved information (labels) based on observed information (data). For example, to predict whether a patient has cancer (label) based on blood test (data). Traditionally, a statistical machine learning algorithm builds prediction rules from a set of labeled data. One of the most important issues in practical applications of statistical machine learning is whether one can improve the performance of a learning algorithm by using unlabeled data. This is because unlabeled data are generally abundant while their labels are very costly to obtain. Methods that use both labeled and unlabeled data are generally referred to as semi-supervised learning. This research attempts to establish a general statistical theory for semi-supervised learning, and applies the theory to improve state of the art machine learning algorithms.