Augmenting supervised neural classifier training using a corpus of unlabeled data

In recent years, there has been growing interest in applying techniques that incorporate knowledge from unlabeled data into systems performing supervised learning. However, disparate results have been presented in the literature, and there is no general consensus that the use of unlabeled examples should always improve classifier performance. This paper proposes a method for incorporating a corpus of unlabeled examples into the supervised training of a neural network classifier and presents results from applying the technique to several datasets from the UCI repository. While the results do not provide support for the claim that unlabeled data can improve overall classification accuracy, a bias-variance decomposition shows that classifiers trained with unlabeled data display lower bias and higher variance than classifiers trained using labeled data alone.