Active Learning with Clustering and Unsupervised Feature Learning

Active learning is a type of semi-supervised learning in which the training algorithm is able to obtain the labels of a small portion of the unlabeled dataset by interacting with an external source (e.g. a human annotator). One strategy employed in active learning is based on the exploration of the cluster structure in the data, by using the labels of a few representative samples in the classification of the remaining points. In this paper we show that unsupervised feature learning can improve the "purity" of clusters found, and how this can be combined with a simple but effective active learning strategy. The proposed method shows state-of-the art performance in MNIST digit recognition in the semi-supervised setting.