Unlabeled data: Now it helps, now it doesn't

T22

A. Singh, R. Nowak and X. Zhu ­ University of Wisconsin, Madison

Asymptotic analysis does not fully reveal distinctions between supervised and semi-supervised learning (SSL). Finite sample analysis and minimax bounds resolve apparent conflicts in recent papers.

Do unlabeled data help? If (1) the label process is smooth on certain subsets of the feature space, and (2) these subsets can be learned from the distribution of unlabeled features, then sometimes unlabeled data helps , and sometimes it doesn't .