Paper titled The Information-Theoretic Value of Unlabeled Data in Semi-Supervised Learning by Alexander Golovnev, Dávid Pál and Balázs Szörényi accepted at ICML 2019.
The paper proves that unlabeled data beneficial for supervised learning tasks. We formalized the problem in the Probably Approximately Correct (PAC) model and we show that for learning projections over the Boolean hypercube {0,1}n one needs less labeled examples by a multiplicative factor Θ(log n) if one has access to unlabeled data.