In many classi cation learning tasks, labeled data may be expensive or scarce. At the same time, unlabeled or \weakly labeled" samples, may be available in abundance. We consider three algorithmic p
