On the Minimal Supervision for Training Any Binary Classifier from Only Unlabeled Data