We use ensembles of highly sensitive to highly precise models and show that their combination can i) be used to learn from unlabeled data, ii) used for active learning procedures and iii) support human-machine interaction.