Learning from imperfect data