Be More with Less: Scaling Deep-learning with Minimal Supervision