Could be for principled knowledge distillation. No more worries about training doodle monkey models.