Adopting Two Supervisors for Efficient Use of Large-Scale Remote Deep Neural Networks