Towards Energy-Quality Scaling in Deep Neural Networks