Scaling for edge inference of deep neural networks