Efficient and compact neural network models are essential for enabling the
deployment on mobile and embedded devices. In this work, we point out that
typical design metrics for gauging the efficiency of neural network
architectures -- total number of operations and parameters -- are no