Compressing Deep Neural Networks using a Rank-Constrained Topology

   Abstract