CirCNN: Accelerating and Compressing Deep Neural Networks Using Block-CirculantWeight Matrices