recurrent neural networks (RNNs) have recently achieved remarkable successes
in a number of applications. However, the huge sizes and computational burden
of these models make it difficult for their deployment on edge devices. A
practically effective approach is to reduce the overall s