Efficient Neural Network Training via Forward and Backward Propagation Sparsification