Communication-Efficient Distributed Deep Learning with Merged Gradient Sparsification on GPUs | IEEE Conference Publication | IEEE Xplore