【论文学习8】TernGrad: Ternary Gradients to Reduce Communication in Distributed Deep Learning