Reviews: QSGD: Communication-Efficient SGD via Gradient Quantization and Encoding

Neural Information Processing Systems 

Update: I decrease slightly the grade due to the mismatch between theoretical and practical results that could be better covered. Still this paper has strong experimental results and some theoretical results. I would encourage the authors to improve on the gap between the two. In this paper the author introduce Quantized SGD (QSGD), a scheme for reducing the communication cost of SGD when performing distributed optimization. The quantization scheme is useful as soon as one has to transmit gradients between different machines.