NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:8002
Title:PowerSGD: Practical Low-Rank Gradient Compression for Distributed Optimization


		
The authors propose a very interesting technique that is all-reduce compatible, for communication efficient learning. During the rebuttal phase the authors addressed most of the comments raised by the reviewers. The authors are strongly encouraged to address before the camera ready: - reporting end-to-end speedup of PowerSGD - adding a comparison of speedup curves comparing to various baselines on converging to certain accuracies.