NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2810
Title:E2-Train: Training State-of-the-art CNNs with Over 80% Less Energy

This paper addresses an important topic of energy-efficient training of CNNs. They investigate this in three different levels and report promising results on hardware testbed. The reviewers did raise a number of critical questions, especially on theoretical understanding of *why* the proposed bag of tricks work and if good results can be easily generalized and reproduced. The rebuttal provided informative and convincing to the reviewers comments and promise to an number of improvement. We recommend accept the paper, but strongly encourage the authors to carefully address the reviewers' concerns. We also strongly encourage the authors to make their code publicly available; we think this is especially critical given that this work is mainly empirically and the results can not be verified or reproduced without code available.