NeurIPS 2020

FracTrain: Fractionally Squeezing Bit Savings Both Temporally and Spatially for Efficient DNN Training


Meta Review

All three referees support accept. The authors did a good job clarifying the concerns of the reviewers in the rebuttal period. AC also thinks that the experimental results are solid and the idea is worth to share to the community. A potential weakness is that dynamic precision training on bit-parallel hardware is hard, and AC suggests the authors to provide some comments on challenging hardware design aspects in the final draft. This would make the paper stronger and more useful in practice. Irrespectively, AC recommend acceptance.