NeurIPS 2020

Bayesian Bits: Unifying Quantization and Pruning

Meta Review

After a discussion with the reviewers, I converged towards recommending to accept this submission. The reviewers have mixed opinions about this paper. While two reviewers suggest accept and two suggest reject, one of the accept and one of the reject reviewers did not engage in the discussion and did not defend their positions given the authors' response. Reducing the weight of these two reviews in the final decision, this leaves two reviewers, one suggesting to accept and the other to reject. R3 sees the presented solution as an elegant way to solve the issue regarding the computational cost for the residual tensors. At the same time, R1 feels that it is not possible to confirm the promise of the proposed method given the lack of theoretical analysis and the weak empirical results. They further feel that the rebuttal did not address their concerns, and that a resubmission addressing the criticism is appropriate.