NeurIPS 2020

Bad Global Minima Exist and SGD Can Reach Them

Meta Review

3 of 4 reviewers reached a consensus that this paper deserves acceptance to neurips. Two additional feedbacks: (1) one of the reviewer thinks "Explicit Regularization is Stronger than Implicit Bias" is too strong of a claim, and the meta-reviewers agree that it might be better to make it "Explicit Regularization can be Stronger than Implicit Bias" (2) Explicit regualrization may trade with the learning rate when there is a batchnormlization involved. The paper should clarify or experiment with it.