NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1457
Title:Asymmetric Valleys: Beyond Sharp and Flat Local Minima

This paper introduces the concept of asymmetric valleys (AVs), studies the generalization performance of the flat side, and provides an algorithm that is biased toward the flat side. Initially the paper received mixed reviews, with two positive and one negative reviews, the latter one arguing the need for more explicit motivation for AVs. The rebuttal successfully addressed a large number of questions raised by the reviewers, and the negative reviewer updated his/her score. Upon discussion, the reviewers agreed that the paper should be accepted. Overall, this paper addresses a very important and timely problem (understanding the relationships between landscape, optimization and generalization).