NeurIPS 2020

Improving Auto-Augment via Augmentation-Wise Weight Sharing

Meta Review

The simple and elegant insight into framing data augmentation as a bilevel optimization problem as is often done in NAS and observing that later stages result in the most consistent gains and stable rankings thus allowing the design of a fast but accurate proxy task for the inner loop of bilevel opt. was well appreciated. Thorough experiments and clear writing also makes it stand out. Authors are encouraged to take all reviewer feedback to further improve the paper for publication.