NeurIPS 2020

Measuring Robustness to Natural Distribution Shifts in Image Classification

Meta Review

This paper is a meta-analysis that explores the combinations (1) network architectures, (2) various (adversarial) robust training methods and (3) variants of ImageNet dataset with the goal of quantifying how robust these networks are to various perturbations. All reviewers appreciated the comprehensiveness of the study and belielve that this paper will serve as a starting point / benchmark of studying robustness of image classification models.