Measuring Robustness to Natural Distribution Shifts in Image Classification

TL;DR

This paper evaluates how well current ImageNet models handle natural distribution shifts, revealing that most models lack robustness to real-world variations and highlighting the need for new approaches.

Contribution

The study provides a comprehensive benchmark of 204 models across 213 natural distribution shifts, showing limited transferability of synthetic robustness to real-world data.

Findings

Most models do not transfer robustness from synthetic to natural shifts.

Training on larger datasets can improve robustness but is insufficient.

Natural distribution shifts remain an open challenge in image classification.

Abstract

We study how robust current ImageNet models are to distribution shifts arising from natural variations in datasets. Most research on robustness focuses on synthetic image perturbations (noise, simulated weather artifacts, adversarial examples, etc.), which leaves open how robustness on synthetic distribution shift relates to distribution shift arising in real data. Informed by an evaluation of 204 ImageNet models in 213 different test conditions, we find that there is often little to no transfer of robustness from current synthetic to natural distribution shift. Moreover, most current techniques provide no robustness to the natural distribution shifts in our testbed. The main exception is training on larger and more diverse datasets, which in multiple cases increases robustness, but is still far from closing the performance gaps. Our results indicate that distribution shifts arising in real data are currently an open research problem. We provide our testbed and data as a resource for future work at https://modestyachts.github.io/imagenet-testbed/ .