AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty

TL;DR

AugMix is a simple data augmentation method that enhances the robustness and uncertainty estimation of image classifiers against unforeseen corruptions, with minimal computational cost.

Contribution

This paper introduces AugMix, a novel data processing technique that significantly improves model robustness and uncertainty estimates under distribution shifts.

Findings

AugMix improves robustness on challenging image classification benchmarks.

AugMix enhances uncertainty estimation in neural networks.

AugMix reduces performance gap to optimal in some cases.

Abstract

Modern deep neural networks can achieve high accuracy when the training distribution and test distribution are identically distributed, but this assumption is frequently violated in practice. When the train and test distributions are mismatched, accuracy can plummet. Currently there are few techniques that improve robustness to unforeseen data shifts encountered during deployment. In this work, we propose a technique to improve the robustness and uncertainty estimates of image classifiers. We propose AugMix, a data processing technique that is simple to implement, adds limited computational overhead, and helps models withstand unforeseen corruptions. AugMix significantly improves robustness and uncertainty measures on challenging image classification benchmarks, closing the gap between previous methods and the best possible performance in some cases by more than half.