Revisiting Batch Normalization for Improving Corruption Robustness

TL;DR

This paper proposes a simple method to improve the robustness of deep neural networks against common corruptions by adapting batch normalization statistics without retraining, significantly enhancing performance on benchmark datasets.

Contribution

The authors introduce a novel approach of rectifying batch normalization statistics to enhance corruption robustness without retraining the entire model.

Findings

Adapting BN statistics improves ImageNet-C top-1 accuracy from 39.2% to 48.7%.

The method boosts robustness of state-of-the-art models from 58.1% to 63.3%.

Simple BN statistics adaptation yields large robustness gains across multiple architectures.

Abstract

The performance of DNNs trained on clean images has been shown to decrease when the test images have common corruptions. In this work, we interpret corruption robustness as a domain shift and propose to rectify batch normalization (BN) statistics for improving model robustness. This is motivated by perceiving the shift from the clean domain to the corruption domain as a style shift that is represented by the BN statistics. We find that simply estimating and adapting the BN statistics on a few (32 for instance) representation samples, without retraining the model, improves the corruption robustness by a large margin on several benchmark datasets with a wide range of model architectures. For example, on ImageNet-C, statistics adaptation improves the top1 accuracy of ResNet50 from 39.2% to 48.7%. Moreover, we find that this technique can further improve state-of-the-art robust models from 58.1% to 63.3%.