Are Deep Neural Networks "Robust"?

TL;DR

This paper discusses the robustness of deep neural networks in computer vision, highlighting that they do not exhibit robustness as traditionally defined in separating outliers from inliers.

Contribution

It clarifies the distinction between deep neural networks and traditional robust estimators, emphasizing their lack of robustness by classical standards.

Findings

Deep neural networks are not robust by traditional outlier-inlier separation standards.

They differ fundamentally from classical robust estimators in their handling of outliers.

The paper clarifies misconceptions about neural network robustness.

Abstract

Separating outliers from inliers is the definition of robustness in computer vision. This essay delineates how deep neural networks are different than typical robust estimators. Deep neural networks not robust by this traditional definition.