GRODIN: Improved Large-Scale Out-of-Domain detection via Back-propagation

Abstract

Uncertainty estimation and out-of-doman (OOD) input detection are critical for improving the safety and robustness of machine learning. Unfortunately, most methods for detecting OOD examples have been evaluated on small tasks while typical methods are computationally expensive. In this paper we propose a new gradient-based method called GRODIN for OOD detection. The proposed method is conceptually simple, computationally cheaper than ensemble methods and can be directly applied to any existing and deployed model without re-training. We evaluate GRODIN on models trained on CIFAR-10 and ImageNet datasets, and show it's strong performance on various OOD ImageNet datasets such as ImageNet-O, ImageNet-A, ImageNet-R, ImageNet-C.

Tasks

Reproductions