On the Importance of Gradients for Detecting Distributional Shifts in the Wild

By Rui Huang, Andrew Geng, Yixuan Li. Neural Information Processing Systems, 2021

498 citations recorded by the source corpus.

Detecting out-of-distribution (OOD) data has become a critical component in ensuring the safe deployment of machine learning models in the real world. Existing OOD detection approaches primarily rely on the output or feature space for deriving OOD scores, while largely overlooking information from the gradient space. In this paper, we present GradNorm, a simple and effective approach for detecting OOD inputs by utilizing information extracted from the gradient space. GradNorm directly employs the vector norm of gradients, backpropagated from the KL divergence between the softmax output and a uniform probability distribution. Our key idea is that the magnitude of gradients is higher for in-distribution (ID) data than that for OOD data, making it informative for OOD detection. GradNorm demonstrates superior performance, reducing the average FPR95 by up to 16.33% compared to the previous best method.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/3b316377f1d7cbcf3c907c4d8b08c05f4521b541

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

OpenOOD: Benchmarking Generalized Out-of-Distribution Detection (2022, background): “Later works explore other simple and more efﬁcient indicators to distinguish ID and OOD, such as ODIN [13] that uses temperature scaling with gradient-based input perturbations, MDS [18] that measures minimum mahalanobis distance from class centroids, EBO [14] that uses energy-based functions, GRAM ”
VOS: Learning What You Don't Know by Virtual Outlier Synthesis (2022, background): “, 2020a), Gram matrices based score (Sastry & Oore, 2020), and GradNorm score (Huang et al., 2021).”
ViM: Out-Of-Distribution with Virtual-logit Matching (2022, methodology): “The gradients are also used as evidence for ID and OOD distinction in [17].”
Out-of-distribution Detection with Deep Nearest Neighbors (2022, background): “The norm of the feature vector u and v could notably affect the value of the Euclidean distance. recent studies share the observation in Figure 4 (a) that the ID data has a larger L 2 feature norm than OOD data (Tack et al., 2020; Huang et al., 2021).”
Is Out-of-Distribution Detection Learnable? (2022, methodology): “Additionally, researchers also consider using the information contained in gradients to help improve the performance of OOD detection [18].”
ReAct: Out-of-distribution Detection With Rectified Activations (2021, background): “Indeed, much of the prior work on OOD detection focused on defining more suitable measures of OOD uncertainty [17, 19, 28, 29, 31, 33].”
Mitigating Neural Network Overconfidence with Logit Normalization (2022, methodology): “Beyond MSP, we show that our method not only out-performs, but also boosts more advanced post-hoc OOD scoring functions, such as ODIN (Liang et al., 2018), energy score (Liu et al., 2020), and GradNorm score (Huang et al., 2021).”