A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection
By Jie Jessie Ren, Stanislav Fort, J. Liu, Abhijit Guha Roy, Shreyas Padhy, Balaji Lakshminarayanan. 2021
310 citations recorded by the source corpus.
Mahalanobis distance (MD) is a simple and popular post-processing method for detecting out-of-distribution (OOD) inputs in neural networks. We analyze its failure modes for near-OOD detection and propose a simple fix called relative Mahalanobis distance (RMD) which improves performance and is more robust to hyperparameter choice. On a wide selection of challenging vision, language, and biology OOD benchmarks (CIFAR-100 vs CIFAR-10, CLINC OOD intent detection, Genomics OOD), we show that RMD meaningfully improves upon MD performance (by up to 15% AUROC on genomics OOD).
Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Anomaly & Outlier Detection
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
- Ziwei Liu
- Luc Van Gool
- Chen Change Loy
- Xiaogang Wang
- Ming-Hsuan Yang
- Yu Qiao
- Dacheng Tao
- Trevor Darrell
In Anomaly & Outlier Detection
Citations in the corpus
- Adversarial vulnerability of powerful near out-of-distribution detection (2022, methodology): “[15], generally improving performance while being more agnostic to hyperparameter choice.”
- Generalized Out-of-Distribution Detection: A Survey (2021, unknown): “A subsequent work splits the images into foreground and background and then calculates the Mahalanobis distance ratio between the two spaces (Ren et al., 2021).”
- OpenOOD: Benchmarking Generalized Out-of-Distribution Detection (2022, methodology): “Distance-based methods usually compute distance in the high-dimension space such as feature space and gradient space to distinguish ID and OOD [18, 19].”