Research map / Label-efficient and robust learning / OOD and anomaly detection
Data Invariants to Understand Unsupervised Out-of-Distribution Detection
Lars Doorenbos, Raphael Sznitman, Pablo Márquez-Neila
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MahaAD instead of MSCL (best prior method) as the detector | U-OOD detectors · U-OOD benchmark, task mean | AUC: 84.1 -> 86.5 | Improved |
| swap ResNet-101 for EfficientNet-b4 as MahaAD backbone (ablation) | MahaAD · U-OOD benchmark, task mean | AUC: 86.5 -> 89.3 | Improved |
| swap ResNet-101 for EfficientNet-b4 as MahaAD backbone (ablation) | MahaAD · shift-high-res | AUC: 78.6 -> 83.8 | Improved |
| swap ResNet-101 for EfficientNet-b4 as MahaAD backbone (ablation) | MahaAD · uni-ano | AUC: 91.3 -> 94.4 | Improved |
| use MahaAD instead of SSD on low-resolution domain shift | U-OOD detectors · shift-low-res (CIFAR10:SVHN) | AUC: 99.6 vs 94.3 | Hurt |
| use MahaAD instead of CFlow on anomaly detection | U-OOD detectors · uni-ano (MVTec) | AUC: 95.7 vs 91.3 | Hurt |
| use MahaAD instead of MSCL on one-vs-rest CIFAR | U-OOD detectors · uni-class | AUC: 96.3 vs 92.4 | Hurt |
| use only the lowest-variance principal components (invariants) in the Mahalanobis score (ablation) | MahaAD · four U-OOD experiments, ResNet50 | AUC: converges at about 3% of variance explained vs over 80% for high-variance components | Improved |
| use MahaAD instead of training-based detectors for training time | MahaAD · CIFAR10:SVHN training time | training time: about 90 seconds vs more than 12 hours for SSD | Improved |
| increase number of CIFAR10 training classes, effect on SVHN/MNIST detection (ablation) | MahaAD · CIFAR10 multi-class training | AUC | No clear effect |
| increase number of CIFAR10 training classes, effect on SVHN/MNIST detection (ablation) | DN2 and MSCL · CIFAR10 multi-class training | AUC | Hurt |
| add shrinkage to covariance estimate (ablation) | MahaAD · U-OOD benchmark | AUC | No clear effect |
Closest papers by meaning
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.88)
- Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection (AAAI 2024 · similarity 0.88)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.88)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.87)
- MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse (NeurIPS 2026 · similarity 0.87)
- Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
