Research map / Label-efficient and robust learning / OOD and anomaly detection
Out-of-Distribution Detection with Deep Nearest Neighbors
Yiyou Sun, Yifei Ming, Xiaojin Zhu, Yixuan Li
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 |
|---|---|---|---|
| Replace Mahalanobis scoring with nearest-neighbor distance | CIFAR-10 KNN with cross-entropy embeddings · CIFAR-10, average across OOD datasets | FPR95: 37.94 -> 29.15 | Improved |
| Use nearest-neighbor scoring instead of SSD+'s Mahalanobis scoring | CIFAR-10 SupCon model · CIFAR-10, average across OOD datasets | FPR95: 16.52 -> 11.07 | Improved |
| Train the KNN detector with SupCon rather than cross-entropy embeddings (ablation) | CIFAR-10 KNN · CIFAR-10 OOD datasets | FPR95 | Improved |
| Use 1% rather than 100% of ID training data for neighbor search (ablation) | ImageNet-1k KNN+ · ImageNet-1k, average across OOD datasets | FPR95: 38.47 -> 39.18 | Hurt |
| Normalize penultimate-layer features before calculating neighbor distances (ablation) | ImageNet KNN · ImageNet-1k OOD datasets | FPR95: 61.05% | Improved |
| Use penultimate-layer features rather than projection-head features (ablation) | ImageNet KNN+ · ImageNet-1k OOD datasets | FPR95 | Improved |
| Use the k-th-neighbor distance rather than the averaged k-neighbor distance (ablation) | ImageNet KNN+ · ImageNet-1k OOD datasets | FPR95 | No clear effect |
| Apply activation rectification (ReAct) (ablation) | ImageNet KNN+ · ImageNet-1k, average across OOD datasets | FPR95: 38.47 -> 26.45 | Improved |
| Replace SSD+ scoring with nearest-neighbor scoring | CIFAR-10 SupCon model · LSUN-FIX, ImageNet-FIX, ImageNet-R, and CIFAR-100 OOD | FPR95: 29.86 -> 21.52; 32.26 -> 25.92; 45.62 -> 29.92; 45.50 -> 38.83 | Improved |
| Replace Mahalanobis scoring with nearest-neighbor scoring | ViT-B/16 fine-tuned on ImageNet-1k · iNaturalist, SUN, Places, and Textures | FPR95: 17.56 -> 7.30; 80.51 -> 48.40; 84.12 -> 56.46; 70.51 -> 39.91 | Improved |
Closest papers by meaning
- How to Overcome Curse-of-Dimensionality for Out-of-Distribution Detection? (AAAI 2024 · similarity 0.91)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.90)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.90)
- Mahalanobis++: Improving OOD Detection via Feature Normalization (ICML 2025 · similarity 0.90)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.90)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.89)
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.
