Research map / Label-efficient and robust learning / OOD and anomaly detection
DOS: Diverse Outlier Sampling for Out-of-Distribution Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DOS, a diverse outlier sampling strategy that clusters normalized features each iteration and picks the most informative outlier per cluster for outlier-exposure training with an absent-category loss.
Problem
Outlier exposure for OOD detection relies on sampling auxiliary outliers; greedy uncertainty-based sampling yields outliers biased toward certain regions, failing to capture the full auxiliary outlier distribution and producing a locally compact decision boundary.
Approach
Cluster normalized penultimate features of candidate outliers with K-means at each iteration and select the single highest-scoring (lowest absent-category probability) outlier from each cluster, so selected outliers are both diverse and hard negatives, trained with an absent-category (K+1) cross-entropy loss.
Key results
- On CIFAR-100 with TI-300K, DOS reduces average FPR95 from 50.15% (NTOM) to 24.36%, a 25.79% improvement.
- On ImageNet-10 vs ImageNet-990, DOS achieves average FPR95 of 20.71%, surpassing OE by 3.07 points.
- With CLIP ViT-B/16 fine-tuning on IN-10/IN-990, DOS achieves average FPR95 of 3.41%, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DOS instead of greedy NTOM sampling | CIFAR-100-TI-300K OOD detection · CIFAR-100-TI-300K, OOD average | FPR95: 50.15% to 24.36% | Helped |
| Use DOS instead of Share random sampling | CIFAR-100-INRC OOD detection · CIFAR-100-INRC, OOD average | FPR95: 4.4% | Helped |
| Use DOS instead of NTOM greedy sampling | CIFAR-100-INRC OOD detection · CIFAR-100-INRC, OOD average | FPR95: 4.8% | Helped |
| Use DOS instead of POEM | CIFAR-100-INRC OOD detection · CIFAR-100-INRC, OOD average | FPR95: 3.7% | Helped |
| Use DOS instead of OE | ImageNet-10-IN990 OOD detection · ImageNet-10-IN990, OOD average | FPR95: 3.07% | Helped |
| Use DOS sampling instead of random under energy loss | CIFAR-100-TI-300K energy-loss training · CIFAR-100-TI-300K, energy loss | FPR95: 37.72% to 29.43%; 8.29% | Helped |
| Replace normalized features with softmax probabilities or raw latent features for clustering (ablation) | DOS · CIFAR-100-TI-300K | FPR95 | Hurt |
| Replace normalized features with PCA or whitening for clustering (ablation) | DOS · CIFAR-100-TI-300K | FPR95 | Hurt |
| Replace K-Means with agglomerative clustering (ablation) | DOS · ImageNet-10-IN990 | FPR95: 20.71 -> 38.39 | Hurt |
| Replace K-Means with K-Means++ seeding (ablation) | DOS · ImageNet-10-IN990 | FPR95: 20.71 -> 48.19 | Hurt |
| Replace K-Means with spherical K-Means (ablation) | DOS · ImageNet-10-IN990 | FPR95: 20.71 -> 24.11 | Hurt |
| Decrease the number of clustering centers (ablation) | DOS · ImageNet-10-IN990 | FPR95 | Hurt |
Surprising findings
- Greedy uncertainty-based sampling (NTOM) can perform worse than random sampling when paired with energy loss, with FPR95 rising to 84.68% vs 37.72% for random.
Limitations
- Clustering adds computational overhead (about 12% per iteration on CIFAR-100 with TI-300K).
- On CIFAR-100 with TI-300K, DOS is not best on every test set (e.g., Places FPR95 59.6 vs 53.2 for energy loss).
Experimental setup
Datasets: CIFAR-100 TI-300K ImageNet-RC ImageNet-1K ImageNet-10 ImageNet-990 SVHN LSUN-C LSUN-R DTD Places365 iSUN
Benchmarks: CIFAR-100 with TI-300K/ImageNet-RC OOD benchmark (SVHN, LSUN-C, DTD, Places, LSUN-R, iSUN) ImageNet-10 vs ImageNet-990 (iNaturalist, SUN, Places, Texture)
Metrics: FPR95 AUROC ACC Calinski-Harabasz index
Baselines: MSP ODIN Mahalanobis Energy score ReAct DICE OE Energy loss NTOM Share POEM ACET
Builds on: Outlier Exposure NTOM Energy regularization POEM ATOM
Closest papers by meaning
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · similarity 0.91)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.90)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.90)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.89)
- Mining In-distribution Attributes in Outliers for Out-of-distribution Detection (AAAI 2025 · similarity 0.89)
- ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.88)
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.