Research map / Label-efficient and robust learning / OOD and anomaly detection
Mining In-distribution Attributes in Outliers 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 MVOL, an outlier-exposure OOD detection framework that mines in-distribution attributes in auxiliary outliers via an extended multi-view data model, using MaxLogit scoring and a soft-label multi-view learning objective.
Problem
Outlier-exposure OOD detection methods suppress logits on auxiliary outliers uniformly, ignoring that outliers often contain in-distribution attributes, which can degrade detection and waste useful signal.
Approach
Extends the multi-view data model so outliers consist of minor ID features plus noise; justifies MaxLogit as an interpretable and effective OOD score; and replaces the uniform confidence loss with a soft-label objective that adaptively aligns logits of categories with minor ID features rather than suppressing them blindly.
Key results
- In the ensemble distillation setting on CIFAR-100, MVOL reduces FPR95 to 36.62% versus 46.92% for OE + MaxLogit (about 10 points), with AUROC 90.37 vs 90.79.
- On CIFAR-10 single-model setting, MVOL achieves FPR95 3.30% and AUROC 98.70%, the best among compared methods.
- In the wild-dataset setting, MVOL's FPR95 advantage over OE + MaxLogit grows with noise level, reaching a 19.61-point reduction at alpha=0.5 in the single-model setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, single model · CIFAR-10 OOD | FPR95: 4.12{ \ $\pm$ 0.20} -> 3.30{ \ $\pm$ 0.19} | Helped |
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, single model · CIFAR-100 OOD | FPR95: 46.20{ \ $\pm$ 3.53} -> 42.96{ \ $\pm$ 0.86} | Helped |
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, ensemble distillation · CIFAR-10 OOD | AUROC: 98.62{ \ $\pm$ 0.06} -> 98.61{ \ $\pm$ 0.06} | No clear effect |
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, ensemble distillation · CIFAR-100 OOD | FPR95: 46.92{ \ $\pm$ 0.75} -> 36.62{ \ $\pm$ 1.36} | Helped |
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, ensemble distillation · CIFAR-100 OOD | AUROC: 90.79{ \ $\pm$ 0.26} -> 90.37{ \ $\pm$ 0.43} | Hurt |
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, single model · Wild dataset, α=0.5 | FPR95: 45.14{ \ $\pm$ 2.94} -> 25.53{ \ $\pm$ 1.23} | Helped |
| Use MVOL instead of OE + MaxLogit | Wide ResNet-40-2, ensemble distillation · Wild dataset, α=0.5 | FPR95: 54.23{ \ $\pm$ 3.77} -> 14.85{ \ $\pm$ 1.86} | Helped |
| Use ensemble distillation instead of single-model training | MVOL · CIFAR-100 OOD | FPR95: 42.96{ \ $\pm$ 0.86} -> 36.62{ \ $\pm$ 1.36} | Helped |
| Use ensemble distillation instead of single-model training | MVOL · CIFAR-10 OOD | FPR95 | No clear effect |
Surprising findings
- Models trained only on ID data produce unexpectedly high MaxLogit for certain outliers, and OOD data often carry ID attributes that prior OE methods suppress.
- WOODS, the customized wild-dataset baseline, degrades in the ensemble distillation setting and falls below plain MaxLogit.
Limitations
- Theoretical analysis relies on a simplified two-layer convolutional network and idealized assumptions (e.g., an ideal classifier with balanced feature sums).
- Experiments are limited to CIFAR-10/100 with Wide ResNet-40-2 and a single auxiliary dataset family.
- The method still requires auxiliary outlier data; the paper leaves ID-only calibration to future work.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 300K RandomImages SVHN Textures Places365 LSUN (crop) LSUN (resize) iSUN
Benchmarks: CIFAR-10 OOD detection CIFAR-100 OOD detection wild-dataset OOD detection (CIFAR-10 with CIFAR-100 auxiliary, noise alpha 0-0.5)
Metrics: FPR95 AUROC ID-Acc
Baselines: MSP Energy MaxLogit ODIN Mahalanobis KNN ASH VOS ATOL Energy w/Aux OE OE + MaxLogit
Builds on: Multi-view data model (Allen-Zhu & Li) Outlier Exposure MaxLogit Energy-based OOD detection WOODS
Closest papers by meaning
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · similarity 0.90)
- When and How Does In-Distribution Label Help Out-of-Distribution Detection? (ICML 2024 · similarity 0.90)
- Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.89)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.89)
- Learning to Shape In-distribution Feature Space for Out-of-distribution Detection (NeurIPS 2024 · similarity 0.89)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · 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.