Research map / Label-efficient and robust learning / OOD and anomaly detection
Out-Of-Distribution Detection with Diversification (Provably)
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 diverseMix, a mixup-based method that generates semantically novel, score-adaptive mixed auxiliary outliers to improve OOD detection with a fixed outlier set, backed by a generalization bound.
Problem
OOD detectors trained with auxiliary outliers generalize poorly to unseen OOD data because the auxiliary outlier set lacks semantic diversity, and collecting more diverse outliers is costly.
Approach
Theoretically links outlier diversity to a lower OOD generalization bound (via distribution-shift error), then enhances diversity of a fixed outlier set by mixing outliers with mixup weights adapted to their current OOD scores so that mixed outliers are novel to the model.
Key results
- On CIFAR-10, diverseMix reduces FPR95 to 1.92% (AUROC 99.42) versus best baseline POEM at 2.54%; on CIFAR-100 FPR95 is 8.51% versus 15.14% for POEM (about 43.8% relative reduction).
- On the ImageNet-200 benchmark, diverseMix achieves average FPR 53.07% and AUROC 89.36%, the best average among compared methods, including MixOE (57.68% FPR).
- Ablation on CIFAR-100 shows diverseMix (FPR 8.51) outperforms vanilla mixup (12.56) and cutmix (15.95).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DiverseMix to auxiliary-outlier training | OOD detector with DenseNet · CIFAR-10 | FPR95: reducing the FPR95 by 0.62% | Helped |
| add DiverseMix to auxiliary-outlier training | OOD detector with DenseNet · CIFAR-100 | FPR95: reducing the FPR95 by 6.63% | Helped |
| add DiverseMix to auxiliary-outlier training | OOD detector with ResNet · ImageNet-200 near-OOD | FPR: 59.12 vs. 59.81 | Mixed |
| add DiverseMix to auxiliary-outlier training | OOD detector with ResNet · ImageNet-200 far-OOD | FPR: 50.51 -> 48.58 | Helped |
| add DiverseMix to auxiliary-outlier training | OOD detector with ResNet · ImageNet-200 average | FPR: 56.61 -> 53.07 | Helped |
| replace DiverseMix augmentation with Gaussian noise (ablation) | DiverseMix OOD detector · CIFAR-10 | FPR: 1.92 -> 6.69 | Hurt |
| replace DiverseMix augmentation with Gaussian noise (ablation) | DiverseMix OOD detector · CIFAR-100 | FPR: 8.51 -> 19.94 | Hurt |
| replace DiverseMix augmentation with Cutout (ablation) | DiverseMix OOD detector · CIFAR-10 | FPR: 1.92 -> 7.20 | Hurt |
| replace DiverseMix augmentation with Cutout (ablation) | DiverseMix OOD detector · CIFAR-100 | FPR: 8.51 -> 19.12 | Hurt |
| replace DiverseMix augmentation with color jitter (ablation) | DiverseMix OOD detector · CIFAR-10 | FPR: 1.92 -> 8.83 | Hurt |
| replace DiverseMix augmentation with color jitter (ablation) | DiverseMix OOD detector · CIFAR-100 | FPR: 8.51 -> 24.36 | Hurt |
| add adaptive semantic interpolation rather than vanilla outlier training (ablation) | DiverseMix OOD detector · CIFAR-10 | FPR: 5.43 -> 1.92 | Helped |
Surprising findings
- Methods using auxiliary outliers can perform worse than outlier-free methods on far-OOD ImageNet test sets, e.g., OE and Energy (w. D_aux) underperform on some far-OOD metrics.
Limitations
- Theoretical bounds rely on strong assumptions (e.g., ReLU width constraints, semantic change under mixup assumption, and a constant beta term)
- Diversity is only tested with fixed auxiliary sets; gains depend on the diversity of the auxiliary data available
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200 ImageNet-RC ImageNet-800 SVHN Textures Places365 LSUN-crop LSUN-resize iSUN SSB-hard
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-200
Metrics: FPR95 AUROC AUPR ID-ACC
Baselines: MSP ODIN Mahalanobis Energy OE SOFL CCU NTOM POEM MixOE DivOE DPN
Builds on: mixup energy-based OOD regularization outlier exposure MixOE
Closest papers by meaning
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.92)
- DOS: Diverse Outlier Sampling for Out-of-Distribution Detection (ICLR 2024 · similarity 0.91)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.90)
- Mining In-distribution Attributes in Outliers for Out-of-distribution Detection (AAAI 2025 · similarity 0.90)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.89)
- How Does Unlabeled Data Provably Help Out-of-Distribution Detection? (ICLR 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.