Research map / Label-efficient and robust learning / OOD and anomaly detection
The Best of Both Worlds: On the Dilemma of 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.
Shows that SOTA OOD detectors sacrifice OOD generalization (accuracy under Gaussian noise), proves a sensitive-robust dilemma for MSP and EBM regularization, and proposes Decoupled Uncertainty Learning (DUL) to raise distributional uncertainty on auxiliary outliers while keeping overall entropy unchanged.
Problem
OOD detection methods that regularize models on auxiliary outliers improve detection but degrade classification accuracy under covariate shift (noise/corruption), creating a sensitive-robust dilemma.
Approach
Theoretically links the OOD detection loss to a lower bound on generalization error via disparity discrepancy, then decouples distributional uncertainty (raised for semantic OOD) from overall predictive entropy (kept unchanged via a KL term), so detection and generalization no longer trade off.
Key results
- On CIFAR-10 with ImageNet-RC auxiliary outliers, DUL achieves OOD-Acc 88.01% versus 87.35% for MSP, with FPR95 5.89% versus 41.96% for MSP, while POEM reaches FPR95 3.32% but OOD-Acc drops to 78.89%.
- On ImageNet-200, DUL attains FPR95 49.14% and AUROC 89.27%, the best among compared methods, with ID-Acc 85.65% and OOD-Acc 75.59%.
- Ablation: Entropy regularization raised OOD error from 9.55% to 18.53% as its weight increased, illustrating the trade-off.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DUL rather than POEM | CIFAR-10 classifier with ImageNet-RC auxiliary outliers · CIFAR-10 / ImageNet-RC | OOD-Acc: 78.89 -> 88.01 | Helped |
| Use DUL rather than POEM | CIFAR-10 classifier with ImageNet-RC auxiliary outliers · CIFAR-10 / ImageNet-RC | FPR95: 3.32 -> 5.89 | Hurt |
| Use DUL rather than the MSP baseline | CIFAR-100 classifier with ImageNet-RC auxiliary outliers · CIFAR-100 / ImageNet-RC | FPR95: 74.63 -> 12.49 | Helped |
| Use DUL rather than the MSP baseline | ImageNet-200 classifier with ImageNet-800 auxiliary outliers · ImageNet-200 / ImageNet-800 | OOD-Acc: 74.84 -> 75.59 | Helped |
| Use Thompson sampling for OOD informativeness mining (ablation) | DUL on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC | OOD-Acc: 88.01 -> 87.53 | Hurt |
| Add DUL regularization to finetuned EBM | EBM on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC | OOD-Acc: 79.03 -> 87.45 | Helped |
| Add DUL regularization to Entropy | Entropy on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC | FPR95: 6.63 -> 29.56 | Hurt |
| Increase Entropy regularization weight from 0 to 5×10⁻⁴ (ablation) | Entropy on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC, Gaussian noise | OOD-Err: 9.55 -> 13.58 | Hurt |
| Increase Entropy regularization weight from 0 to 5×10⁻⁴ (ablation) | Entropy on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC | FPR95: 35.15 -> 8.36 | Helped |
| Increase EBM regularization weight from 0 to 1×10⁻¹ (ablation) | EBM on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC, Gaussian noise | OOD-Err: 9.55 -> 24.38 | Hurt |
| Increase EBM regularization weight from 0 to 1×10⁻¹ (ablation) | EBM on CIFAR-10 / ImageNet-RC · CIFAR-10 / ImageNet-RC | FPR95: 20.57 -> 6.11 | Helped |
| Add the unchanged-overall-uncertainty constraint to OOD regularization (ablation) | DUL on semantic OOD samples · CIFAR-10 / ImageNet-RC | overall predictive uncertainty | Helped |
Surprising findings
- SOTA OOD detectors such as POEM and WOODS lose roughly 10% OOD accuracy under Gaussian noise relative to the MSP baseline while improving detection.
- Adding the DUL term to Entropy regularization did not substantially improve accuracy, since the two objectives directly conflict.
Limitations
- Requires auxiliary outlier data during training, which the authors acknowledge as a cost.
- Theory does not directly cover methods with extra OOD detection output branches.
- Covariate shift evaluated mainly with Gaussian noise in the main text; large variance on some OOD-Acc numbers (e.g., ±3.29 on CIFAR-100).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200 ImageNet-RC TIN-597 ImageNet-800 SVHN Textures Places365 iSUN LSUN-C LSUN-R
Benchmarks: CIFAR-10 OOD detection suite CIFAR-100 OOD detection suite ImageNet-200 OOD detection suite
Metrics: ID-Acc OOD-Acc FPR95 AUROC AUPR
Baselines: MSP Entropy regularization EBM (pretrain) EBM (finetune) Maxlogits Mahalanobis DPN WOODS POEM SCONE
Builds on: Outlier Exposure Energy-based OOD detection Posterior Network / Dirichlet prior networks SCONE POEM
Closest papers by meaning
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.91)
- Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · similarity 0.91)
- Position: Supervised Classifiers Answer the Wrong Questions for OOD Detection (ICML 2025 · similarity 0.90)
- What If the Input is Expanded in OOD Detection? (NeurIPS 2024 · similarity 0.90)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.90)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.90)
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.