Research map / Label-efficient and robust learning / OOD and anomaly detection
Test-Time Linear 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 RTL, a test-time linear regression that calibrates base OOD scores using the linear relation between network features and scores over a batch of test data, plus a sparse-outlier variant RTL++ and an online version.
Problem
Out-of-distribution detection with neural classifiers is unreliable because networks are overconfident on OOD inputs; existing post hoc scores are computed per sample and ignore interactions among test samples.
Approach
Observes an approximately linear relation between base OOD scores (MSP, energy, ODIN, KL) and penultimate features on benchmarks; fits a least-squares linear regression from features to scores over the test batch and uses fitted values as rectified OOD scores. RTL++ adds a LASSO-based incidental-parameter step to drop high-error samples before fitting, and an online variant updates the normal equa
Key results
- On CIFAR-100, RTL raises MSP AUROC from 75.67 to 84.50 and cuts FPR95 from 80.21 to 51.63 (RTL++: 43.73 FPR95).
- On ImageNet-1k with BiT ResNetV2-101, MSP+RTL reduces average FPR95 from 76.96% to 46.85% and RTL++ to 45.97%, with AUROC 79.29 to 87.38.
- Online RTL on CIFAR-100 with batch size 32 gives MSP FPR95 55.84 and AUROC 83.31, close to the full-batch result (51.63 / 84.50).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply RTL to MSP scores | MSP with Wide ResNet · CIFAR-10 OOD | FPR95: 51.37 -> 12.30 | Helped |
| Apply RTL to MSP scores | MSP with Wide ResNet · CIFAR-100 OOD | AUROC: 75.67 -> 84.50 | Helped |
| Apply RTL++ instead of RTL to MSP scores (ablation) | MSP with Wide ResNet · CIFAR-100 OOD | FPR95: 51.63 -> 43.73 | Helped |
| Apply RTL++ instead of RTL to MSP scores (ablation) | MSP with Wide ResNet · CIFAR-10 OOD | FPR95: 12.30 -> 13.50 | Hurt |
| Apply RTL++ to Energy scores | Energy with Wide ResNet · CIFAR-100 OOD | FPR95: 73.46 -> 58.06 | Helped |
| Apply RTL to MSP scores | MSP with ResNetV2-101 · ImageNet-1k OOD average | FPR95: 76.96 -> 46.85 | Helped |
| Apply RTL++ instead of RTL to MSP scores (ablation) | MSP with ResNetV2-101 · ImageNet-1k OOD average | AUROC: 86.71 -> 87.38 | Helped |
| Apply RTL++ instead of RTL to Energy scores (ablation) | Energy with ResNetV2-101 · ImageNet-1k OOD average | FPR95: 57.43 -> 56.06 | Helped |
| Apply RTL++ instead of RTL to ODIN scores (ablation) | ODIN with Wide ResNet · CIFAR-10 OOD | FPR95: 37.94 -> 35.79 | No clear effect |
| Use online RTL instead of full-test-set RTL (ablation) | MSP with Wide ResNet · CIFAR-10 OOD | FPR95: 12.30 -> 15.06 | Hurt |
| Use online RTL instead of full-test-set RTL (ablation) | MSP with Wide ResNet · CIFAR-100 OOD | AUROC: 84.50 -> 83.31 | Hurt |
| Increase online RTL batch size from 32 to 1024 (ablation) | Online RTL with MSP · CIFAR-10 OOD | FPR95: 15.06 -> 14.16 | Helped |
Surprising findings
- Base OOD scores from different algorithms (MSP, energy, ODIN, KL) show an approximately linear relation with penultimate features, regardless of the detector used.
- Traditional outlier detectors (IF, LOF, GMM) perform poorly on OOD test batches where OOD data makes up a large share of the data.
Limitations
- Requires a batch of test data at inference, so it is not a per-sample detector.
- Gains depend on the base score being roughly linear in features; ODIN on CIFAR-10 did not improve in some settings (e.g., online ODIN FPR95 worsened to 40.30).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Textures SVHN Places365 LSUN-Crop LSUN-Resize iSUN ImageNet-1k iNaturalist SUN
Benchmarks: CIFAR-10 OOD suite CIFAR-100 OOD suite ImageNet-1k OOD (iNaturalist, SUN, Places, Textures)
Metrics: FPR95 AUROC
Baselines: MSP ODIN Energy KL GradNorm Isolation Forest Local Outlier Factor Gaussian Mixture Models
Builds on: maximum softmax probability ODIN energy-based OOD detection GradNorm test-time adaptation
Closest papers by meaning
- OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary (CVPR 2025 · similarity 0.91)
- 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 0.89)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.89)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- Exploring Channel-Aware Typical Features for Out-of-Distribution Detection (AAAI 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.