Research map / Label-efficient and robust learning / OOD and anomaly detection
Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail
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 PATT, a long-tailed OOD detection method that combines vMF-based implicit semantic augmentation contrastive learning with temperature-scaled logit adjustment during training and attention-based feature calibration at inference.
Problem
Out-of-distribution detection under long-tailed in-distribution data, where existing methods that reduce head-class semantics improve OOD detection at the cost of ID classification accuracy.
Approach
Model ID features with a mixture of vMF distributions and derive a closed-form infinite-sample implicit semantic augmentation contrastive loss, add temperature-scaled logit adjustment to boost ID confidence, and apply a training-set-derived channel attention weight to calibrate features at inference instead of reducing head-class semantics.
Key results
- On ImageNet-LT, PATT reaches 74.13 AUROC and 55.14 ID accuracy, versus 69.84 AUROC and 46.79 ACC for EAT (+4.29 AUROC, +8.35 ACC).
- On CIFAR100-LT (average over six OOD sets), PATT achieves 76.25 AUROC, 62.94 FPR95 and 50.07 ACC, outperforming EAT's 74.41 AUROC and 46.13 ACC.
- On CIFAR10-LT, PATT achieves 91.62 AUROC and 84.77 ACC, with PATT's AUROC roughly equal to EAT (91.73) and ACC above EAT's 81.31.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PATT instead of EAT | CIFAR10-LT, six-OOD average · CIFAR10-LT | AUROC: 91.73 -> 91.62 | Hurt |
| Use PATT instead of EAT | CIFAR10-LT, six-OOD average · CIFAR10-LT | ACC: 81.31 -> 84.77 | Helped |
| Use PATT instead of EAT | CIFAR100-LT, six-OOD average · CIFAR100-LT | AUROC: 74.41 -> 76.25 | Helped |
| Use PATT instead of EAT | CIFAR100-LT, six-OOD average · CIFAR100-LT | ACC: 46.13 -> 50.07 | Helped |
| Use PATT instead of EAT | ImageNet-LT with ImageNet-1k-OOD · ImageNet-LT | AUROC: 69.84 -> 74.13 | Helped |
| Use PATT instead of EAT | ImageNet-LT with ImageNet-1k-OOD · ImageNet-LT | ACC: 46.79 -> 55.14 | Helped |
| Add ISAC, TLA, and FC to the baseline (ablation) | CIFAR10-LT · CIFAR10-LT, six-OOD average | AUROC: 71.21 -> 91.62 | Helped |
| Add ISAC, TLA, and FC to the baseline (ablation) | CIFAR100-LT · CIFAR100-LT, six-OOD average | AUROC: 71.70 -> 76.25 | Helped |
| Add ISAC, TLA, and FC to the baseline (ablation) | ImageNet-LT · ImageNet-LT | AUROC: 67.87 -> 74.13 | Helped |
| Add FC to ISAC and TLA (ablation) | CIFAR10-LT · CIFAR10-LT, six-OOD average | AUROC: 91.06 -> 91.62 | Helped |
| Add FC to ISAC and TLA (ablation) | CIFAR100-LT · CIFAR100-LT, six-OOD average | AUROC: 75.48 -> 76.25 | Helped |
| Add FC to ISAC and TLA (ablation) | ImageNet-LT · ImageNet-LT | AUROC: 73.77 -> 74.13 | Helped |
Surprising findings
- On CIFAR10-LT, PATT slightly trails EAT in AUROC and AUPR-in, contrary to its gains on larger datasets.
- Post-hoc logit adjustment and tau-norm can degrade results when combined with the balanced TISAC encoder, while feature calibration still helps.
Limitations
- On CIFAR10-LT, OOD detection is not better than EAT in AUROC and AUPR-in.
- Hyperparameters alpha, beta, epsilon and tau are tuned on the evaluation setting, with the best AUROC and best ACC obtained at different settings.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT TinyImages80M ImageNet-Extra Textures SVHN Tiny ImageNet LSUN Places365 CIFAR-100 ImageNet-1k-OOD
Benchmarks: CIFAR10-LT CIFAR100-LT ImageNet-LT SC-OOD benchmark
Metrics: AUROC AUPR-in AUPR-out FPR95 ACC
Baselines: MSP OE EnergyOE PASCL EAT ODIN tau-norm
Builds on: Outlier Exposure ISDA PASCL EnergyOE logit adjustment
Closest papers by meaning
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.95)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.93)
- Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning (AAAI 2024 · similarity 0.92)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.90)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.90)
- Mining In-distribution Attributes in Outliers for Out-of-distribution Detection (AAAI 2025 · 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.