Research map / Label-efficient and robust learning / OOD and anomaly detection
Out-of-Distribution Detection with Prototypical Outlier Proxy
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 Prototypical Outlier Proxy (POP), which fixes a hierarchy-aware classifier and adds virtual outlier prototypes plus a hierarchical similarity boundary loss, enabling OOD detection without real or synthetic outliers.
Problem
Overconfidence of deep classifiers on OOD inputs; existing outlier-exposure and outlier-synthesis methods are costly and biased toward specific outlier characteristics.
Approach
Fix the classifier weights to hierarchy-derived ID prototypes (HAFrame), append a few virtual OOD prototypes placed at a fixed hierarchical distance, and train with a hierarchical similarity boundary loss that penalizes misclassifications in proportion to semantic dissimilarity; score OOD with feature norm times MaxLogit.
Key results
- On CIFAR-10, POP reaches average FPR95 16.53% and AUROC 96.09%, the best among the compared methods (second-best NPOS at 24.23% FPR95).
- On CIFAR-100, POP reaches average FPR95 47.82% and AUROC 84.00%, versus 54.12% FPR95 for the second-best ViM.
- POP trains 7.23x faster than NPOS (9.00s vs 65.03s) and infers 19.5x faster (5.49s vs 107.05s) in the reported timing on CIFAR-10 to Places365.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use POP instead of NPOS | OOD detection on CIFAR-10 ID · CIFAR-10 ID, average OOD | FPR95: 24.23 -> 16.53 | Helped |
| use POP instead of ViM | OOD detection on CIFAR-100 ID · CIFAR-100 ID, average OOD | FPR95: 54.12 -> 47.82 | Helped |
| use POP instead of Energy | OOD detection on ImageNet-200 ID · ImageNet-200 ID, average OOD | AUROC: 87.18 -> 88.70 | Helped |
| add HSBL to the fixed model without outlier proxies (ablation) | Fixed ResNet-18 on CIFAR-10 · CIFAR-10 ID, average OOD | FPR95: 27.45 -> 25.65 | No clear effect |
| add HSBL to the fixed model without outlier proxies (ablation) | Fixed ResNet-18 on CIFAR-100 · CIFAR-100 ID, average OOD | FPR95: 54.56 -> 53.28 | Helped |
| add outlier proxies to the fixed model without HSBL (ablation) | Fixed ResNet-18 · CIFAR-10 and CIFAR-100 | FPR95: 27.45 -> 23.90; 54.56 -> 52.10 | Helped |
| add outlier proxies and HSBL (ablation) | Fixed ResNet-18 · CIFAR-10 and CIFAR-100 | FPR95: 27.45 -> 21.25; 54.56 -> 47.82 | Helped |
| replace hierarchical prototypes with random orthogonalized prototypes (ablation) | POP · CIFAR-10 and CIFAR-100 | FPR95: 16.53 -> 32.43; 47.82 -> 63.06 | Hurt |
| increase the number of outlier proxies (ablation) | POP on CIFAR-10 · CIFAR-10 ID | FPR95 | Hurt |
| vary HSBL scaling factor β (ablation) | POP · CIFAR-10 and CIFAR-100 | FPR95: 23.72; 18.92; 16.53; 18.69; 54.65; 47.82; 53.12; 57.47 | Mixed |
| use POP instead of NPOS | OOD detection · CIFAR-10 training, Places365 inference | Train time: 7.23 times faster | Helped |
| use POP instead of NPOS | OOD detection · CIFAR-10 training, Places365 inference | Inference time: 19.50 times the speed | Helped |
Surprising findings
- Adding more outlier proxies (larger C) generally decreased performance on CIFAR-10, which the authors attribute to excess proxies interfering with use of hierarchical prior information.
- Outlier-exposure and synthesis methods VOS and NPOS performed poorly on the simple MNIST dataset and on NINCO, sometimes worse than training-free post-hoc methods.
Limitations
- Gains on CIFAR-100 are weaker on some far-OOD sets (e.g., MNIST and TIN AUROC), and POP does not win on every dataset.
- Number and placement of outlier proxies (C, d) require a grid search and dataset-specific choices (d_max, beta).
- Experiments are limited to ResNet-18 trained from scratch on small and ImageNet-200 benchmarks.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200
Benchmarks: CIFAR-10 CIFAR-100 TIN MNIST SVHN Textures Places365 SSB-hard NINCO iNaturelist OpenImage-O
Metrics: FPR95 AUROC
Baselines: MSP Energy KNN MaxLogit ViM VOS NPOS
Builds on: HAFrame MaxLogit VOS NPOS OpenOOD
Closest papers by meaning
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.90)
- Out-Of-Distribution Detection with Diversification (Provably) (NeurIPS 2024 · similarity 0.90)
- DOS: Diverse Outlier Sampling for Out-of-Distribution Detection (ICLR 2024 · similarity 0.90)
- Learning with Mixture of Prototypes for Out-of-Distribution Detection (ICLR 2024 · similarity 0.90)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.90)
- Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (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.