Research map / Label-efficient and robust learning / OOD and anomaly detection
Spurious-Aware Prototype Refinement for Reliable 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 SPROD, a post-hoc, hyperparameter-free prototype refinement method that splits each class into majority and minority-group prototypes to detect OOD samples under unknown spurious correlations, and introduces the Animals MetaCoCo benchmark.
Problem
OOD detection is vulnerable to spurious correlations in training data, which make models misclassify spurious OOD samples as in-distribution and hurt reliability; existing methods often need retraining, extra OOD data, or are tested on limited benchmarks.
Approach
A three-stage post-hoc procedure builds class prototypes, adds group prototypes for misclassified training samples, and reassigns training samples to refine majority and minority group prototypes; OOD score is the distance to the nearest group prototype, with no group labels, extra data, or tuning.
Key results
- SPROD reaches 85.1% average AUROC on five SP-OOD benchmarks with ResNet-50, 4.8 points above the second best (KNN, 80.3).
- SPROD reaches 49.0% average FPR@95 on the same benchmarks, versus 58.4 for KNN, the second best.
- On Waterbirds with CLIP ViT-B/16, SPROD reports 99.01 AUROC and 2.94 FPR@95, matching or exceeding CMA and MCM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Refine class prototypes into group-specific prototypes | SPROD · Waterbirds SP-OOD | AUROC: 98.6 -> 98.8 | Helped |
| Refine class prototypes into group-specific prototypes | SPROD · Average SP-OOD benchmarks | AUROC: 80.3 -> 85.1 | Helped |
| Refine class prototypes into group-specific prototypes | SPROD · Average SP-OOD benchmarks | FPR@95: 58.4 -> 49.0 | Helped |
| Use softmax-normalized negative prototype distances instead of negative minimum distance (ablation) | SPROD · Waterbirds, CelebA, and UrbanCars SP-OOD | OOD detection performance | Hurt |
| Use softmax-normalized negative prototype distances instead of negative minimum distance (ablation) | SPROD · Animals MetaCoCo and Spurious ImageNet SP-OOD | OOD detection performance | Hurt |
| Fine-tune the feature backbone on ID data (ablation) | OOD detection methods · Waterbirds SP-OOD | OOD detection performance | Hurt |
| Increase ID spurious correlation from 50% to 90% (ablation) | OOD detection methods · Waterbirds SP-OOD | OOD detection performance | Hurt |
| Use a ResNet-18 instead of ResNet-50 backbone (ablation) | OOD detection methods · Waterbirds SP-OOD | OOD detection performance | No clear effect |
| Reduce the number of ID training samples while preserving the spurious correlation level (ablation) | SPROD and KNN · CelebA SP-OOD | OOD detection performance | Helped |
| Use a post-hoc visual method without text inputs | SPROD vs. text-based methods · Waterbirds | AUROC: 98.36; 98.62; 99.01 | Mixed |
Surprising findings
- Fine-tuning the backbone on ID data generally degrades SP-OOD detection performance, more so for feature-based methods.
- Softmax-based discriminative scoring substantially degrades SPROD performance compared with distance-based generative scoring on several benchmarks.
Limitations
- Relies on ID class labels to construct class-conditional prototypes.
- Not always best on every metric, particularly in conventional (non-spurious) OOD settings where it is only competitive.
- Theoretical analysis uses a simplified binary orthogonal-subspace model with assumed class-conditional distributions.
Experimental setup
Datasets: Waterbirds CelebA UrbanCars Animals MetaCoCo Spurious ImageNet (Sp-ImageNet100) CIFAR-10 CIFAR-100 ImageNet-1k
Benchmarks: Waterbirds CelebA UrbanCars Animals MetaCoCo Spurious ImageNet100 CIFAR-10 CIFAR-100 ImageNet-1k (OpenOOD)
Metrics: AUROC FPR@95 AUPR
Baselines: MSP Energy MLS KLM GEN GradNorm ReAct ViM ASH MDS RMDS KNN
Builds on: Mahalanobis distance OOD detection KNN OOD detection OpenOOD
Closest papers by meaning
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.90)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.90)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.90)
- Position: Supervised Classifiers Answer the Wrong Questions for OOD Detection (ICML 2025 · similarity 0.90)
- Learning with Mixture of Prototypes for Out-of-Distribution Detection (ICLR 2024 · similarity 0.89)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · 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.