Research map / Label-efficient and robust learning / OOD and anomaly detection
Learning with Mixture of Prototypes for 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 PALM, which models each ID class with a mixture of multiple vMF prototypes learned via reciprocal-neighbor soft assignments, trained with an MLE loss and a prototype contrastive loss for distance-based OOD detection.
Problem
Distance-based OOD detection methods model each ID class with a single prototype/centroid, which oversimplifies within-class diversity and yields non-compact embeddings and ID-OOD confusion.
Approach
Model each class's hyperspherical embeddings as a mixture of vMF distributions with K learnable prototypes, assign samples softly via Sinkhorn-based assignment with top-K pruning, update prototypes by EMA, and add a prototype-level contrastive loss alongside the MLE loss.
Key results
- On labeled CIFAR-100 with ResNet-34 and Mahalanobis scoring, PALM reaches average AUROC 93.82 and FPR 28.02, versus 89.34 AUROC for CIDER.
- With KNN scoring (K=300), PALM achieves average AUROC 92.09 and FPR 30.23 on CIFAR-100.
- In the unlabeled CIFAR-100 setting, PALM reaches average AUROC 88.07 versus 86.11 for CSI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace a single prototype per class with a mixture of prototypes | distance-based OOD detection on CIFAR-100 · CIFAR-100 ID; average across SVHN, Places365, LSUN, iSUN, Textures | AUROC | Helped |
| use multiple prototypes rather than one prototype per class (ablation) | PALM · CIFAR-100 ID; number of prototypes varied | OOD detection performance | Helped |
| prune less relevant assignments with top-K selection (ablation) | PALM · CIFAR-100 ID; prototype assignment | OOD detection performance | Helped |
| use soft instead of hard prototype assignments (ablation) | PALM · CIFAR-100 ID; prototype assignment | OOD detection performance | Helped |
| update prototypes with EMA (ablation) | PALM · CIFAR-100 ID; prototype updating | cluster assignment consistency | Helped |
| use KNN distance instead of Mahalanobis distance (ablation) | PALM · CIFAR-100 ID; average across SVHN, Places365, LSUN, iSUN, Textures | AUROC: 92.09 vs 93.82 | Hurt |
| use PALM instead of NPOS | CIFAR-100 supervised OOD detection · Average across SVHN, Places365, LSUN, iSUN, Textures | AUROC: 89.84 -> 93.82 | Helped |
| train PALM without class labels | PALM · Unlabeled CIFAR-100 ID; average across SVHN, Places365, LSUN, iSUN, Textures | AUROC: 86.11 -> 88.07 | Helped |
Surprising findings
- Using only one prototype per class performs worse, and performance increases as the number of prototypes grows, indicating single-prototype modeling is limiting.
Limitations
- Number of prototypes K must be manually set as a hyperparameter.
- Performance on Places365 remains weak (FPR 64.66) relative to some baselines.
- Unsupervised extension is a simple adaptation with only preliminary results.
Experimental setup
Datasets: CIFAR-100 CIFAR-10 ImageNet-100
Benchmarks: SVHN Places365 LSUN iSUN Textures LSUN-FIX ImageNet-FIX ImageNet-RESIZE CIFAR-10 OpenOOD SUN iNaturalist
Metrics: FPR@95TPR AUROC ID accuracy
Baselines: MSP Vim ODIN Energy VOS CSI SSD+ KNN+ NPOS CIDER
Builds on: CIDER SwAV SupCon Mahalanobis distance KNN OOD detection
Closest papers by meaning
- Learning to Shape In-distribution Feature Space for Out-of-distribution Detection (NeurIPS 2024 · similarity 0.91)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.90)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.89)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.89)
- DPU: Dynamic Prototype Updating for Multimodal Out-of-Distribution Detection (CVPR 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.