Research map / Label-efficient and robust learning / OOD and anomaly detection
AP-OOD: Attention Pooling for Out-of- Distribution Detection
Claus Hofmann, Christian Huber, Bernhard Lehner, Daniel Klotz, Sepp Hochreiter, Werner Zellinger
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace mean pooling with learned attention pooling over token embeddings | Mahalanobis · XSUM summarization input OOD CNN/DM | FPR95: 74.20 -> 12.88 | Improved |
| replace mean pooling with attention pooling (AP-OOD vs Mahalanobis) | OOD detector · XSUM summarization input OOD (mean) | AUROC: 88.44 -> 98.91 | Improved |
| replace mean pooling with attention pooling | OOD detector · XSUM summarization output OOD (mean) | AUROC: 88.49 -> 96.39 | Improved |
| use AP-OOD instead of KNN | OOD detector · WMT15 En-Fr input OOD (mean) | AUROC: 71.00 -> 74.81 | Improved |
| use AP-OOD instead of KNN | OOD detector · WMT15 En-Fr output OOD (mean) | AUROC: 74.39 -> 77.38 | Improved |
| use AP-OOD instead of MSP | classifier OOD detector · MIMII-DG audio | FPR95: 36.43 -> 22.35 | Improved |
| add auxiliary OOD samples (semi-supervised AP-OOD) | AP-OOD · XSUM summarization | AUROC | Improved |
| use fully supervised AP-OOD instead of unsupervised AP-OOD | AP-OOD · XSUM summarization input OOD (mean) | AUROC: 98.91 -> 99.94 | Improved |
| use AP-OOD instead of binary logits in supervised setting | OOD detector · XSUM summarization input OOD (mean) | AUROC: 99.74 -> 99.94 | Improved |
| scale AP-OOD to more heads and queries (M=1024, T=16) (ablation) | AP-OOD · XSUM summarization (full XSUM training) | AUROC | Improved |
| use perplexity instead of embedding-based methods for output OOD | Perplexity · WMT15 En-Fr output OOD (mean) | FPR95: 81.10 -> 77.08 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Mahalanobis — Mahalanobis++: Improving OOD Detection via Feature Normalization (ICML 2025) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection (AAAI 2024 · similarity 0.90)
- Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection (ICML 2024 · similarity 0.89)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.88)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.87)
- AHA: Human-Assisted Out-of-Distribution Generalization and Detection (NeurIPS 2024 · similarity 0.87)
- Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · similarity 0.87)
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.
