Research map / Label-efficient and robust learning / OOD and anomaly detection
HYPO: Hyperspherical Out-Of-Distribution Generalization
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 HYPO, a hyperspherical prototype-based loss (intra-class alignment plus inter-prototype separation) for OOD generalization, with a theoretical bound linking the loss to reduced intra-class variation.
Problem
Out-of-distribution generalization requires learning features whose class-conditional distributions are stable across training domains and well separated across classes; prior theory bounds OOD error by intra-class variation and inter-class separation but gives no practical algorithm to optimize the
Approach
Normalize features onto a unit hypersphere and train with a prototype-based loss: a softmax term pulling samples toward their class prototype (variation) plus a term pushing class prototypes apart (separation), interpreted as vMF maximum likelihood and shown theoretically to reduce the variation term in the OOD error bound.
Key results
- On PACS with SGD, HYPO reaches 88.0% average accuracy, 1.1 points above the best loss-based baseline (SelfReg-style methods at 86.9-ish per table; PCL 86.3%).
- On CIFAR-10 to CIFAR-10-C Gaussian noise, OOD accuracy improves from 78.09% (ERM) to 85.21%.
- Averaged over PACS, Office-Home and VLCS, HYPO achieves 79.3% versus 78.1% for the strongest listed baseline (GVRT); with SWAD, PACS accuracy rises to 89.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HYPO instead of ERM | ResNet-18 trained on CIFAR-10 · CIFAR-10-C Gaussian noise | OOD accuracy: 78.09\% to 85.21\% (+\textbf{7.12}\%) | Helped |
| Use HYPO instead of the strongest listed loss-based baseline | ResNet-50 domain generalization · PACS | accuracy: 86.9 to 88.0 | Helped |
| Use HYPO instead of GVRT | ResNet-50 domain generalization · Office-Home | accuracy: 70.1 to 71.7 | Helped |
| Use HYPO instead of GVRT | ResNet-50 domain generalization · VLCS | accuracy: 79.0 to 78.2 | Hurt |
| Use HYPO instead of GVRT | ResNet-50 domain generalization · PACS, Office-Home, VLCS | average accuracy: 78.1 to 79.3 | Helped |
| Use HYPO rather than PCL with SGD | PACS domain generalization · PACS | average accuracy: 86.3 to 88.0 | Helped |
| Use HYPO rather than PCL with SWAD | PACS domain generalization · PACS | average accuracy: 88.7 to 89.0 | Helped |
| Add SWAD optimization to HYPO | HYPO · PACS | average accuracy: 88.0 to 89.0 | Helped |
| Emphasize hard negative pairs from the same domain | HYPO variation loss for multi-source domain generalization · PACS, Office-Home, VLCS | accuracy | No clear effect |
Limitations
- Theoretical bound relies on assumptions such as uniform domain and class distributions and sufficient inter-class separation for learnability.
- Hard-negative modification to the variation loss is used only for multi-source DG and is a heuristic extension.
- Inter-class separation benefit is supported mainly by ablation rather than a full theoretical proof in the main text.
Experimental setup
Datasets: CIFAR-10 PACS Office-Home VLCS Terra Incognita ImageNet-100
Benchmarks: CIFAR-10-C PACS Office-Home VLCS Terra Incognita ImageNet-100
Metrics: OOD accuracy ID accuracy average accuracy Sinkhorn divergence
Baselines: ERM CORAL DANN MLDG MMD IRM GroupDRO I-Mixup RSC ARM VREx Mixstyle
Builds on: PCL SWAD variation-based OOD error bound (Ye et al. 2021) DomainBed
Closest papers by meaning
- Provable Discriminative Hyperspherical Embedding for Out-of-Distribution Detection (AAAI 2025 · similarity 0.88)
- ProHOC: Probabilistic Hierarchical Out-of-Distribution Classification via Multi-Depth Networks (CVPR 2025 · similarity 0.86)
- CRoFT: Robust Fine-Tuning with Concurrent Optimization for OOD Generalization and Open-Set OOD Detection (ICML 2024 · similarity 0.86)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.86)
- Towards Robust Out-of-Distribution Generalization Bounds via Sharpness (ICLR 2024 · similarity 0.85)
- OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable? (ICML 2025 · similarity 0.85)
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.