Research map / Label-efficient and robust learning / OOD and anomaly detection
The Value of Out-of-Distribution Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Adding small amounts of OOD data to target training data | Fisher's Linear Discriminant on synthetic Gaussian mixture · synthetic FLD, n=100 | target generalization error | Mixed |
| Adding OOD samples from rotated CIFAR-10 (Bird vs Cat, large rotations) | WRN-10-2 trained OOD-agnostic · Split-CIFAR10 T2 with rotated OOD | target generalization error | Hurt |
| Adding Gaussian-blurred OOD samples (high blur) to Frog vs Horse target | WRN-10-2 OOD-agnostic · Split-CIFAR10 T4 with blurred OOD | target generalization error | Mixed |
| Adding ImageNet OOD samples to CIFAR-10 target (CINIC-10) | network trained on CINIC-10 subset · CIFAR-10 target with ImageNet OOD | target generalization error | No clear effect |
| Data augmentation (padding, random crop, flips) in OOD-agnostic training (ablation) | WRN-10-2 · Split-CIFAR10 T2 target, T5 OOD, n=100 | target generalization error vs OOD count | No clear effect |
| Pre-training on target-like ImageNet images then fine-tuning on combined data (ablation) | WRN-10-2 · Split-CIFAR10 T2 target, T5 OOD, n=100 | target generalization error vs OOD count | No clear effect |
| Hyperparameter tuning on target-only validation set (ablation) | WRN-10-2 OOD-agnostic · Split-CIFAR10 T2 target, T5 OOD | target generalization error vs OOD count | No clear effect |
| Alpha-weighted objective (alpha=1/2, naive) with known OOD identities (ablation) | WRN-10-2 / FLD · Split-CIFAR10 with known OOD samples | target generalization error vs OOD count | Helped |
| Tuning alpha via grid search (alpha*) in weighted objective with known OOD identities (ablation) | WRN-10-2 networks · CIFAR-10 and PACS target-OOD pairs | target test error | Helped |
| Biased mini-batch sampling with fraction beta target samples per batch (ablation) | alpha-weighted training · PACS and CIFAR-10 target-OOD pairs | target test error | Helped |
| Data augmentation combined with alpha-weighted objective (known OOD) (ablation) | WRN-10-2 · CIFAR-10 target-OOD pair with known OOD identities | target generalization error vs OOD count | Helped |
Closest papers by meaning
- Aggregation Hides Out-of-Distribution Generalization Failures from Spurious Correlations (NeurIPS 2025 · similarity 0.91)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.89)
- In Search of Forgotten Domain Generalization (ICLR 2025 · similarity 0.89)
- OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable? (ICML 2025 · similarity 0.88)
- Unraveling the Key Components of OOD Generalization via Diversification (ICLR 2024 · similarity 0.88)
- Towards Robust Out-of-Distribution Generalization Bounds via Sharpness (ICLR 2024 · similarity 0.88)
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.