Research map / Label-efficient and robust learning / Recognition and long-tail
Contextual Reliability: When Different Features Matter in Different Contexts
Gaurav Rohit Ghosal, Amrith Setlur, Daniel S. Brown, Anca Dragan, Aditi Raghunathan
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 |
|---|---|---|---|
| Standard group DRO (group definitions without context) instead of ERM | Corrupted Waterbirds · Corrupted Waterbirds | worst-group accuracy: 0.67 to 0.60 | Hurt |
| Add ENP feature annotations on only 10% of training data | Corrupted Waterbirds · Corrupted Waterbirds | worst-group accuracy: 0.728 vs 0.73 for conDRO and GT-Aug. with full context labels | No clear effect |
| Add ENP (feature-annotation-based two-stage) over ERM | Corrupted Waterbirds · Corrupted Waterbirds | worst-group accuracy: 0.67 to 0.728 | Improved |
| Add ENP feature-based augmentation of the previous action on noisy MountainCar | Noisy MountainCar · Noisy MountainCar | test reward: -139.5 vs best baseline -170.4 | Improved |
| Provide previous action as input to the imitation policy (no augmentation) | Noisy MountainCar · Noisy MountainCar | test reward: -170.4 to -194 | Hurt |
| Add conDRO with ground-truth context on all points | Noisy MountainCar · Noisy MountainCar | test reward: -188.3 vs -194 with previous action | No clear effect |
| Add annotated-spurious-agent augmentation (20% annotated) to MultiPath++ | WOMD MultiPath++ · Waymo Open Motion Dataset | minADE (perturbed, lower is better): 0.817 to 0.801 | Improved |
| Random agent-deletion augmentation without annotations | WOMD MultiPath++ · Waymo Open Motion Dataset | minADE (perturbed, lower is better): 0.817 to 0.815 | No clear effect |
| ENP: predicted spuriousness labels from a small feature predictor used to augment all trajectories | WOMD MultiPath++ · Waymo Open Motion Dataset | minADE (perturbed, lower is better): 0.801 to 0.774 | Improved |
| Increase fraction of non-spurious feature annotations in the simplified linear/deep setup (ablation) | ENP on simplified two-context setup · Simplified two-context setup | test accuracy | Improved |
| Use ENP feature-based context structure instead of ERM/IRM/conDRO on simplified setup (asymptotic) | ENP vs conDRO, IRM, ERM · Simplified two-context setup (theory) | asymptotic test accuracy | Improved |
Closest papers by meaning
- How Spurious Features are Memorized: Precise Analysis for Random and NTK Features (ICML 2024 · similarity 0.83)
- Complexity Matters: Feature Learning in the Presence of Spurious Correlations (ICML 2024 · similarity 0.83)
- Contextures: Representations from Contexts (ICML 2025 · similarity 0.83)
- On the Foundations of Shortcut Learning (ICLR 2024 · similarity 0.83)
- Testing Semantic Importance via Betting (NeurIPS 2024 · similarity 0.82)
- Quantifying Uncertainty in Error Consistency: Towards Reliable Behavioral Comparison of Classifiers (NeurIPS 2025 · similarity 0.82)
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.
