Research map / Label-efficient and robust learning / Recognition and long-tail
Testing Using Privileged Information by Adapting Features With Statistical Dependence
Kwang In Kim, James Tompkin
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 |
|---|---|---|---|
| add TUPI denoising with other feature sets as test time information | initial predictor · MFeat (F1 as baseline features) | ranking accuracy (%): 77.85 -> 81.97 | Improved |
| add TUPI denoising with other feature sets as test time information | initial predictor · MFeat (F4 as baseline features) | ranking accuracy (%): 70.88 -> 74.33 | Improved |
| add TUPI denoising with other feature sets as test time information | initial predictor · MFeat (F6 as baseline features) | ranking accuracy (%): 77.10 -> 82.25 | Improved |
| add 10 random features to the test time information (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 81.97 -> 82.14 | No clear effect |
| add 10 random features to the test time information (ablation) | TUPI · MFeat (F5 as baseline features) | ranking accuracy (%): 78.19 -> 78.05 | No clear effect |
| use only 1 of 5 feature sets instead of all 5 (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 81.97 -> 79.88 | Hurt |
| use only 3 of 5 feature sets instead of all 5 (ablation) | TUPI · MFeat (F6 as baseline features) | ranking accuracy (%): 82.25 -> 80.71 | Hurt |
| use only 3 of 5 feature sets instead of all 5 (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 81.97 -> 82.15 | No clear effect |
| supply noise-free ground-truth targets as test time features (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 77.85 -> 99.56 | Improved |
| increase noise std of ground-truth test time features from 0.2 to 1 (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 87.84 -> 78.84 | Hurt |
| add copies of the initial prediction to the test time features (11 copies) (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 81.97 -> 78.96 | Hurt |
| add a single copy of the initial prediction to the test time features (ablation) | TUPI · MFeat (F1 as baseline features) | ranking accuracy (%): 81.97 -> 82.81 | No clear effect |
Closest papers by meaning
- Spuriosity Didn’t Kill the Classifier: Using Invariant Predictions to Harness Spurious Features (NeurIPS 2023 · similarity 0.85)
- Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent (NeurIPS 2025 · similarity 0.85)
- Testing Semantic Importance via Betting (NeurIPS 2024 · similarity 0.85)
- Entropy is not Enough for Test-Time Adaptation: From the Perspective of Disentangled Factors (ICLR 2024 · similarity 0.84)
- Sample Selection via Contrastive Fragmentation for Noisy Label Regression (NeurIPS 2024 · similarity 0.83)
- Prediction-Powered Active Testing (NeurIPS 2026 · similarity 0.83)
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.
