Research map / Label-efficient and robust learning / OOD and anomaly detection
Predicting Is Not Understanding: Recognizing and Addressing Underspecification in Machine Learning
Damien Teney, Maxime Peyrard, Ehsan Abbasnejad
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 independence constraint to ERM (ablation) | linear classifier on frozen features · WILDS-Camelyon17 test-OOD | best accuracy (%): 68.4 -> 74.6 | Improved |
| add on-manifold hard projection (PCA) on top of independence (ablation) | linear classifier on frozen features · WILDS-Camelyon17 test-OOD | best accuracy (%): 74.6 -> 81.6 | Improved |
| add fine-tuning and distillation (ablation) | full method · WILDS-Camelyon17 test-OOD | best accuracy (%): 81.6 -> 82.5 | Improved |
| use PCA instead of VAE for the manifold with hard projection (ablation) | Camelyon17 classifier · WILDS-Camelyon17 test-OOD | best accuracy (%): 76.3 -> 81.6 | Improved |
| add independence constraint to ERM (ablation) | MLP ensemble · Collages (average of 4 test sets) | accuracy (%): 62.5 -> 63.0 | No clear effect |
| add on-manifold constraint to independence loss (PCA, 8 models) (ablation) | Collages MLPs · Collages (average of 4 test sets) | accuracy (%): 64.2 -> 72.3 | Improved |
| use VAE instead of PCA for the manifold (ablation) | Collages MLPs · Collages (average of 4 test sets) | accuracy (%): 72.3 -> 76.2 | Improved |
| add fine-tuning on masked inputs (ablation) | Collages proposed method · Collages (average of 4 test sets) | accuracy (%): 76.2 -> 84.8 | Improved |
| add pairwise combinations (3x) (ablation) | Collages proposed method + FT · Collages (average of 4 test sets) | accuracy (%): 84.8 -> 85.4 | Improved |
| add independence constraint alone on 16 models (ablation) | GQA yes/no model · GQA yes/no val (best of N models) | accuracy (%): 68.6 -> 68.1 | No clear effect |
| add on-manifold constraint to independence on 16 models (ablation) | GQA yes/no model · GQA yes/no val (best of N models) | accuracy (%): 68.1 -> 69.3 | Improved |
| add on-manifold constraint to independence on 16 models (ablation) | GQA yes/no model · GQA-OOD Val-tail (best of N models) | accuracy (%): 69.6 -> 71.5 | 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.
- FixMatch — FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) · better on 1 of 1 shared result
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
- Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity (NeurIPS 2025 · similarity 0.85)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.84)
- OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable? (ICML 2025 · similarity 0.84)
- A Closer Look at AUROC and AUPRC under Class Imbalance (NeurIPS 2024 · similarity 0.83)
- Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods (ICML 2025 · similarity 0.83)
- Unraveling the Key Components of OOD Generalization via Diversification (ICLR 2024 · 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.
