Research map / Label-efficient and robust learning / Recognition and long-tail
Spuriosity Didn’t Kill the Classifier: Using Invariant Predictions to Harness Spurious Features
Cian Eastwood, Shashank Singh, Andrei Liviu Nicolicioiu, Marin Vlastelica, Julius von Kügelgen, Bernhard Schölkopf
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 |
|---|---|---|---|
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · Synthetic AC | accuracy: 74.7 ± 1.2 -> 89.2 ± 2.9 | Improved |
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · Synthetic CE-DD | accuracy: 74.9 ± 3.6 -> 88.6 ± 1.4 | Improved |
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · CMNIST | accuracy: 70.6 ± 1.8 -> 88.1 ± 1.8 | Improved |
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · PACS P | accuracy: 93.7 ± 0.6 -> 95.8 ± 0.6 | Improved |
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · PACS A | accuracy: 78.1 ± 1.1 -> 80.4 ± 1.3 | Improved |
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · PACS C | accuracy: 73.7 ± 0.6 -> 76.6 ± 0.6 | Improved |
| adapt the unstable-feature predictor using unlabeled test-domain data (ablation) | SFB · PACS S | accuracy: 69.7 ± 2.3 -> 71.8 ± 2.0 | Improved |
| add bias correction to test-domain adaptation (ablation) | SFB · CMNIST test domains with varying color-label correlation | accuracy | Improved |
| add post-hoc calibration of the stable predictor (ablation) | SFB · CMNIST test domains with varying color-label correlation | accuracy | Improved |
| use multiple pseudo-labeling rounds (ablation) | SFB · CMNIST test domains with varying color-label correlation | accuracy | Improved |
| use Stable Feature Boosting instead of ERM | domain generalization · Synthetic AC | accuracy: 9.9 ± 0.1 -> 89.2 ± 2.9 | Improved |
| use Stable Feature Boosting instead of ERM | domain generalization · Synthetic CE-DD | accuracy: 11.6 ± 0.7 -> 88.6 ± 1.4 | Improved |
Closest papers by meaning
- Learning Stable Classifiers by Transferring Unstable Features (ICML 2022 · similarity 0.87)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.85)
- Complexity Matters: Feature Learning in the Presence of Spurious Correlations (ICML 2024 · similarity 0.85)
- Probabilistic Stability Guarantees for Feature Attributions (NeurIPS 2025 · similarity 0.84)
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.84)
- Building a stable classifier with the inflated argmax (NeurIPS 2024 · similarity 0.84)
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.
