Research map / Architectures and efficiency / RNNs and normalization
Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
Polina Kirichenko, Pavel Izmailov, Andrew Gordon Wilson
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 |
|---|---|---|---|
| Retrain the last layer on group-balanced validation data | ERM feature extractor · Waterbirds | worst-group accuracy: 74.9 -> 92.9 | Improved |
| Retrain the last layer on group-balanced validation data | ERM feature extractor · CelebA | worst-group accuracy: 46.9 -> 88.3 | Improved |
| Retrain the last layer on group-balanced validation data | ERM feature extractor · MultiNLI | worst-group accuracy: 65.9 -> 74.7 | Improved |
| Retrain the last layer on group-balanced validation data | ERM feature extractor · CivilComments | worst-group accuracy: 55.6 -> 70.1 | Improved |
| Use held-out validation data rather than training data for last-layer retraining (ablation) | DFR · Waterbirds | worst-group accuracy: 90.2 -> 92.9 | Improved |
| Use held-out validation data rather than training data for last-layer retraining (ablation) | DFR · CelebA | worst-group accuracy: 80.7 -> 88.3 | Improved |
| Retrain only the last linear layer instead of the last two layers (ablation) | DFR · Waterbirds | worst-group accuracy: 93.1 -> 93.1 | No clear effect |
| Retrain only the last linear layer instead of the last residual block (ablation) | DFR · Waterbirds | worst-group accuracy: 90.7 -> 93.1 | Improved |
| Retrain only the last linear layer instead of all layers (ablation) | DFR · Waterbirds | worst-group accuracy: 35.6 -> 93.1 | Improved |
| Retrain only the last linear layer instead of the last two layers (ablation) | DFR · CelebA | worst-group accuracy: 87.2 -> 88.3 | Improved |
| Retrain only the last linear layer instead of the last residual block (ablation) | DFR · CelebA | worst-group accuracy: 83.0 -> 88.3 | Improved |
| Retrain only the last linear layer instead of all layers (ablation) | DFR · CelebA | worst-group accuracy: 77.8 -> 88.3 | Improved |
Closest papers by meaning
- On Feature Learning in the Presence of Spurious Correlations (NeurIPS 2022 · similarity 0.87)
- Complexity Matters: Feature Learning in the Presence of Spurious Correlations (ICML 2024 · similarity 0.86)
- ImageNet-E: Benchmarking Neural Network Robustness via Attribute Editing (CVPR 2023 · similarity 0.85)
- Severing Spurious Correlations with Data Pruning (ICLR 2025 · similarity 0.85)
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.85)
- Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks (ICLR 2025 · 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.
