Research map / Label-efficient and robust learning / Recognition and long-tail
Spurious Correlation-Aware Embedding Regularization for Worst-Group Robustness
Subeen Park, JOOWANG KIM, Hakyung Lee, Sunjae yoo, Kyungwoo Song
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 embedding regularization (SCER) to worst-group training | ResNet-50 on Waterbirds · Waterbirds | worst-group accuracy: 90.3 -> 91.2 (vs PDE); ERM 69.1 | Improved |
| add SCER | ResNet-50 on CelebA · CelebA | worst-group accuracy: 91.4 -> 91.4 (vs ElRep) | No clear effect |
| add SCER | ResNet-50 on ColorMNIST (rho=80%) · ColorMNIST (rho=80%) | worst-group accuracy: 73.2 -> 73.6 (vs LISA) | No clear effect |
| add SCER | ResNet-50 on ColorMNIST (rho=99%) · ColorMNIST (rho=99%) | worst-group accuracy: 47.5 -> 56.0 (vs PDE) | Improved |
| add SCER when one subpopulation is absent from training | ResNet-50 on ColorMNIST · ColorMNIST (one subpopulation omitted) | worst-group accuracy: 44.1 -> 59.6 (vs GroupDRO) | Improved |
| use SCER as second stage after EIIL environment inference | ColorMNIST · ColorMNIST (two train envs) | average accuracy: 68.2 -> 72.6 (vs EIIL + DRO) | Improved |
| add SCER to BERT | BERT on CivilComments · CivilComments | worst-group accuracy: 73.7 -> 74.0 (vs LISA) | No clear effect |
| add SCER | BERT on MultiNLI · MultiNLI | worst-group accuracy: 76.0 -> 76.8 (vs GroupDRO) | Improved |
| add core and spurious embedding losses (ablation) | SCER · ColorMNIST (rho=95%) | worst-group accuracy: 70.7 -> 72.8 | Improved |
| add core loss alone (lambda_core 0.5) (ablation) | SCER · ColorMNIST (rho=95%) | worst-group accuracy: 71.6 -> 72.2 | Improved |
| increase spurious loss weight from 0.5 to 1.0 (ablation) | SCER · ColorMNIST (rho=95%) | worst-group accuracy: 71.6 -> 72.8 | Improved |
| use Sigma-norm (covariance-based) instead of Euclidean norm (ablation) | SCER · ColorMNIST (rho=95%) | worst-group accuracy: 70.0 -> 72.8 | Improved |
Closest papers by meaning
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.87)
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.87)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.87)
- Calibrating Multi-modal Representations: A Pursuit of Group Robustness without Annotations (CVPR 2024 · similarity 0.87)
- Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference (ICML 2024 · similarity 0.87)
- ExMap: Leveraging Explainability Heatmaps for Unsupervised Group Robustness to Spurious Correlations (CVPR 2024 · similarity 0.86)
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.
