Research map / Learning theory and algorithms / Sample complexity
Is Spurious Correlation Removal Always Learnable?
Yibo Zhou, Bo Li, Hai-Miao Hu, Hanzi Wang, Xiaokang Zhang, Ruifan Zhang
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 |
|---|---|---|---|
| increase environment diversity gamma from 0.1 to 0.6 (ablation) | exhaustive/invariance feature selection · Synthetic (d=20,k=4) | feature selection accuracy: .20 -> 1.0 (+.80) | Improved |
| increase samples per environment from 200 to 2000 (ablation) | feature selection · Synthetic | feature selection accuracy: .70 -> 1.0 (+.30) | Improved |
| reduce invariant dimension k from 8 to 3 (ablation) | feature selection · Synthetic | feature selection accuracy: .73 -> 1.0 (+.27) | Improved |
| increase SNR from 1 to 10 (ablation) | feature selection · Synthetic | feature selection accuracy: .75 -> .93 (+.18) | Improved |
| increase training size from N=500 to N=2000 (IRM gap to oracle) | IRM · ColoredMNIST (gamma=0.4) | gap vs oracle: 0.100 -> 0.020 | Improved |
| increase diversity from low to high (ablation) | IRM · CMNIST | worst-group accuracy: .61 -> .84 | Improved |
| increase diversity from low to high (ablation) | ERM · CMNIST | worst-group accuracy: .54 -> .76 | Improved |
| reduce shared spurious bias (high diversity) (ablation) | GDRO · CelebA | worst-group accuracy: .85 -> .91 | Improved |
| reduce shared spurious bias (high diversity) (ablation) | GDRO · Waterbirds | worst-group accuracy: .79 -> .87 | Improved |
| use group-aware GDRO instead of IRM at low diversity | GDRO · CelebA | worst-group accuracy: IRM .82 vs GDRO .85 | Improved |
| use IRM instead of ERM at low diversity | IRM · ColoredMNIST | worst-group accuracy: ERM .54 vs IRM .61 | Improved |
| use IRM instead of ERM at low diversity | IRM · CelebA | overall accuracy: .90 vs .89 | No clear effect |
Closest papers by meaning
- The Implicit Bias of Heterogeneity towards Invariance: A Study of Multi-Environment Matrix Sensing (NeurIPS 2024 · similarity 0.85)
- Sparsely Supervised Diffusion (NeurIPS 2026 · similarity 0.85)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.84)
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · similarity 0.84)
- On Sparse Canonical Correlation Analysis (NeurIPS 2024 · similarity 0.84)
- The Price of Sparsity: Sufficient Conditions for Sparse Recovery using Sparse and Sparsified Measurements (NeurIPS 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.
