Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Optimal Features via Partial Invariance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use a recent-time partition (1930-1970) for P-IRM instead of all training years | IRM on House Prices · House Prices, 1970-2010 OoD test | Avg. MSE: 0.522 (0.015) -> 0.427 (0.009) | Helped |
| Use a recent-time partition (1930-1970) for P-IRM instead of all training years | IRM on House Prices · House Prices, 1970-2010 OoD test | Worst Group MSE: 1.129 (0.038) -> 0.873 (0.024) | Helped |
| Partition environments and add nearby samples up to p=10 | ERM on MetaShift · MetaShift dog(shelf) domain generalization | Avg. Performance (accuracy): 0.62425 -> 0.6865 | Helped |
| Apply IRM to a selected partition | IRM on MetaShift · MetaShift dog(shelf) domain generalization | Avg. Performance (accuracy): 0.6695 -> 0.7285 | Helped |
| Add samples from nearby communities to the P-IRM partition (p=10) (ablation) | P-IRM on MetaShift · MetaShift dog(shelf) domain generalization | Avg. Performance (accuracy): 0.7285 -> 0.735 | Helped |
| Partition the training years to 1990-2009 | ERM on SciERC · SciERC NER, 2010-2016 | Accuracy: 0.800 (0.012) -> 0.804 (0.020) | No clear effect |
| Train P-IRM on the 2000-2009 partition | IRM on SciERC · SciERC NER, 2010-2016 | Accuracy: 0.795 (0.005) -> 0.807 (0.005) | Helped |
| Apply the invariance constraint to 1990-2009 while retaining all training data | IRM on SciERC · SciERC NER, 2010-2016 | Accuracy: 0.795 (0.005) -> 0.812 (0.008) | Helped |
| Use P-IRM (partitioned) rather than IRM | IRM on MetaShift · MetaShift subpopulation shift, m=0.12 | Worst Group Acc.: 0.336(0.160) -> 0.604(0.227) | Helped |
| Partition P-IRM and add samples up to p=25 rather than using p=0 (ablation) | P-IRM on MetaShift · MetaShift subpopulation shift, m=0.12 | Worst Group Acc.: 0.560(0.230) -> 0.604(0.227) | Helped |
| Partitioning with little available training data (ablation) | P-IRM on MetaShift · MetaShift subpopulation shift, m=0.01 | Worst Group Acc. | Mixed |
| Use partitioning rather than conditioning in underparameterized linear regression (ablation) | P-IRM on House Prices · House Prices | MSE: 0.427 (0.009) vs 0.490 (0.014) | Helped |
Closest papers by meaning
- The Risks of Invariant Risk Minimization (ICLR 2021 · similarity 0.93)
- Invariant Risk Minimization Is A Total Variation Model (ICML 2024 · similarity 0.91)
- What Is Missing in IRM Training and Evaluation? Challenges and Solutions (ICLR 2023 · similarity 0.91)
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · similarity 0.87)
- The Implicit Bias of Heterogeneity towards Invariance: A Study of Multi-Environment Matrix Sensing (NeurIPS 2024 · similarity 0.86)
- Learning Conditional Invariances through Non-Commutativity (ICLR 2024 · similarity 0.85)
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.