Research map / Learning theory and algorithms / Prediction and decision losses
The Risks of Invariant Risk Minimization
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 IRM with a linear featurizer when the number of training environments exceeds the environmental feature dimension | IRM · Linear classification under distribution shift | worst-case risk | Helped |
| Use IRM with a linear featurizer when the number of training environments is at most the environmental feature dimension | IRM · Linear classification on training environments | logistic risk and 0-1 risk | Hurt |
| Use a feasible predictor that relies only on environmental features | IRM · Linear classification on training environments | 0-1 risk | Hurt |
| Use the constructed nonlinear featurizer that matches the invariant predictor on most training inputs but switches to ERM outside that region | Penalized IRM · Nonlinear classification, training environments | gradient penalty: O(p_epsilon^2 (c_epsilon d_e + average squared norm of environmental means)) | Hurt |
| Use the constructed nonlinear featurizer under a sufficiently shifted test distribution | Penalized IRM · Nonlinear classification, shifted test environments | 0-1 risk: greater than .975 - q | Hurt |
| Reverse the correlation between environmental features and labels at test time | IRM and ERM · Synthetic linear classification, shifted test environments | accuracy | Hurt |
| Increase the number of observed environments beyond the environmental feature dimension | IRM · Synthetic linear classification, shifted test environments | accuracy | Helped |
| Tune the penalty coefficient and optimizer learning rate carefully | IRM · Synthetic linear classification | stability | Helped |
Closest papers by meaning
- Learning Optimal Features via Partial Invariance (AAAI 2023 · similarity 0.93)
- Bayesian Invariant Risk Minimization (CVPR 2022 · similarity 0.92)
- Invariant Risk Minimization Is A Total Variation Model (ICML 2024 · similarity 0.92)
- Out-of-distribution Generalization for Total Variation based Invariant Risk Minimization (ICLR 2025 · similarity 0.86)
- Diagnosing and Rectifying Fake OOD Invariance: A Restructured Causal Approach (AAAI 2024 · similarity 0.86)
- Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization (NeurIPS 2025 · 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.