Research map / Label-efficient and robust learning / Domain adaptation and generalization
Learning to Generalize: Meta-Learning for Domain Generalization
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 |
|---|---|---|---|
| Simulate train/test domain shift within minibatches and optimize for meta-test performance after a meta-train update | AlexNet · PACS Ave. | accuracy: 67.24 -> 70.01 | Helped |
| Set the meta-train step size alpha to zero, removing the meta-optimization adaptation step (ablation) | MLDG on AlexNet · PACS Ave. | accuracy: 70.01 -> 67.93 | Hurt |
| Apply MLDG only to fully connected layers rather than all layers (ablation) | MLDG on AlexNet · PACS Ave. | accuracy: 70.01 -> 68.89 | Hurt |
| Use normalized gradient cosine similarity instead of the vanilla MLDG objective (ablation) | MLDG · PACS Ave. | accuracy: 70.01 -> 68.43 | Hurt |
| Use the meta-test gradient-norm objective instead of the vanilla MLDG objective (ablation) | MLDG · PACS Ave. | accuracy: 70.01 -> 67.48 | Hurt |
| Use the gradient-norm objective variant instead of vanilla MLDG (ablation) | RL-MLDG · Cart-Pole: pole length | return: $165.34\pm 3.38$ -> $175.25 \pm 3.16$ | Helped |
| Use the gradient-cosine objective variant instead of vanilla MLDG (ablation) | RL-MLDG · Cart-Pole: pole length | return: $165.34\pm 3.38$ -> $129.56\pm 2.51$ | Hurt |
| Use the gradient-cosine objective variant instead of vanilla MLDG (ablation) | RL-MLDG · Mountain Car | return (successful cases): $-125.73\pm 2.76$ -> $ -311.80 \pm 3.92$ | Hurt |
| Use the gradient-cosine objective variant instead of vanilla MLDG (ablation) | RL-MLDG · Mountain Car | failure rate: $0.05\pm 0.02$ -> $0.0 \pm 0.0$ | Helped |
| Use the gradient-norm objective variant instead of vanilla MLDG (ablation) | RL-MLDG · Mountain Car | failure rate: $0.05\pm 0.02$ -> $1.0\pm 0.0$ | Hurt |
| Train on a single random source domain rather than aggregating all source domains | RL policy · Cart-Pole: pole length | return: $97.39\pm 73.49$ -> $133.74 \pm 6.79$ | Helped |
| Train on a single random source domain rather than aggregating all source domains | RL policy · Mountain Car | failure rate: $0.05\pm 0.02$ -> $0.55\pm 0.07$ | Hurt |
Closest papers by meaning
- MetaReg: Towards Domain Generalization using Meta-Regularization (NeurIPS 2018 · similarity 0.93)
- Meta-Causal Learning for Single Domain Generalization (CVPR 2023 · similarity 0.91)
- Learning to Generalize across Domains on Single Test Samples (ICLR 2022 · similarity 0.91)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.90)
- Improving Domain Generalization with Domain Relations (ICLR 2024 · similarity 0.89)
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 2025 · similarity 0.89)
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.