Research map / Learning theory and algorithms / Kernels and regression theory
Provable Meta-Learning of Linear Representations
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 |
|---|---|---|---|
| Aggregate squared-response covariate moments across tasks and use their top-r eigenspace | Shared linear feature learning with Gaussian covariates · Multi-task linear regression | subspace angle error: \tilde{O} \left(\sqrt{\frac{\tilde{\kappa}}{\tilde{\nu}} \frac{d r}{n_1}} \right) | Helped |
| Transfer the learned representation and fit the new task in its r-dimensional space | New-task linear regression · New linear regression task | parameter recovery error squared: \tilde{O} \left( \delta^2 + \frac{r}{n_2} \right) | Helped |
| Use an inaccurate learned representation when its squared error is larger than the new-task sample benefit | Meta-learned linear regression · New linear regression task with limited feature-learning data | new-task parameter recovery error | Hurt |
| Increase the number of training tasks while keeping n_t=5 | LF-MoM and LF-FO feature learning · Simulation with d=100, r=5, n_t=5 | subspace angle distance | Mixed |
| Transfer learned features when feature learning is poor and n₂=2500 | Meta-LR-FO and meta-LR-MoM · Simulation with d=100, r=5, n_t=5, n₂=2500 | new-task parameter error | Hurt |
| Transfer learned features when new-task data are scarce | Meta-LR-FO and meta-LR-MoM · Simulation with d=100, r=5, n_t=25, n₂=25 | new-task parameter error | Helped |
| Increase samples per training task while holding t=20 | Meta-learned regression · Simulation with d=100, r=5, t=20, n₂=50 | new-task parameter error | Helped |
Closest papers by meaning
- Conditional Meta-Learning of Linear Representations (NeurIPS 2022 · similarity 0.92)
- Subspace Learning for Effective Meta-Learning (ICML 2022 · similarity 0.91)
- Statistically and Computationally Efficient Linear Meta-representation Learning (NeurIPS 2021 · similarity 0.91)
- Meta-learning for Mixed Linear Regression (ICML 2020 · similarity 0.90)
- Learning To Learn Around A Common Mean (NeurIPS 2018 · similarity 0.90)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.88)
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.