Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Subspace Learning for Effective Meta-Learning
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 |
|---|---|---|---|
| Replace single global meta-model with subspace mixture (MUSML, K subspaces) | MAML-style meta-learning · Synthetic 5-shot regression | MSE: 0.74 to 0.07 | Helped |
| Subspace mixture (MUSML) vs global meta-model on real data | Meta-learning · Meta-Dataset-BTAF 5-way 5-shot | accuracy (%): 66.18 vs 57.78 (MAML) / 62.29 (ProtoNet) | Helped |
| Increase number of clusters in TSA-MAML instead of subspace mixture (ablation) | TSA-MAML · Synthetic 5-shot regression | MSE: 0.58 (2 clusters) to 0.12 (20 clusters), then 0.14 (40) and 0.13 (80) | Mixed |
| Use constant temperature instead of linear annealing for softmax subspace selection (ablation) | MUSML · Meta-Dataset-BTAF 5-way 5-shot | accuracy (%): best constant 63.16 (gamma=0.1) vs 66.18 with annealing | Hurt |
| Increase number of subspaces K beyond the selected value (ablation) | MUSML · Meta-Dataset-BTAF 5-way 5-shot | meta-testing accuracy: training accuracy rises with K, but K=5 is not advantageous at meta-validation and meta-testing | Mixed |
| Increase subspace dimension m from 1 to 2, then to 3-5 (ablation) | MUSML · Meta-Dataset-BTAF 5-way 5-shot | meta-validation and meta-testing accuracy: improves from m=1 to m=2; m=3,4,5 worse | Mixed |
| Apply MUSML subspace mixture to Meta-Curvature (ablation) | Meta-Curvature · Meta-Dataset-BTAF 5-way 5-shot | accuracy (%): 60.02 to 66.10 | Helped |
| Apply MUSML subspace mixture to Meta-SGD (ablation) | Meta-SGD · Meta-Dataset-BTAF 5-way 5-shot | accuracy (%): 58.93 to 65.72 | Helped |
| Subspace mixture (MUSML) on Omniglot dataset | MUSML · Meta-Dataset-CIO 5-way 5-shot (Omniglot) | accuracy (%): 95.13 vs 97.92 (BMG) and 96.99 (TSA-MAML) | Hurt |
| Subspace mixture (MUSML) on 1-shot Meta-Dataset-BTAF | MUSML · Meta-Dataset-BTAF 5-way 1-shot | accuracy (%): 50.53 vs 50.34 (ARML), smaller margin than 5-shot | Helped |
| Subspace mixture (MUSML) for cross-domain transfer BTAF to CIO | MUSML · Cross-domain BTAF to CIO 5-way 5-shot | accuracy (%): 67.41 vs 66.92 (TSA-ProtoNet) | Helped |
Closest papers by meaning
- Provable Meta-Learning of Linear Representations (ICML 2021 · similarity 0.91)
- Gradient-Based Meta-Learning with Learned Layerwise Metric and Subspace (ICML 2018 · similarity 0.89)
- A Representation Learning Perspective on the Importance of Train-Validation Splitting in Meta-Learning (ICML 2021 · similarity 0.89)
- Hierarchically Structured Meta-learning (ICML 2019 · similarity 0.89)
- A Sample Complexity Separation between Non-Convex and Convex Meta-Learning (ICML 2020 · similarity 0.89)
- More Flexible PAC-Bayesian Meta-Learning by Learning Learning Algorithms (ICML 2024 · similarity 0.87)
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.