Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
How to Train Your MAML to Excel in Few-Shot Classification
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 |
|---|---|---|---|
| Increase inner-loop updates from 5 to 15 (ablation) | MAML · MiniImageNet 5-way 1-shot | accuracy: 62.90% -> 64.42% | Helped |
| Increase inner-loop updates from 5 to 15 (ablation) | MAML · TieredImageNet 5-way 1-shot | accuracy: 59.08% -> 65.72% | Helped |
| Initialize every class head vector from one shared meta-trained vector | MAML · MiniImageNet 5-way 1-shot | accuracy: 64.42 ± 0.20 -> 65.17 ± 0.20 | Helped |
| Initialize every class head vector from one shared meta-trained vector | MAML · MiniImageNet 5-way 5-shot | accuracy: 83.44 ± 0.14 -> 84.30 ± 0.14 | Helped |
| Initialize every class head vector from one shared meta-trained vector | MAML · TieredImageNet 5-way 1-shot | accuracy: 65.72 ± 0.20 -> 69.24 ± 0.20 | Helped |
| Initialize every class head vector from one shared meta-trained vector | MAML · TieredImageNet 5-way 5-shot | accuracy: 84.37 ± 0.16 -> 86.06 ± 0.16 | Helped |
| Choose permutations using support-set accuracy or loss (ablation) | MAML · MiniImageNet and TieredImageNet 5-way 1/5-shot | accuracy: 64.42 ± 0.20; 83.44 ± 0.13; 65.06 ± 0.20; 84.37 ± 0.16; 64.67 ± 0.20; 84.05 ± 0.13; 65.43 ± 0.20; 84.22 ± 0.16 | No clear effect |
| Ensemble updated predictions over all permutations (ablation) | MAML · MiniImageNet 5-way 1/5-shot | accuracy: 64.42 -> 65.50; 83.44 -> 84.43 | Helped |
| Ensemble updated predictions over rotated permutations (ablation) | MAML · MiniImageNet 5-way 1/5-shot | accuracy: 64.42 -> 65.37; 83.44 -> 84.40 | Helped |
| Average classifier initialization vectors during meta-testing (ablation) | MAML · MiniImageNet and TieredImageNet 5-way 1/5-shot | accuracy: 64.40 ± 0.21; 84.24 ± 0.13; 66.24 ± 0.24; 84.52 ± 0.16 | Mixed |
| Freeze rather than adapt the feature extractor during inner-loop updates (ablation) | unicorn-MAML · MiniImageNet and TieredImageNet 5-way 1/5-shot | accuracy | Mixed |
| Use the shared-vector head initialization rather than separate learned class vectors | MAML · CUB 5-way 1/5-shot | accuracy: 77.67 ± 0.20 -> 78.07 ± 0.20; 90.35 ± 0.16 -> 91.67 ± 0.16 | Helped |
Closest papers by meaning
- How to train your MAML (ICLR 2019 · similarity 0.93)
- Repurposing Pretrained Models for Robust Out-of-domain Few-Shot Learning (ICLR 2021 · similarity 0.91)
- Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML (ICLR 2020 · similarity 0.90)
- MAML is a Noisy Contrastive Learner in Classification (ICLR 2022 · similarity 0.90)
- Hypernetwork Approach to Bayesian MAML (Student Abstract) (AAAI 2025 · similarity 0.86)
- Enabling Few-Shot Learning with PID Control: A Layer Adaptive Optimizer (ICML 2024 · similarity 0.86)
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.