Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
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 |
|---|---|---|---|
| train an initialization to perform well after task-specific gradient adaptation | MAML · Sinusoid regression, 5-shot, 1 gradient step | mean squared error: 2.41 -> 0.67 | Helped |
| train an initialization to perform well after task-specific gradient adaptation | MAML · MiniImagenet, 5-way 1-shot | accuracy: 43.44 ± 0.77% -> 48.70 ± 1.84% | Helped |
| omit second derivatives in the meta-gradient (ablation) | MAML · MiniImagenet, 5-way 1-shot | accuracy: 48.70 ± 1.84% -> 48.07 ± 1.75% | No clear effect |
| omit second derivatives in the meta-gradient (ablation) | MAML · MiniImagenet, 5-way 5-shot | accuracy: 63.11 ± 0.92% -> 63.15 ± 0.91% | No clear effect |
| adapt only an input context vector instead of model parameters (ablation) | MAML · Omniglot, 5-way 1-shot | accuracy: 98.7 ± 0.4% -> 94.9 ± 0.9% | Hurt |
| adapt only an input context vector instead of model parameters (ablation) | MAML · Omniglot, 5-way 5-shot | accuracy: 99.9 ± 0.1% -> 97.7 ± 0.3% | Hurt |
| adapt only an input context vector instead of model parameters (ablation) | MAML · 2D Pointmass, 1 gradient step | average return: -11.68 -> -13.90 | Hurt |
| adapt only an input context vector instead of model parameters (ablation) | MAML · 2D Pointmass, 3 gradient steps | average return: -3.23 -> -3.18 | Helped |
| adapt only an input context vector instead of model parameters (ablation) | MAML · Half-cheetah forward/backward, 1 gradient step | average return: 293.19 -> -44.08 | Hurt |
| adapt only an input context vector instead of model parameters (ablation) | MAML · Half-cheetah forward/backward, 3 gradient steps | average return: 315.65 -> -42.50 | Hurt |
| use the MAML initialization rather than conventional pretraining | MAML policy · 2D navigation | adaptation performance | Helped |
| use MAML initialization rather than random initialization or pretraining | MAML policy · Locomotion tasks | adaptation performance | Helped |
Closest papers by meaning
- Bayesian Model-Agnostic Meta-Learning (NeurIPS 2018 · similarity 0.91)
- Meta-Learning with Neural Tangent Kernels (ICLR 2021 · similarity 0.90)
- Probabilistic Model-Agnostic Meta-Learning (NeurIPS 2018 · similarity 0.90)
- Task Agnostic Meta-Learning for Few-Shot Learning (CVPR 2019 · similarity 0.90)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.88)
- First-order ANIL provably learns representations despite overparametrisation (ICLR 2024 · 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.