Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML
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 |
|---|---|---|---|
| Restrict inner-loop updates to the task-specific head (ANIL) | MAML · Omniglot and MiniImageNet few-shot classification; HalfCheetah-Direction, HalfCheetah-Velocity, and 2D-Navigation | accuracy or average return: Omniglot-20way-1shot: 93.7 $\pm$ 0.7 -> 96.2 $\pm$ 0.5; Omniglot-20way-5shot: 96.4 $\pm$ 0.1 -> 98.0 $\pm$ 0.3; MiniImag | No clear effect |
| Restrict inner-loop updates to the task-specific head (ANIL) | MAML · HalfCheetah-Direction | average return: 170.4 $\pm$ 21.0 -> 363.2 $\pm$ 14.8 | Helped |
| Restrict inner-loop updates to the task-specific head (ANIL) | MAML · HalfCheetah-Velocity and 2D-Navigation | average return: HalfCheetah-Velocity: -139.0 $\pm$ 18.9 -> -120.9 $\pm$ 6.3; 2D-Navigation: -20.3 $\pm$ 3.2 -> -20.1 $\pm$ 2.3 | No clear effect |
| Restrict inner-loop updates to the task-specific head (ANIL) | MAML · MiniImageNet training and inference | speedup: Training: 1.79; 1.84. Inference: 4.15; 4.87. | Helped |
| Remove the head at test time and classify using cosine similarity of frozen features (NIL) (ablation) | ANIL-trained network · Omniglot and MiniImageNet few-shot classification | accuracy: Omniglot-20way-1shot: ANIL 96.2 $\pm$ 0.5 -> NIL 96.7 $\pm$ 0.3; Omniglot-20way-5shot: 98.0 $\pm$ 0.3 -> 98.0 $\pm$ 0.04 | No clear effect |
| Freeze all four convolutional body layers during test-time adaptation (ablation) | MAML · MiniImageNet-5way-1shot and MiniImageNet-5way-5shot | accuracy: 1-shot: 46.9 $\pm$ 0.2 -> 46.3 $\pm$ 0.4; 5-shot: 63.1 $\pm$ 0.4 -> 61.0 $\pm$ 0.6 | No clear effect |
| Train the body using task-specific MAML/ANIL objectives rather than multiclass classification | Few-shot feature learning · MiniImageNet-5way-1shot and MiniImageNet-5way-5shot | accuracy: MAML training-NIL head: 48.4 $\pm$ 0.3 vs Multiclass training-NIL head: 39.7 $\pm$ 0.3; 61.5 $\pm$ 0.8 vs 54.4 $\pm$ 0.5 | Helped |
| Train jointly across tasks without a task-specific head | Few-shot feature learning · MiniImageNet-5way-1shot and MiniImageNet-5way-5shot | accuracy: Multitask training-NIL head: 26.5 $\pm$ 1.1 and 34.2 $\pm$ 3.5; Random features-NIL head: 32.9 $\pm$ 0.6 and 43.2 $\pm$ | Hurt |
| Use NIL training rather than task-specific MAML/ANIL training (ablation) | Few-shot feature learning · MiniImageNet-5way-1shot and MiniImageNet-5way-5shot | accuracy: NIL training-NIL head: 38.3 $\pm$ 0.6 and 43.0 $\pm$ 0.2; ANIL training-NIL head: 48.0 $\pm$ 0.7 and 62.2 $\pm$ 0.5 | Hurt |
| Replace the NIL test-time head with a task-adapted MAML/ANIL head (ablation) | MAML/ANIL trained features · MiniImageNet-5way-1shot and MiniImageNet-5way-5shot | accuracy: MAML training: 48.4 $\pm$ 0.3 -> 46.9 $\pm$ 0.2 (1-shot), 61.5 $\pm$ 0.8 -> 63.1 $\pm$ 0.4 (5-shot); ANIL training: 48.0 | No clear effect |
Closest papers by meaning
- MAML is a Noisy Contrastive Learner in Classification (ICLR 2022 · similarity 0.91)
- BOIL: Towards Representation Change for Few-shot Learning (ICLR 2021 · similarity 0.91)
- How to Train Your MAML to Excel in Few-Shot Classification (ICLR 2022 · similarity 0.90)
- MAML and ANIL Provably Learn Representations (ICML 2022 · similarity 0.90)
- Sharp-MAML: Sharpness-Aware Model-Agnostic Meta Learning (ICML 2022 · similarity 0.90)
- Meta-Learning with Implicit Gradients (NeurIPS 2019 · similarity 0.90)
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.