Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples
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 |
|---|---|---|---|
| Initialize the MAML task-specific classifier from class prototypes | fo-MAML (Proto-MAML) · Meta-Dataset | average rank: 3.7 -> 1.85; 4.25 -> 1.5 | Helped |
| Train on all datasets rather than ImageNet only | few-shot classifiers · Omniglot, Quick Draw, Aircraft | accuracy | Helped |
| Train on all datasets rather than ImageNet only | few-shot classifiers · remaining test sources | accuracy | Mixed |
| Initialize the embedding from ImageNet-trained k-NN weights instead of from scratch (ablation) | meta-learners · natural-image test datasets | accuracy | Helped |
| Initialize the embedding from ImageNet-trained k-NN weights instead of from scratch (ablation) | meta-learners · Omniglot and Quick Draw | accuracy | Hurt |
| Meta-train the embedding rather than train non-episodically and use the same inference procedure (ablation) | ProtoNet, MatchingNet, and fo-Proto-MAML inference baselines · Meta-Dataset (ImageNet-only training) | accuracy | Helped |
| Meta-train the embedding rather than train non-episodically and use the same inference procedure (ablation) | ProtoNet, MatchingNet, and fo-Proto-MAML inference baselines · Meta-Dataset (all-dataset training) | accuracy | Mixed |
| Increase the fo-MAML inner-loop learning rate from 0.01 to approximately 0.1 (ablation) | fo-MAML · ILSVRC (ImageNet-only training) | accuracy (%): 36.09 -> 45.51 | Helped |
| Increase the fo-MAML inner-loop learning rate from 0.01 to approximately 0.1 (ablation) | fo-MAML · Omniglot (ImageNet-only training) | accuracy (%): 38.67 -> 55.55 | Helped |
| Increase the fo-MAML inner-loop learning rate from 0.01 to approximately 0.1 (ablation) | fo-MAML · Aircraft (ImageNet-only training) | accuracy (%): 34.50 -> 56.24 | Helped |
| Use more support examples per class | few-shot classifiers · Meta-Dataset test episodes | class precision | Helped |
| Increase the number of classes per episode | few-shot classifiers · Meta-Dataset test episodes | accuracy | Hurt |
Closest papers by meaning
- Unraveling Meta-Learning: Understanding Feature Representations for Few-Shot Tasks (ICML 2020 · similarity 0.92)
- A Theoretical Analysis of the Number of Shots in Few-Shot Learning (ICLR 2020 · similarity 0.91)
- Meta-Learning for Semi-Supervised Few-Shot Classification (ICLR 2018 · similarity 0.91)
- The Role of Global Labels in Few-Shot Classification and How to Infer Them (NeurIPS 2021 · similarity 0.91)
- Improving Generalization via Meta-Learning on Hard Samples (CVPR 2024 · similarity 0.89)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 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.