Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Unraveling Meta-Learning: Understanding Feature Representations for Few-Shot Tasks
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 |
|---|---|---|---|
| Add feature-clustering regularization to classical R2-D2-backbone training | Classical transfer learning · mini-ImageNet, CIFAR-FS (1-shot and 5-shot) | accuracy: 48.39±0.29% -> 50.39±0.30%; 68.24±0.26% -> 69.58 ±0.26%; 62.9±0.3% -> 65.5±0.4%; 82.8±0.3% -> 83.3±0.3% | Helped |
| Add feature-clustering regularization to classical MetaOptNet-backbone training | Classical transfer learning · mini-ImageNet, CIFAR-FS (1-shot and 5-shot) | accuracy: 56.18±0.31% -> 59.38±0.31%; 76.72±0.24% -> 78.15±0.24%; 69.5±0.3% -> 72.3±0.4%; 85.7±0.2% -> 86.3±0.2% | Helped |
| Use hyperplane-variation regularization instead of feature-clustering regularization (ablation) | Classical R2-D2-backbone training · mini-ImageNet 1-shot and 5-shot | accuracy: 48.39±0.29% -> 50.16±0.30%; 68.24±0.26% -> 69.54±0.26% | Helped |
| Add feature-clustering regularization | Classical MetaOptNet-backbone training · CIFAR-FS 5-shot | accuracy: 85.7±0.2% -> 86.3±0.2% | Helped |
| Add consensus weight-clustering regularization to Reptile | Reptile · mini-ImageNet 1-shot and 5-shot | accuracy: 49.97±0.32% -> 51.94 ±0.23%; 65.99±0.58% -> 68.02±0.22% | Helped |
| Increase the weight-clustering coefficient to 2.5×10^-4 (ablation) | Reptile with Weight-Clustering · mini-ImageNet 5-shot | accuracy: 65.99±0.58% -> 65.95±0.23% | No clear effect |
| Use feature-clustering training rather than classical training | R2-D2-backbone transfer learning · CIFAR-FS 5-shot | accuracy: 82.8±0.3% -> 83.3±0.3% | Helped |
| Use meta-learned rather than classically trained feature representations | MetaOptNet and R2-D2 feature extractors · 1-shot mini-ImageNet | accuracy: MetaOptNet-M vs MetaOptNet-C: 62.64±0.31% vs 56.18±0.31%; R2-D2-M vs R2-D2-C: 51.80±0.20% vs 48.39±0.29% (SVM) | Helped |
| Add the feature-clustering regularizer to classical training (ablation) | R2-D2 and MetaOptNet backbones · mini-ImageNet and CIFAR-FS | runtime: R2-D2 mini-ImageNet: 20s/1.7h -> 20s/1.7h; MetaOptNet mini-ImageNet: 1.4m/7.0h -> 1.5m/7.4h | No clear effect |
| Use weight-clustering regularization instead of vanilla Reptile (ablation) | Reptile · mini-ImageNet 1-shot and 5-shot | accuracy: 49.97±0.32% -> 51.94 ±0.23%; 65.99±0.58% -> 68.02±0.22% | Helped |
Closest papers by meaning
- Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples (ICLR 2020 · similarity 0.92)
- Meta-Learning With Differentiable Convex Optimization (CVPR 2019 · similarity 0.91)
- The Role of Global Labels in Few-Shot Classification and How to Infer Them (NeurIPS 2021 · similarity 0.91)
- A Theoretical Analysis of the Number of Shots in Few-Shot Learning (ICLR 2020 · similarity 0.91)
- On sensitivity of meta-learning to support data (NeurIPS 2021 · similarity 0.90)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.89)
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.