Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
A Representation Learning Perspective on the Importance of Train-Validation Splitting in Meta-Learning
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 |
|---|---|---|---|
| Use train-validation rather than train-train meta-training | RepLearn and iMAML · Omniglot 5-way 1-shot | accuracy: 89.00 $\pm$ 0.01 -> 97.30 $\pm$ 0.01; 49.20 $\pm$ 1.91 -> 97.90 $\pm$ 0.58 | Helped |
| Use train-validation rather than train-train meta-training | RepLearn and iMAML · Omniglot 20-way 1-shot | accuracy: 88.20 $\pm$ 0.01 -> 92.30 $\pm$ 0.01; 18.30 $\pm$ 0.67 -> 91.00 $\pm$ 0.54 | Helped |
| Use train-validation objective instead of train-train objective | Linear representation learning · Simulation test tasks with 5, 15, and 25 samples | loss: 1.04 $\pm$ 0.08 -> 0.72 $\pm$ 0.06; 0.85 $\pm$ 0.06 -> 0.40 $\pm$ 0.03; 0.82 $\pm$ 0.06 -> 0.38 $\pm$ 0.03 | Helped |
| Use train-validation rather than train-train training | RepLearn · MiniImageNet 5-way 1-shot and 5-way 5-shot | accuracy: 25.53 $\pm$ 1.43 -> 46.16 $\pm$ 1.67; 33.49 $\pm$ 0.82 -> 65.36 $\pm$ 0.91 | Helped |
| Increase FCN width from $64^*1$ to $64^*32$ (ablation) | RepLearn tr-val · Omniglot 5-way 1-shot | accuracy: 87.8 $\pm$ 1.1 -> 91.6 $\pm$ 1.0 | Helped |
| Increase FCN width from $64^*1$ to $64^*32$ (ablation) | RepLearn tr-tr · Omniglot 5-way 1-shot | accuracy: 80.6 $\pm$ 1.4 -> 67.8 $\pm$ 1.6 | Hurt |
| Increase CNN capacity from $num\_filters^*0.5$ to $num\_filters^*8$ (ablation) | RepLearn tr-val · MiniImageNet 5-way 1-shot | accuracy: 46.66 $\pm$ 1.69 -> 52.25 $\pm$ 1.71 | Helped |
| Increase CNN capacity from $num\_filters^*0.5$ to $num\_filters^*8$ (ablation) | RepLearn tr-tr · MiniImageNet 5-way 1-shot | accuracy: 26.25 $\pm$ 1.45 -> 25.28 $\pm$ 1.37 | Hurt |
| Add Frobenius regularization to the representation (tr-tr $\lambda=0$ to $\lambda=1.0$) (ablation) | RepLearn · MiniImageNet 5-way 5-shot | accuracy: 33.49 $\pm$ 0.82 -> 40.19 $\pm$ 1.12 | Helped |
| Add Frobenius regularization to the representation (tr-tr $\lambda=0$ to $\lambda=1.0$) (ablation) | RepLearn · MiniImageNet 5-way 1-shot | accuracy: 25.53 $\pm$ 1.43 -> 25.88 $\pm$ 1.45 | No clear effect |
| Increase representation capacity while retaining tr-val training (ablation) | FCN representations · Omniglot supervised 413-way | accuracy: 87.4 -> 100.0 | No clear effect |
Closest papers by meaning
- How Important is the Train-Validation Split in Meta-Learning? (ICML 2021 · similarity 0.92)
- Understanding Train-Validation Split in Meta-Learning with Neural Networks (ICLR 2023 · similarity 0.91)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.87)
- Improving Generalization via Meta-Learning on Hard Samples (CVPR 2024 · similarity 0.86)
- First-order ANIL provably learns representations despite overparametrisation (ICLR 2024 · similarity 0.86)
- Learning to Learn with Contrastive Meta-Objective (NeurIPS 2025 · similarity 0.85)
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.