Research map / Learning theory and algorithms / Kernels and regression theory
Meta-learning with differentiable closed-form solvers
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 differentiable ridge regression for per-episode adaptation, trained through meta-learning | few-shot classification · CIFAR-FS and miniImageNet | accuracy: +4.3% relative average accuracy over GNN on the four evaluated settings | Helped |
| Use Woodbury identity to solve using an episode-sized matrix | R2-D2 · few-shot episodes | computation cost | Helped |
| Use LR-D2 with five IRLS iterations rather than one (ablation) | LR-D2 · miniImageNet 5-way 5-shot | accuracy: 65.6±0.2% -> 68.7±0.2% | Helped |
| Use LR-D2 with five IRLS iterations rather than one (ablation) | LR-D2 · CIFAR-FS 5-way 5-shot | accuracy: 75.8±0.2% -> 78.3±0.2% | Helped |
| Use LR-D2 with five IRLS iterations rather than one (ablation) | LR-D2 · miniImageNet 5-way 1-shot | accuracy: 51.0±0.2% -> 51.9±0.2% | Helped |
| Use LR-D2 with five IRLS iterations rather than one (ablation) | LR-D2 · CIFAR-FS 5-way 1-shot | accuracy: 64.5±0.2% -> 65.3±0.2% | Helped |
| Use a 64-channel embedding rather than the full embedding (ablation) | R2-D2 · miniImageNet 5-way 1-shot | accuracy: 51.8±0.2% -> 49.5±0.2% | Hurt |
| Use a 64-channel embedding rather than the full embedding (ablation) | R2-D2 · CIFAR-FS 5-way 1-shot | accuracy: 65.4±0.2% -> 62.3±0.2% | Hurt |
| Learn calibration scale α instead of leaving it fixed (ablation) | R2-D2 · few-shot classification | accuracy | Helped |
| Learn regularization λ instead of leaving it fixed (ablation) | R2-D2 · few-shot classification | accuracy | No clear effect |
| Use a diagonal, per-parameter regularization prior instead of scalar λ (ablation) | R2-D2 · few-shot classification | accuracy | No clear effect |
| Train the representation with adaptation included rather than standard pretraining followed by task-specific final layers (ablation) | R2-D2 feature extractor · miniImageNet and CIFAR-FS | accuracy: -13.8% and -11.6% accuracy on miniImageNet; -11.5% and -5.9% on CIFAR | Helped |
Closest papers by meaning
- Optimization as a Model for Few-Shot Learning (ICLR 2017 · similarity 0.90)
- Meta-Transfer Learning for Few-Shot Learning (CVPR 2019 · similarity 0.88)
- On Enforcing Better Conditioned Meta-Learning for Rapid Few-Shot Adaptation (NeurIPS 2022 · similarity 0.88)
- MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning (AAAI 2024 · similarity 0.86)
- Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective (ICML 2025 · similarity 0.86)
- Neural Fine-Tuning Search for Few-Shot Learning (ICLR 2024 · 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.