Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
On Enforcing Better Conditioned Meta-Learning for Rapid Few-Shot Adaptation
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 a conditioning loss based on variance of log Hessian eigenvalues | MAML · Few-shot classification, first inner-loop adaptation step | accuracy: 20.12±0.06 -> 68.60±0.71 (miniImageNet 5-way 5-shot, ResNet18, Table 4) | Helped |
| Add a conditioning loss based on variance of log Hessian eigenvalues | MAML · Five-shot and one-shot classification after five steps, across the five reported datasets | accuracy: 71.06±0.74 -> 73.28±0.68 (miniImageNet 5-way 5-shot, ResNet18, Table 4) | Helped |
| Use conditioning loss on the classifier rather than no conditioning loss (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 20.04±0.02 -> 61.78±0.75 | Helped |
| Apply the conditioning loss to the embedding layer instead of the classifier (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 59.85±0.80 vs 61.78±0.75 | Mixed |
| Apply the conditioning loss to embedding batch normalization instead of the classifier (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 62.22±0.79 vs 61.78±0.75 | Helped |
| Apply conditioning to classifier and embedding together (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 64.27±0.77 vs 61.78±0.75 | Helped |
| Apply conditioning to classifier and embedding batch normalization together (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 55.54±0.83 vs 61.78±0.75 | Hurt |
| Apply conditioning to embedding and embedding batch normalization together (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 63.46±0.76 vs 61.78±0.75 | Helped |
| Apply conditioning to classifier, embedding, and embedding batch normalization together (ablation) | Conv4 · CUB 5-way 5-shot, step 1 | accuracy: 60.93±0.81 vs 61.78±0.75 | Hurt |
| Train with the conditioning loss, then take one inference step with a larger learning rate | MAML with conditioning constraint · miniImageNet 5-way 5-shot, one inference step | accuracy: 65.96 ± 0.71 -> 66.54±0.70 (5 steps - lr MAML vs 1 step - 5·lr ours) | Helped |
| Continue adaptation beyond the five-step training horizon | Conditioning constraint (ours) · CUB, ResNet18, 5-way 5-shot | accuracy: 84.66±0.54 -> 85.01±0.53 | Helped |
Closest papers by meaning
- Optimization as a Model for Few-Shot Learning (ICLR 2017 · similarity 0.90)
- Efficient Meta Learning via Minibatch Proximal Update (NeurIPS 2019 · similarity 0.90)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.88)
- MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning (AAAI 2024 · similarity 0.88)
- Memory-Reduced Meta-Learning with Guaranteed Convergence (AAAI 2025 · similarity 0.88)
- Task-Specific Preconditioner for Cross-Domain Few-Shot Learning (AAAI 2025 · similarity 0.87)
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.