Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Meta-Learning With Differentiable Convex Optimization
Check 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 a discriminative linear SVM instead of a nearest-class-mean classifier with high-dimensional features | MetaOptNet with ResNet-12 embeddings · miniImageNet and tieredImageNet 5-way few-shot classification | accuracy: 59.25 ± 0.64 -> 62.64 ± 0.61 (miniImageNet 1-shot); 61.74 ± 0.77 -> 65.99 ± 0.72 (tieredImageNet 1-shot) | Helped |
| Use a discriminative linear SVM instead of nearest-class-mean classification with a low-dimensional 4-layer CNN | MetaOptNet · miniImageNet 5-way few-shot classification | accuracy: 53.47 ± 0.63 -> 52.87 ± 0.57 (1-shot); 70.68 ± 0.49 -> 68.76 ± 0.48 (5-shot) | Hurt |
| Replace the SVM base learner with ridge regression | MetaOptNet · CIFAR-FS 5-way few-shot classification | accuracy: 72.0 ± 0.7 -> 72.6 ± 0.7 (1-shot); 84.2 ± 0.5 -> 84.3 ± 0.5 (5-shot) | Mixed |
| Train on the union of meta-training and meta-validation classes (ablation) | MetaOptNet-SVM · miniImageNet, CIFAR-FS, and FC100 few-shot classification | accuracy: 62.64 ± 0.61 -> 64.09 ± 0.62 (miniImageNet 1-shot); 41.1 ± 0.6 -> 47.2 ± 0.6 (FC100 1-shot) | Helped |
| Train on the union of meta-training and meta-validation classes (ablation) | MetaOptNet-SVM · tieredImageNet few-shot classification | accuracy: 65.99 ± 0.72 -> 65.81 ± 0.74 (1-shot); 81.56 ± 0.53 -> 81.75 ± 0.53 (5-shot) | No clear effect |
| Add data augmentation (ablation) | MetaOptNet-SVM with ResNet-12 · miniImageNet 5-way few-shot classification | accuracy: 51.13 -> 55.80 (1-shot); 70.88 -> 75.76 (5-shot) | Helped |
| Add weight decay after data augmentation (ablation) | MetaOptNet-SVM with ResNet-12 · miniImageNet 5-way few-shot classification | accuracy: 55.80 -> 60.33 (1-shot); 75.76 -> 76.61 (5-shot) | Helped |
| Add DropBlock after data augmentation and weight decay (ablation) | MetaOptNet-SVM with ResNet-12 · miniImageNet 5-way few-shot classification | accuracy: 60.33 -> 61.11 (1-shot); 76.61 -> 77.40 (5-shot) | Helped |
| Add label smoothing alongside data augmentation, weight decay, and DropBlock (ablation) | MetaOptNet-SVM with ResNet-12 · miniImageNet 5-way few-shot classification | accuracy: 61.11 -> 62.64 (1-shot); 77.40 -> 78.63 (5-shot) | Helped |
| Add the merged meta-training and meta-validation dataset to existing regularization (ablation) | MetaOptNet-SVM with ResNet-12 · miniImageNet 5-way few-shot classification | accuracy: 62.64 -> 64.09 (1-shot); 78.63 -> 80.00 (5-shot) | Helped |
| Increase the number of meta-training shots above the meta-testing shot count (ablation) | MetaOptNet-SVM · miniImageNet 5-way few-shot classification | accuracy | Helped |
| Stop the QP solver after three iterations rather than running it to convergence (ablation) | MetaOptNet-SVM · few-shot classification | accuracy | Helped |
Closest papers by meaning
- Unraveling Meta-Learning: Understanding Feature Representations for Few-Shot Tasks (ICML 2020 · similarity 0.91)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.90)
- Statistically and Computationally Efficient Linear Meta-representation Learning (NeurIPS 2021 · similarity 0.90)
- Optimization as a Model for Few-Shot Learning (ICLR 2017 · similarity 0.90)
- Few-Shot Learning With Localization in Realistic Settings (CVPR 2019 · similarity 0.90)
- Meta-Learning for Semi-Supervised Few-Shot Classification (ICLR 2018 · 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.