Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Repurposing Pretrained Models for Robust Out-of-domain Few-Shot 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 the full method instead of vanilla MAML | Pretrained MAML checkpoint with SGD · 5-way 5-shot, Mini/Birds/Flowers/Signs average | All accuracy: 40.34\tstd{0.39} -> 46.42\tstd{0.44} | Helped |
| Use the full method instead of vanilla MAML | Pretrained MAML checkpoint with SGD · 5-way 1-shot, Mini/Birds/Flowers/Signs average | All accuracy: 35.50\tstd{0.49} -> 35.42\tstd{0.44} | No clear effect |
| Use the full method instead of vanilla MAML | Pretrained MAML checkpoint with SGD · 5-way 1-shot, Mini/Birds/Flowers/Signs average | Top-40% accuracy: 42.94\tstd{0.70} -> 44.52\tstd{0.68} | Helped |
| Add adversarial training only (ablation) | MAML test-time adaptation · 5-way 1-shot, Mini/Birds/Flowers/Signs average | Top-1 accuracy: 44.54\tstd{0.77} -> 45.38\tstd{0.74} | Helped |
| Add adversarial training only (ablation) | MAML test-time adaptation · 5-way 1-shot, Mini/Birds/Flowers/Signs average | All accuracy: 35.50\tstd{0.49} -> 33.91\tstd{0.41} | Hurt |
| Add USA (uncertainty-based layer-wise stepsizes) (ablation) | MAML test-time adaptation · 5-way 1-shot, Mini/Birds/Flowers/Signs average | Top-40% accuracy: 42.94\tstd{0.70} -> 43.22\tstd{0.43} | No clear effect |
| Add UFGSM to USA (ablation) | MAML test-time adaptation · 5-way 1-shot, Mini/Birds/Flowers/Signs average | All accuracy: 36.01\tstd{0.36} -> 37.05\tstd{0.27} | Helped |
| Add UFGSM rather than ordinary FGSM (ablation) | MAML with USA · 5-way 1-shot, Mini/Birds/Flowers/Signs average | Top-1 accuracy: 42.93\tstd{0.47} -> 44.73\tstd{0.41} | Helped |
| Use higher stepsizes on high-uncertainty layers (inverse USA) (ablation) | MAML test-time adaptation · 5-way 1-shot, Mini/Birds/Flowers/Signs average | All accuracy: 36.01 -> 28.03 | Hurt |
| Use less perturbation on high-uncertainty input gradients (inverse UFGSM) (ablation) | MAML with USA · 5-way 1-shot, Mini/Birds/Flowers/Signs average | All accuracy: 37.05 -> 35.98 | Hurt |
| Add adversarial training and ensemble augmentation to USA+UFGSM (ablation) | MAML test-time adaptation · 5-way 1-shot, Mini/Birds/Flowers/Signs average | Top-40% accuracy: 43.83\tstd{0.39} -> 44.52\tstd{0.68} | Helped |
| Freeze BatchNorm scale and shift parameters during adaptation (ablation) | SGD+All · MAML test-time adaptation, especially base stepsizes >0.1 | All accuracy | Helped |
Closest papers by meaning
- How to train your MAML (ICLR 2019 · similarity 0.91)
- How to Train Your MAML to Excel in Few-Shot Classification (ICLR 2022 · similarity 0.91)
- Global Convergence of MAML and Theory-Inspired Neural Architecture Search for Few-Shot Learning (CVPR 2022 · similarity 0.91)
- On Fast Adversarial Robustness Adaptation in Model-Agnostic Meta-Learning (ICLR 2021 · similarity 0.90)
- Sharp-MAML: Sharpness-Aware Model-Agnostic Meta Learning (ICML 2022 · similarity 0.89)
- Mixture of Adversarial LoRAs: Boosting Robust Generalization in Meta-Tuning (NeurIPS 2024 · 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.