Research map / Architectures and efficiency / Neural architecture search
Which Model To Transfer? Finding the Needle in the Growing Haystack
Cedric Renggli, André Susano Pinto, Luka Rimanic, Joan Puigcerver, Carlos Riquelme, Ce Zhang, Mario Lučić
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Combine the top task-agnostic model with the top task-aware linear-evaluation model | Model search across the All, ResNet-50, and Expert pools · VTAB-1K | relative regret | Improved |
| Include the task-agnostic pick alongside the linear-evaluation pick | Linear evaluation, ResNet-50 pool · dSpr-Loc | relative regret: 36% | Improved |
| Include the task-agnostic pick alongside the linear-evaluation pick | Linear evaluation, Expert pool · KITTI-Dist | relative regret: 11% | Improved |
| Use linear evaluation rather than task-agnostic ranking | Model search, Expert pool · VTAB-1K | relative regret | Improved |
| Use linear evaluation rather than task-agnostic ranking | Model search, All pool · VTAB-1K | relative regret | Mixed |
| Use linear evaluation rather than task-agnostic ranking | Model search, ResNet-50 or Expert pool · Structured datasets | relative regret | Mixed |
| Use kNN instead of linear evaluation as the task-aware proxy (ablation) | Model search across model pools · VTAB-1K | relative regret | Mixed |
| Use kNN rather than linear evaluation in the hybrid strategy (ablation) | Hybrid model search, All pool · VTAB-1K | relative regret | Improved |
| Use kNN rather than linear evaluation in the hybrid strategy (ablation) | Hybrid model search, Expert pool · VTAB-1K | relative regret | Hurt |
| Replace linear evaluation with early stopping after 10 SGD iterations (ablation) | Task-aware model search · VTAB-1K | relative regret | No clear effect |
| Increase kNN neighborhood size from k=1 to k=3 (ablation) | kNN evaluation · VTAB-1K | accuracy | No clear effect |
| Use ImageNet accuracy ranking instead of task-aware search (ablation) | Model search, ImageNet-accuracy pool · VTAB-1K | number of optimal model picks: 12 out of 19 for B=1; 17 out of 19 for B=2 | Improved |
Closest papers by meaning
- Scalable Transfer Learning with Expert Models (ICLR 2021 · similarity 0.89)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.87)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.87)
- Features are fate: a theory of transfer learning in high-dimensional regression (ICML 2025 · similarity 0.87)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.86)
- Fantastic Gains and Where to Find Them: On the Existence and Prospect of General Knowledge Transfer between Any Pretrained Model (ICLR 2024 · similarity 0.86)
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.
