Research map / Label-efficient and robust learning / Domain adaptation and generalization
Learning What and Where to Transfer
Yunhun Jang, Hankook Lee, Sung Ju Hwang, Jinwoo Shin
Read the paperCheck 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 |
|---|---|---|---|
| learn sample-conditioned channel and layer-transfer weights jointly | L2T-ww · CUB200, MIT67, Stanford40, Stanford Dogs | accuracy: 58.90 ± 1.32 -> 65.05 ± 1.19 (CUB200); 61.42 ± 1.68 -> 64.85 ± 2.75 (MIT67); 60.20 ± 1.34 -> 63.08 ± 0.88 (Stanford40); | Improved |
| learn sample-conditioned channel importance rather than using unweighted feature matching | L2T-w · CUB200, MIT67, Stanford40 | accuracy: 47.60 ± 0.31 -> 51.95 ± 0.83 (CUB200); 55.15 ± 0.93 -> 60.41 ± 0.37 (MIT67); 42.93 ± 1.48 -> 46.25 ± 3.66 (Stanford40) | Improved |
| learn layer associations over all candidate pairs instead of using hand-crafted one-to-one matching | L2T-ww · CIFAR-100, STL-10, CUB200, MIT67, Stanford40, Stanford Dogs | accuracy: 70.02 ± 0.19 -> 70.96 ± 0.61 (CIFAR-100); 76.42 ± 0.52 -> 78.31 ± 0.21 (STL-10); 56.61 ± 0.20 -> 65.05 ± 1.19 (CUB200); | Improved |
| use meta-networks to produce sample-specific transfer weights instead of constant meta-weights (ablation) | L2T-ww · CUB200 | accuracy: 61.75 -> 65.05 | Improved |
| use meta-networks to produce sample-specific transfer weights instead of constant meta-weights (ablation) | L2T-ww · MIT67 | accuracy: 64.10 -> 64.85 | Improved |
| use meta-networks to produce sample-specific transfer weights instead of constant meta-weights (ablation) | L2T-ww · Stanford40 | accuracy: 58.88 -> 63.08 | Improved |
| use the proposed three-stage bilevel scheme instead of the original scheme (ablation) | L2T-ww · CUB200 | accuracy: 35.38 -> 65.05 | Improved |
| use the proposed three-stage bilevel scheme instead of the original scheme (ablation) | L2T-ww · MIT67 | accuracy: 54.18 -> 64.85 | Improved |
| use the proposed three-stage bilevel scheme instead of the original scheme (ablation) | L2T-ww · Stanford40 | accuracy: 53.47 -> 63.08 | Improved |
| transfer from TinyImageNet and CIFAR-10 sources simultaneously instead of TinyImageNet alone | L2T-ww · STL-10 | accuracy: 78.31 ± 0.21 -> 80.52 ± 0.29 | Improved |
| reduce the target training set to 50 samples per class while using L2T-ww rather than the strongest listed baseline | L2T-ww · CIFAR-10, 50 samples per class | accuracy: 53.76% -> 64.91% | Improved |
Closest papers by meaning
- Meta-Transfer Learning Through Hard Tasks (TPAMI 2020 · similarity 0.89)
- Meta-Transfer Learning for Few-Shot Learning (CVPR 2019 · similarity 0.88)
- Auto-Transfer: Learning to Route Transferable Representations (ICLR 2022 · similarity 0.87)
- What is being transferred in transfer learning? (NeurIPS 2020 · similarity 0.87)
- Fantastic Gains and Where to Find Them: On the Existence and Prospect of General Knowledge Transfer between Any Pretrained Model (ICLR 2024 · similarity 0.87)
- A High-Dimensional Statistical Method for Optimizing Transfer Quantities in Multi-Source Transfer Learning (NeurIPS 2025 · 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.
