Research map / Label-efficient and robust learning / Domain adaptation and generalization
Task-aware Adaptive Learning for Cross-domain Few-shot Learning
Yurong Guo, Ruoyi Du, Yuan Dong, Timothy Hospedales, Yi-Zhe Song, Zhanyu Ma
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 |
|---|---|---|---|
| replace fixed task-specific adapter with task-adaptive RL-learned adapter configuration | TA2-Net vs TSA · Meta-Dataset multi-domain, unseen average | accuracy (%): 78.1 -> 79.1 | Improved |
| replace fixed task-specific adapter with task-adaptive configuration in single-domain training | TA2-Net vs TSA · Meta-Dataset single-domain, unseen average | accuracy (%): 71.9 -> 75.0 | Improved |
| train adapter gate with reinforcement learning instead of episodic training (ablation) | TA2-Net · Meta-Dataset multi-domain, unseen | accuracy (%): 67.9 -> 72.6 | Improved |
| widen the task-specific adapter (parallel 1x1 conv) instead of deepening it (series) (ablation) | Task-specific adapter · Meta-Dataset multi-domain | accuracy | Improved |
| attach fixed task-specific adapter to backbone (no adaptive selection) (ablation) | ResNet-18 backbone · Meta-Dataset seen domains | average accuracy | Hurt |
| attach fixed task-specific adapter to backbone (no adaptive selection) (ablation) | ResNet-18 backbone · Meta-Dataset unseen domains | average accuracy | Improved |
| increase max adapter width K from 1 to 3 (ablation) | TA2-Net · Meta-Dataset five-shot, unseen | accuracy (%): 66.6 -> 70.5 | Improved |
| increase max adapter width K from 3 to 4 (ablation) | TA2-Net · Meta-Dataset five-shot, unseen | accuracy (%): 70.5 -> 68.2 | Hurt |
| increase max adapter width K on seen domains (ablation) | TA2-Net · Meta-Dataset five-shot, seen | accuracy (%): 75.4 -> 75.1 (K=1 to K=4) | No clear effect |
| TA2-Net versus TSA on CIFAR-10 and Textures | TA2-Net · Meta-Dataset multi-domain | accuracy (%): CIFAR-10 82.9 -> 82.1; Textures 77.5 -> 77.4 | Mixed |
Closest papers by meaning
- Cross-Domain Few-Shot Learning With Task-Specific Adapters (CVPR 2022 · similarity 0.89)
- Task-Adaptive Prompted Transformer for Cross-Domain Few-Shot Learning (AAAI 2024 · similarity 0.88)
- Towards Few-Shot Adaptation of Foundation Models via Multitask Finetuning (ICLR 2024 · similarity 0.87)
- Task-Specific Preconditioner for Cross-Domain Few-Shot Learning (AAAI 2025 · similarity 0.87)
- Leveraging Normalization Layer in Adapters with Progressive Learning and Adaptive Distillation for Cross-Domain Few-Shot Learning (AAAI 2024 · similarity 0.87)
- SemTra: A Semantic Skill Translator for Cross-Domain Zero-Shot Policy Adaptation (AAAI 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.
