Research map / Label-efficient and robust learning / Domain adaptation and generalization
Omni-Training: Bridging Pre-Training and Meta-Training for Few-Shot Learning
Yang Shu, Zhangjie Cao, Jinghan Gao, Jianmin Wang, Philip S. Yu, Mingsheng Long
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 |
|---|---|---|---|
| add self-distillation regularization R (ablation) | Omni-Training (pre+meta+joint losses) · mini-ImageNet 5-shot | accuracy (%): 80.79 -> 81.26 | Improved |
| add self-distillation regularization R (ablation) | Omni-Training (pre+meta+joint losses) · CUB 5-shot | accuracy (%): 88.54 -> 91.09 | Improved |
| add self-distillation regularization R (ablation) | Omni-Training (pre+meta+joint losses) · mini-ImageNet->CUB 5-shot | accuracy (%): 70.12 -> 71.30 | Improved |
| add joint loss L_joint (ablation) | Omni-Training (pre+meta losses) · mini-ImageNet 5-shot | accuracy (%): 78.12 -> 80.79 | Improved |
| add joint loss L_joint (ablation) | Omni-Training (pre+meta losses) · CUB 5-shot | accuracy (%): 87.34 -> 88.54 | Improved |
| add joint loss L_joint (ablation) | Omni-Training (pre+meta losses) · mini-ImageNet->CUB 5-shot | accuracy (%): 69.66 -> 70.12 | Improved |
| add meta-training loss to pre-training loss (ablation) | Omni-Net · mini-ImageNet->CUB 5-shot | accuracy (%): 65.57 -> 69.66 | Improved |
| add extra 1x1 conv parameters only (ProtoNet*) (ablation) | ProtoNet ResNet-18 · mini-ImageNet 5-shot | accuracy (%): 73.68 -> 73.44 | No clear effect |
| use Omni-Net tri-flow instead of just enlarging model (ablation) | ProtoNet* ResNet-18 · mini-ImageNet 5-shot | accuracy (%): 73.44 -> 81.26 | Improved |
| replace FT-GNN with OT-GNN in Multi-domain | FT-GNN · Multi-domain CUB 5-shot (train on mini-ImageNet) | accuracy (%): 66.98 -> 70.24 | Improved |
| turn all ResNet blocks into Omni-Layers instead of sharing early blocks (ablation) | Omni-Net · CUB 5-shot | accuracy (%) | Improved |
| remove self-distillation (lambda=0) (ablation) | Omni-Training · mini-ImageNet 1-shot | accuracy (%) | Hurt |
Closest papers by meaning
- Self-training For Few-shot Transfer Across Extreme Task Differences (ICLR 2021 · similarity 0.86)
- Meta-Transfer Learning for Few-Shot Learning (CVPR 2019 · similarity 0.86)
- LGM-Net: Learning to Generate Matching Networks for Few-Shot Learning (ICML 2019 · similarity 0.86)
- Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.85)
- Learning to Forget for Meta-Learning via Task-and-Layer-Wise Attenuation (TPAMI 2021 · similarity 0.85)
- Task-Level Self-Supervision for Cross-Domain Few-Shot Learning (AAAI 2022 · 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.
