Research map / Label-efficient and robust learning / Recognition and long-tail
Co-Tuning for Transfer 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 |
|---|---|---|---|
| add Co-Tuning (probabilistic source-label supervision via task-specific layers) to vanilla fine-tuning | ResNet-50 fine-tuning · Stanford Cars 15% | accuracy (%): 36.77 -> 46.02 | Helped |
| add Co-Tuning | ResNet-50 fine-tuning · CUB-200-2011 100% data | accuracy (%): 78.01 -> 81.24 | Helped |
| add Co-Tuning on large dataset | DenseNet-121 fine-tuning · COCO-70 100% data | accuracy (%): 84.41 -> 85.65 | Helped |
| add regularizers L2-SP/DELTA/BSS to fine-tuning with abundant data | DenseNet-121 fine-tuning · COCO-70 100% data | accuracy (%): 84.41 -> 84.78 (L2-SP) | No clear effect |
| add Co-Tuning to BERT NER fine-tuning | BERT-base · CoNLL 2003 NER | F1: 90.81 -> 91.27 | Helped |
| add Co-Tuning on dissimilar scene dataset | ResNet-50 fine-tuning · Places365 100 imgs/class | top-1 accuracy (%): 42.49 -> 43.56 | Helped |
| add calibration of the pre-trained model predictions (ablation) | Co-Tuning · CUB-200-2011 15% | accuracy (%): 50.67 -> 52.58 | Helped |
| learn category relationship with reverse approach instead of direct approach (ablation) | Co-Tuning · CUB-200-2011 15% | accuracy (%): 51.20 -> 52.58 | Helped |
| use two-layer network instead of softmax regression for g (ablation) | Co-Tuning · CUB-200-2011 15% | accuracy (%): 52.58 -> 52.03 | No clear effect |
| replace Co-Tuning with LwF to reuse task-specific layers (ablation) | ResNet-50 fine-tuning · CUB-200-2011 15% | accuracy (%): 52.58 -> 47.67 | Hurt |
| add new layer on top of source logits (AddLayer) instead of Co-Tuning (ablation) | ResNet-50 fine-tuning · CUB-200-2011 15% | accuracy (%): 45.25 -> 42.68 | Hurt |
| add L2-SP regularization at 100% data | ResNet-50 fine-tuning · Stanford Cars 100% | accuracy (%): 87.20 -> 86.58 | Hurt |
Closest papers by meaning
- Improved Regularization and Robustness for Fine-tuning in Neural Networks (NeurIPS 2021 · similarity 0.90)
- Borrowing Treasures From the Wealthy: Deep Transfer Learning Through Selective Joint Fine-Tuning (CVPR 2017 · similarity 0.90)
- Explicit Inductive Bias for Transfer Learning with Convolutional Networks (ICML 2018 · similarity 0.89)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 2024 · similarity 0.87)
- Quick-Tune: Quickly Learning Which Pretrained Model to Finetune and How (ICLR 2024 · similarity 0.86)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · 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.