Research map / Label-efficient and robust learning / OOD and anomaly detection
Fast Trainable Projection for Robust Fine-tuning
Junjiao Tian, Yen-Cheng Liu, James Smith, Zsolt Kira
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 TPGM with FTP (single-pass trainable projection) | Projected fine-tuning · DomainNet MoCo-V3 ResNet50 100% | training time (s/it): 0.80 -> 0.51 | Improved |
| replace TPGM with FTP | Projected fine-tuning · DomainNet MoCo-V3 ResNet50 100% | OOD avg accuracy: 36.86 -> 37.62 | Improved |
| replace TPGM with FTP | Projected fine-tuning · DomainNet CLIP ResNet50 100% | OOD avg accuracy: 39.65 -> 39.94 | Improved |
| replace TPGM with FTP | Projected fine-tuning · DomainNet CLIP ResNet50 10% | OOD avg accuracy: 30.42 -> 29.83 | No clear effect |
| replace TPGM with FTP | Projected fine-tuning · DomainNet CLIP ResNet50 10% | training time (s/it): 1.10 -> 0.69 | Improved |
| apply FTP instead of vanilla fine-tuning | Fine-tuning · DomainNet CLIP ResNet50 100% | ID accuracy (Real): 80.93 -> 84.22 | Improved |
| apply FTP instead of vanilla fine-tuning | Fine-tuning · PASCAL semantic segmentation | OOD avg mIoU: 44.00 -> 58.45 | Improved |
| apply FTP instead of vanilla fine-tuning | Fine-tuning · PASCAL surface normal | OOD avg RMSE: 23.48 -> 20.48 | Improved |
| apply FTP instead of TPGM | Projected fine-tuning · PASCAL human parts segmentation | OOD avg mIoU: 51.04 -> 52.12 | Improved |
| replace TPGM with FTP | Projected fine-tuning · PASCAL semantic segmentation | training time (s/it): 0.611 -> 0.401 | Improved |
| compare MARS-SP with vanilla fine-tuning | Fine-tuning · PASCAL semantic segmentation | OOD avg mIoU: 44.00 -> 43.59 | No clear effect |
| use FTP and linear-probe style baseline LP instead of vanilla fine-tuning | Fine-tuning · DomainNet CLIP ResNet50 100% | ID accuracy (Real): 80.93 -> 52.56 | Hurt |
Closest papers by meaning
- Tuning Stable Rank Shrinkage: Aiming at the Overlooked Structural Risk in Fine-tuning (CVPR 2024 · similarity 0.87)
- Directional Gradient Projection for Robust Fine-Tuning of Foundation Models (ICLR 2025 · similarity 0.86)
- Rethinking Weight Decay for Robust Fine-Tuning of Foundation Models (NeurIPS 2024 · similarity 0.86)
- Towards Calibrated Robust Fine-Tuning of Vision-Language Models (NeurIPS 2024 · similarity 0.85)
- PROFIT: A Specialized Optimizer for Deep Fine Tuning (NeurIPS 2025 · similarity 0.85)
- Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation (CVPR 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.
