CoLin introduces a 1% parameter low-rank complex adapter for vision foundation models. It resolves convergence issues in composite matrices with tailored loss. Surpasses full fine-tuning and delta-tuning on detection, segmentation, and classification.
Key Points
- 1.Complex linear projection optimization
- 2.Theoretical fix for low-rank convergence
- 3.Code released on GitHub
Impact Analysis
Enables efficient deployment of vision models, reducing costs dramatically while boosting performance.
Technical Details
Low-rank complex adapter architecture; proven loss addresses training instability; excels in remote sensing too.