Reproduce “Safe Self-Refinement for Transformer-based Domain Adaptation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tao Sun, Cheng Lu, Tianshuo Zhang, Haibin Ling
A new framework, Safe Self-Refinement for Transformer-based Domain Adaptation (SSRT), integrates Vision Transformers with a robust self-refinement and adaptive 'safe training' mechanism to improve performance in Unsupervised Domain Adaptation. This approach achieved 45.2% average accuracy on the challenging DomainNet dataset, demonstrating enhanced stability and effectiveness across large domain gaps.
Key insights
- Vision Transformers learn highly transferable features that provide a strong foundation for Unsupervised Domain Adaptation, often outperforming CNNs even with basic adaptation techniques.
- Applying perturbations to latent feature representations within transformer blocks is more effective for consistency regularization than perturbing raw input images, enhancing robustness for high-capacity models.
- Adaptive monitoring of prediction diversity and dynamic adjustment of training parameters are crucial for stable training, effectively preventing model collapse in challenging UDA tasks with large domain shifts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.