핵심 개념
The author introduces a unified Source-Free Domain Adaptation problem and proposes a novel approach called Latent Causal Factors Discovery (LCFD) to address it. By focusing on causality relationships between latent variables and model decisions, LCFD aims to enhance model reliability against domain shifts.
초록
In the pursuit of transferring a source model to a target domain without access to the source training data, Source-Free Domain Adaptation (SFDA) has been extensively explored. Existing methods have limitations in practical utility and deployability due to focusing on specific scenarios. To address this, the author introduces Unified SFDA and proposes LCFD as a novel approach that emphasizes causality perspective for enhanced model robustness.
Key points:
- SFDA challenges due to strict data access controls.
- Unified SFDA introduced to comprehensively address various scenarios.
- LCFD proposed as an approach focusing on causal relationships for better model reliability.
- External and internal causal factors discovered through ViL models like CLIP.
- Self-supervised information bottleneck used for causal factor discovery.
- Extensive experiments show LCFD achieving state-of-the-art results in various SFDA settings.
통계
Existing works focus on specific SFDA settings, limiting usability and generality substantially in practice.
Proposed LCFD achieves new state-of-the-art results in distinct SFDA scenarios.
인용구
"Extensive experiments demonstrate that LCFD can achieve new state-of-the-art results in distinct SFDA settings."