Scholarlink

<p>Source-free domain adaptation (SFDA) aims to effectively transfer the knowledge learned by pre-trained models from the source distribution to the target distribution. A common approach is to use pseudo-labels (PL) for self-supervised training. However, the PL for the target domain data have a high level of noise. Therefore, it is advisable to avoid relying solely on PL for model training. In this paper, we design an innovative energy-guided active SFDA method (EASDA), which applies active learning to SFDA. It leverages an active selection strategy based on free energy and uncertainty to select active samples with high information. In addition, we design an energy-based PL approach for training unlabeled samples. Furthermore, to promote model robustness and diversity, we introduce an information maximization (IM) loss to implicitly mine knowledge from the target domain. To demonstrate the superiority of our method, we conduct experiments on multiple benchmark datasets for object classification. The results indicate that our method outperforms the current state-of-the-art approaches, achieving the highest accuracy. Our code can be found at <a href="https://github.com/qingfann/EASDA-code.">https://github.com/qingfann/EASDA-code.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Energy-guided active source-free domain adaptation

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Md Gulzar Hussain,
  • Qing Tian,
  • Liangyu Zhou

摘要

Source-free domain adaptation (SFDA) aims to effectively transfer the knowledge learned by pre-trained models from the source distribution to the target distribution. A common approach is to use pseudo-labels (PL) for self-supervised training. However, the PL for the target domain data have a high level of noise. Therefore, it is advisable to avoid relying solely on PL for model training. In this paper, we design an innovative energy-guided active SFDA method (EASDA), which applies active learning to SFDA. It leverages an active selection strategy based on free energy and uncertainty to select active samples with high information. In addition, we design an energy-based PL approach for training unlabeled samples. Furthermore, to promote model robustness and diversity, we introduce an information maximization (IM) loss to implicitly mine knowledge from the target domain. To demonstrate the superiority of our method, we conduct experiments on multiple benchmark datasets for object classification. The results indicate that our method outperforms the current state-of-the-art approaches, achieving the highest accuracy. Our code can be found at https://github.com/qingfann/EASDA-code.