Scholarlink

<p>Domain Generalization (DG) seeks to learn models that perform well on unseen target domains. While Dynamic Domain Generalization (DDG) introduces instance-wise adaptability, existing solutions often depend on data-level mixing and explicit domain labels, limiting their practicality. To address this, we propose Weight Perturbation Training (WPT), a novel training strategy that enhances the DDG framework by shifting the augmentation paradigm from the data space to the weight space. Without requiring domain labels, WPT injects controlled uncertainty by perturbing the dynamic coefficients of the network with both random and class-aware noise, thereby simulating diverse domain characteristics. This enables the learning of more robust representations without any data mixing. Extensive experiments on five established DG benchmarks—PACS, Office-Home, VLCS, TerraIncognita, and DomainNet—show that WPT consistently improves the strong DDG baseline and achieves competitive or leading performance against recent DG methods. Crucially, our approach preserves the training-free adaptation mechanism of DDG during inference, enabling robust generalization to unseen domains without fine-tuning.</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

Enhancing dynamic domain generalization with weight perturbation training

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Zhishu Sun,
  • Shuqi Yu,
  • Dongxing Wang,
  • Ling Xu,
  • Luojun Lin,
  • Yuanlong Yu

摘要

Domain Generalization (DG) seeks to learn models that perform well on unseen target domains. While Dynamic Domain Generalization (DDG) introduces instance-wise adaptability, existing solutions often depend on data-level mixing and explicit domain labels, limiting their practicality. To address this, we propose Weight Perturbation Training (WPT), a novel training strategy that enhances the DDG framework by shifting the augmentation paradigm from the data space to the weight space. Without requiring domain labels, WPT injects controlled uncertainty by perturbing the dynamic coefficients of the network with both random and class-aware noise, thereby simulating diverse domain characteristics. This enables the learning of more robust representations without any data mixing. Extensive experiments on five established DG benchmarks—PACS, Office-Home, VLCS, TerraIncognita, and DomainNet—show that WPT consistently improves the strong DDG baseline and achieves competitive or leading performance against recent DG methods. Crucially, our approach preserves the training-free adaptation mechanism of DDG during inference, enabling robust generalization to unseen domains without fine-tuning.