Scholarlink

<p>Vision-language models (VLMs) have achieved notable success in zero-shot image recognition; however, transferring these models to downstream tasks remains challenging. This study introduces Hierarchical Prompt Learning (HPL), a novel framework that incorporates semantic hierarchies into prompt learning. HPL learns multi-level prompts at coarse, mid, and fine granularities, dynamically fusing them to enhance transferability and robustness. Experiments on ImageNet, CIFAR-100, Caltech101, and DomainNet demonstrate HPL’s superior generalization performance, harmonic-mean accuracy, and out-of-distribution robustness compared to state-of-the-art methods. Here, we show a 5% improvement in harmonic-mean accuracy on CIFAR-100, highlighting HPL’s potential in real-world open-world tasks.</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

Hierarchical prompt learning for vision-language models with enhanced generalization

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Varuna Kumara,
  • Manjunatha Badiger,
  • Akshatha Naik

摘要

Vision-language models (VLMs) have achieved notable success in zero-shot image recognition; however, transferring these models to downstream tasks remains challenging. This study introduces Hierarchical Prompt Learning (HPL), a novel framework that incorporates semantic hierarchies into prompt learning. HPL learns multi-level prompts at coarse, mid, and fine granularities, dynamically fusing them to enhance transferability and robustness. Experiments on ImageNet, CIFAR-100, Caltech101, and DomainNet demonstrate HPL’s superior generalization performance, harmonic-mean accuracy, and out-of-distribution robustness compared to state-of-the-art methods. Here, we show a 5% improvement in harmonic-mean accuracy on CIFAR-100, highlighting HPL’s potential in real-world open-world tasks.