Scholarlink

We propose DDTL (Double Disentangled Transfer Learning), a novel parameter-efficient framework for adapting vision transformers. DDTL enhances the recently introduced DTL by incorporating dual compact side networks, placed before and in each transformer block, enabling more expressive and fine-grained task adaptation. Unlike traditional PETL approaches that entangle the backbone layers, DDTL remains highly memory-efficient while achieving superior accuracy. We further explore block-dropping strategies, including loss-guided and cosine similarity-based methods, to reduce computation cost during training. Experiments on VTAB-1K across multiple ViT scales demonstrate that DDTL consistently outperforms strong PETL baselines with low parameter increase and memory footprint. Our method provides a scalable and flexible solution for efficient transfer learning in low-resource settings.

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

Dual DTL: Double-Side Compact Transfer Networks for Efficient ViT Adaptation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zonghao Ding,
  • Minghao Fu,
  • Jianxin Wu

摘要

We propose DDTL (Double Disentangled Transfer Learning), a novel parameter-efficient framework for adapting vision transformers. DDTL enhances the recently introduced DTL by incorporating dual compact side networks, placed before and in each transformer block, enabling more expressive and fine-grained task adaptation. Unlike traditional PETL approaches that entangle the backbone layers, DDTL remains highly memory-efficient while achieving superior accuracy. We further explore block-dropping strategies, including loss-guided and cosine similarity-based methods, to reduce computation cost during training. Experiments on VTAB-1K across multiple ViT scales demonstrate that DDTL consistently outperforms strong PETL baselines with low parameter increase and memory footprint. Our method provides a scalable and flexible solution for efficient transfer learning in low-resource settings.