Scholarlink

Multimodal learning holds immense potential in video understanding, semantic matching, and human-computer interaction, yet existing methods still suffer from modal information asymmetry and training imbalance issues. This paper presents DAGMP (Dual-Aware Guided Multimodal Plug-in), a plug-and-play solution for addressing modal information asymmetry and training imbalance in multimodal learning. DAGMP enhances feature alignment and noise suppression with the EEDF (Enhanced Energy-based Dynamic Fusion) strategy, and optimizes gradient consistency using the DEA (Gradient Energy with Angle) mechanism. It can be integrated into any multimodal framework without structural changes. Experiments on CREMA-D, MEAD, and AVE datasets show significant performance improvements, demonstrating its effectiveness and versatility.

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

DAGMP: A Multimodal Learning Approach Jointly Driven by Feature Fusion and Gradient Modulation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xinying Zhou,
  • Leixiao Li,
  • Hao Lin

摘要

Multimodal learning holds immense potential in video understanding, semantic matching, and human-computer interaction, yet existing methods still suffer from modal information asymmetry and training imbalance issues. This paper presents DAGMP (Dual-Aware Guided Multimodal Plug-in), a plug-and-play solution for addressing modal information asymmetry and training imbalance in multimodal learning. DAGMP enhances feature alignment and noise suppression with the EEDF (Enhanced Energy-based Dynamic Fusion) strategy, and optimizes gradient consistency using the DEA (Gradient Energy with Angle) mechanism. It can be integrated into any multimodal framework without structural changes. Experiments on CREMA-D, MEAD, and AVE datasets show significant performance improvements, demonstrating its effectiveness and versatility.