Scholarlink

<p>Existing multimodal fusion methods have shortcomings in cross-modal semantic alignment, dynamic scene adaptation, and long-term memory retention. This study proposes a flexible control model of an intelligent synaptic fusion network designed for dynamic multimodal fusion. The model employs a flexible control mechanism at its core to establish a three-layer integrated framework. The data processing layer first uniformly performs vectorization on multi-source, multi-modal information and assigns priorities using an attention mechanism. The flexible control layer incorporates a Dynamic Adaptive Neural Network (DANN) as an intelligent control agent to enable global dynamic regulation and adaptive optimization. The scene generation layer further constructs a fine-grained feature extractor based on CNNs and the Laplace transform. We then employed a dual-path fusion strategy based on KCCA and DCCA with hierarchical attention to capture strong and weak modal correlations, and integrated a Feedback Adaptive Memory Network (FAMN) to enhance historical feature replay and long-term memory learning. The experiments were tested on the CMU-MOSI, CH-SIMS, BraTS2020, and NYU Depth v2 datasets, as well as our own Full Modal dataset. In comparison with other models, the proposed method achieved improvements of 6.8%, 2.2%, 4.6%, 3.7%, and 0.3% in respective accuracy, F1 score, recall, AUC, and IoU.</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

An intelligent synapse fusion network with dynamic adaptive control for multimodal fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Weichen Zhai,
  • Keda Chen,
  • Shengwei Wang

摘要

Existing multimodal fusion methods have shortcomings in cross-modal semantic alignment, dynamic scene adaptation, and long-term memory retention. This study proposes a flexible control model of an intelligent synaptic fusion network designed for dynamic multimodal fusion. The model employs a flexible control mechanism at its core to establish a three-layer integrated framework. The data processing layer first uniformly performs vectorization on multi-source, multi-modal information and assigns priorities using an attention mechanism. The flexible control layer incorporates a Dynamic Adaptive Neural Network (DANN) as an intelligent control agent to enable global dynamic regulation and adaptive optimization. The scene generation layer further constructs a fine-grained feature extractor based on CNNs and the Laplace transform. We then employed a dual-path fusion strategy based on KCCA and DCCA with hierarchical attention to capture strong and weak modal correlations, and integrated a Feedback Adaptive Memory Network (FAMN) to enhance historical feature replay and long-term memory learning. The experiments were tested on the CMU-MOSI, CH-SIMS, BraTS2020, and NYU Depth v2 datasets, as well as our own Full Modal dataset. In comparison with other models, the proposed method achieved improvements of 6.8%, 2.2%, 4.6%, 3.7%, and 0.3% in respective accuracy, F1 score, recall, AUC, and IoU.