Scholarlink

<p>The increasing reliance on large and computationally intensive deep learning models has intensified the demand for architectures that balance performance, efficiency, and adaptability. Dynamic neural networks, which enable input-dependent or learning-driven adaptation of computational behavior, have emerged as a compelling paradigm for addressing these challenges. Despite growing interest, existing research remains fragmented, with dynamic mechanisms explored primarily along structural dimensions (e.g., depth, width, and path) and, more recently, within functional components (e.g., activation, normalization, and pooling). This paper presents a systematic review and conceptual synthesis of dynamic neural architectures, introducing a unified taxonomy that organizes existing approaches into parametric mechanisms, intermediate modules, auxiliary networks, and specialized architectures. Beyond categorization, we analyze the fundamental trade-offs governing dynamic design, highlighting how adaptivity, efficiency, representational capacity, and optimization stability interact. Our analysis indicates that while dynamic mechanisms offer substantial benefits in computational efficiency and modeling flexibility, they also introduce persistent challenges, including training instability, architectural overhead, and limited interpretability. Furthermore, the adoption of dynamic architectures in risk-sensitive domains remains constrained by reliability and robustness considerations. By consolidating existing knowledge, clarifying design trade-offs, and identifying emerging research directions, this work provides a conceptual framework for advancing scalable, interpretable, and trustworthy dynamic neural networks.</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

Exploits and advances in dynamic neural networks

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Kwang Jerh Chia,
  • Tomas Maul,
  • Zhiyuan Chen,
  • Veronica Lestari Jauw,
  • Yee Jiun Yap

摘要

The increasing reliance on large and computationally intensive deep learning models has intensified the demand for architectures that balance performance, efficiency, and adaptability. Dynamic neural networks, which enable input-dependent or learning-driven adaptation of computational behavior, have emerged as a compelling paradigm for addressing these challenges. Despite growing interest, existing research remains fragmented, with dynamic mechanisms explored primarily along structural dimensions (e.g., depth, width, and path) and, more recently, within functional components (e.g., activation, normalization, and pooling). This paper presents a systematic review and conceptual synthesis of dynamic neural architectures, introducing a unified taxonomy that organizes existing approaches into parametric mechanisms, intermediate modules, auxiliary networks, and specialized architectures. Beyond categorization, we analyze the fundamental trade-offs governing dynamic design, highlighting how adaptivity, efficiency, representational capacity, and optimization stability interact. Our analysis indicates that while dynamic mechanisms offer substantial benefits in computational efficiency and modeling flexibility, they also introduce persistent challenges, including training instability, architectural overhead, and limited interpretability. Furthermore, the adoption of dynamic architectures in risk-sensitive domains remains constrained by reliability and robustness considerations. By consolidating existing knowledge, clarifying design trade-offs, and identifying emerging research directions, this work provides a conceptual framework for advancing scalable, interpretable, and trustworthy dynamic neural networks.