Scholarlink

This systematic review presents a taxonomy of five pivotal neural network architectures—Convolutional Neural Networks (CNNs), Spiking Neural Networks (SNNs), Graph Neural Networks (GNNs), Recurrent Neural Networks (RNNs), and Involutional Neural Networks (INNs)—based on an analysis of 142 peer-reviewed publications from 2010 to 2023. We propose a multi-dimensional comparison framework evaluating: (1) theoretical underpinnings, (2) computational complexity, (3) training dynamics, (4) domain suitability, (5) energy efficiency, and (6) interpretability. Key findings indicate that CNNs dominate computer vision applications (78% market share), SNNs achieve 8.7 \(\times \) energy efficiency in edge computing, and GNNs exhibit 62% annual growth in relational learning tasks. A decision matrix guides architecture selection across 12 domains, supported by examples like ResNet-50, Loihi, and GraphSAGE. Four research frontiers—hybrid architectures, neuromorphic scaling, efficient training, and unified frameworks—are identified to shape future AI designs. This review serves as a technical reference and strategic roadmap for advancing neural network designs.

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

A Systematic Taxonomy of Neural Network Architectures: Principles, Trade-Offs, and Future Directions

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sowad Rahman,
  • Raisha Rafa

摘要

This systematic review presents a taxonomy of five pivotal neural network architectures—Convolutional Neural Networks (CNNs), Spiking Neural Networks (SNNs), Graph Neural Networks (GNNs), Recurrent Neural Networks (RNNs), and Involutional Neural Networks (INNs)—based on an analysis of 142 peer-reviewed publications from 2010 to 2023. We propose a multi-dimensional comparison framework evaluating: (1) theoretical underpinnings, (2) computational complexity, (3) training dynamics, (4) domain suitability, (5) energy efficiency, and (6) interpretability. Key findings indicate that CNNs dominate computer vision applications (78% market share), SNNs achieve 8.7 \(\times \) energy efficiency in edge computing, and GNNs exhibit 62% annual growth in relational learning tasks. A decision matrix guides architecture selection across 12 domains, supported by examples like ResNet-50, Loihi, and GraphSAGE. Four research frontiers—hybrid architectures, neuromorphic scaling, efficient training, and unified frameworks—are identified to shape future AI designs. This review serves as a technical reference and strategic roadmap for advancing neural network designs.