Scholarlink

<p>Collaborative computing between edge devices and cloud servers over wireless communication is critical for energy-constrained edge devices to perform complex tasks that exceed their processing capacities. However, current wireless collaborative systems face challenges in terms of energy efficiency and latency due to the separation of memory and computing, the separation of signal processing and transmission and/or reception, and the separation of neural networks and wireless communication. Here we report communication-aware in-memory wireless neural networks. The approach uses analogue in-memory computing technology to implement both edge computing and wireless communication, and integrates wireless communication as a learnable module of the wireless neural network. We build a prototype that comprises an edge inference accelerator and a wireless communication system. The prototype exhibits an experimental inference accuracy of 93.71% on the Street View House Numbers dataset, and can maintain inference accuracy when using low-resolution analogue-to-digital converters in wireless communication. We also show that the approach can adapt to various wireless conditions and can reduce communication costs.</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

Communication-aware in-memory wireless neural networks

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Zai-Zheng Yang,
  • Cong Wang,
  • Yichen Zhao,
  • Gong-Jie Ruan,
  • Xing-Jian Yangdong,
  • Yuekun Yang,
  • Chen Pan,
  • Bin Cheng,
  • Shi-Jun Liang,
  • Feng Miao

摘要

Collaborative computing between edge devices and cloud servers over wireless communication is critical for energy-constrained edge devices to perform complex tasks that exceed their processing capacities. However, current wireless collaborative systems face challenges in terms of energy efficiency and latency due to the separation of memory and computing, the separation of signal processing and transmission and/or reception, and the separation of neural networks and wireless communication. Here we report communication-aware in-memory wireless neural networks. The approach uses analogue in-memory computing technology to implement both edge computing and wireless communication, and integrates wireless communication as a learnable module of the wireless neural network. We build a prototype that comprises an edge inference accelerator and a wireless communication system. The prototype exhibits an experimental inference accuracy of 93.71% on the Street View House Numbers dataset, and can maintain inference accuracy when using low-resolution analogue-to-digital converters in wireless communication. We also show that the approach can adapt to various wireless conditions and can reduce communication costs.