Scholarlink

<p>Edge computing requires real-time processing of high-throughput analog signals, posing a major challenge to conventional electronics. Although integrated photonics offers low-latency processing, it struggles to directly handle raw analog data. Here, we present a photonic edge intelligence chip (PEIC) that fuses multiple analog modalities—images, spectra, and radio-frequency signals—into broad optical spectra for single-fiber input. After transmission onto the chip, these spectral inputs are processed by an arrayed waveguide grating (AWG) that performs both spectral sensing and energy-efficient convolution (29 fJ/OP). A subsequent nonlinear activation layer and a fully connected layer form an end-to-end optical neural network, achieving on-chip inference with a measured response time of 1.33 ns. We demonstrate both supervised and unsupervised learning on three tasks: drug spectral recognition, image classification, and radar target classification. Our work paves the way for on-chip solutions that unify analog signal acquisition and optical computation for edge intelligence.</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

Photonic edge intelligence chip for multi-modal sensing, inference and learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月19日

查看全文
  • Shiji Zhang,
  • Xueyi Jiang,
  • Bo Wu,
  • Haojun Zhou,
  • Wenguang Xu,
  • Hailong Zhou,
  • Zhichao Ruan,
  • Jianji Dong,
  • Xinliang Zhang

摘要

Edge computing requires real-time processing of high-throughput analog signals, posing a major challenge to conventional electronics. Although integrated photonics offers low-latency processing, it struggles to directly handle raw analog data. Here, we present a photonic edge intelligence chip (PEIC) that fuses multiple analog modalities—images, spectra, and radio-frequency signals—into broad optical spectra for single-fiber input. After transmission onto the chip, these spectral inputs are processed by an arrayed waveguide grating (AWG) that performs both spectral sensing and energy-efficient convolution (29 fJ/OP). A subsequent nonlinear activation layer and a fully connected layer form an end-to-end optical neural network, achieving on-chip inference with a measured response time of 1.33 ns. We demonstrate both supervised and unsupervised learning on three tasks: drug spectral recognition, image classification, and radar target classification. Our work paves the way for on-chip solutions that unify analog signal acquisition and optical computation for edge intelligence.