Scholarlink

<p>Traditional optical neural networks necessitate separate components for linear operations and nonlinear activations, increasing system complexity and energy consumption. Here, we demonstrate a dual-function electro-optical modulator based on graphene-coated silicon waveguides that simultaneously performs both weight mapping and nonlinear activation functions within a single device. By exploiting the voltage-dependent optical transmission characteristics of graphene, our modulator achieves a transmission range from 0.14 to 1.0 with a voltage-response curve that closely approximates a sigmoid function (fitting error &lt; ±6%). Assessment analysis on standard machine learning datasets reveals reconstruction accuracies of approximately 86.7% after 60 training iterations, with accuracy remaining above 80% under significant noise perturbations. This approach substantially reduces hardware complexity and energy overhead, representing a significant advancement toward practical integrated photonic neural computing systems.</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

Dual-function electro-optical modulation for fully hardware-implemented photonic restricted Boltzmann machine

  • 期刊论文

  • 发表日期：

    2025年11月6日

查看全文
  • Chengwang Yang,
  • Chengyan Zhong,
  • Yapeng Zhang,
  • Wenbin Shen,
  • Lingfei Li,
  • Yu Liu

摘要

Traditional optical neural networks necessitate separate components for linear operations and nonlinear activations, increasing system complexity and energy consumption. Here, we demonstrate a dual-function electro-optical modulator based on graphene-coated silicon waveguides that simultaneously performs both weight mapping and nonlinear activation functions within a single device. By exploiting the voltage-dependent optical transmission characteristics of graphene, our modulator achieves a transmission range from 0.14 to 1.0 with a voltage-response curve that closely approximates a sigmoid function (fitting error < ±6%). Assessment analysis on standard machine learning datasets reveals reconstruction accuracies of approximately 86.7% after 60 training iterations, with accuracy remaining above 80% under significant noise perturbations. This approach substantially reduces hardware complexity and energy overhead, representing a significant advancement toward practical integrated photonic neural computing systems.