Scholarlink

<p>Real-time intelligent systems increasingly require hardware that can adapt continuously to evolving inputs, yet most existing processors rely on static-weight inference, making them vulnerable to distribution shifts and error accumulation in dynamic environments. Although adaptive weight updates can, in principle, address this limitation, their implementation on electronic hardware is hindered by the stability–plasticity trade-off, as well as by the memory wall and clocking bottlenecks that become particularly severe in sequential processing. Here, we present a Temporally Plastic Photonic Processor (TPPP) that enables ultra-fast in situ adaptation by combining multi-timescale photonic kernels with a recursive optical delay memory. The architecture integrates a slow, reconfigurable kernel for stable long-term processing and a fast, dynamic kernel for transient adaptation, enabling time-varying weights to be embedded directly in the photonic domain without repeated electronic memory access. We experimentally validate the TPPP on linear and nonlinear sequential tasks. In both regimes, data-driven temporal plasticity enables the TPPP to outperform conventional static photonic baselines in robustness and accuracy. Under an operation-matched INT8 comparison, scaling analysis projects up to 16 × higher per-operation energy efficiency and up to 100 × lower intrinsic single-pass compute delay than advanced electronic processors, establishing the TPPP as a promising hardware framework for real-time adaptive photonic computing.</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

Temporally plastic photonic processor for real-time adaptive computing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月23日

查看全文
  • Lingzhi Luo,
  • Yizhi Wang,
  • Zhiwei Xue,
  • Yanzhi Chen,
  • Chunhui Yao,
  • Senbiao Qin,
  • Peng Bao,
  • Jing Zhang,
  • Kangning Xu,
  • Minjia Chen,
  • Ting Yan,
  • Yuxiao Ye,
  • Liang Ming,
  • Gunther Roelkens,
  • Jianji Dong,
  • Tawfique Hasan,
  • Ian White,
  • Richard Penty,
  • Lu Fang,
  • Qixiang Cheng

摘要

Real-time intelligent systems increasingly require hardware that can adapt continuously to evolving inputs, yet most existing processors rely on static-weight inference, making them vulnerable to distribution shifts and error accumulation in dynamic environments. Although adaptive weight updates can, in principle, address this limitation, their implementation on electronic hardware is hindered by the stability–plasticity trade-off, as well as by the memory wall and clocking bottlenecks that become particularly severe in sequential processing. Here, we present a Temporally Plastic Photonic Processor (TPPP) that enables ultra-fast in situ adaptation by combining multi-timescale photonic kernels with a recursive optical delay memory. The architecture integrates a slow, reconfigurable kernel for stable long-term processing and a fast, dynamic kernel for transient adaptation, enabling time-varying weights to be embedded directly in the photonic domain without repeated electronic memory access. We experimentally validate the TPPP on linear and nonlinear sequential tasks. In both regimes, data-driven temporal plasticity enables the TPPP to outperform conventional static photonic baselines in robustness and accuracy. Under an operation-matched INT8 comparison, scaling analysis projects up to 16 × higher per-operation energy efficiency and up to 100 × lower intrinsic single-pass compute delay than advanced electronic processors, establishing the TPPP as a promising hardware framework for real-time adaptive photonic computing.