arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2501.16604physics.optics
收藏

利用纳米光子光学参量振荡器实现超快类脑计算

Ultrafast neuromorphic computing with nanophotonic optical parametric oscillators

Midya Parto, Gordon H. Y. Li, Ryoto Sekine, Robert M. Gray, Luis L. Ledezma, James Williams, Arkadev Roy, Alireza Marandi

AI总结:

本文提出了一种基于薄膜铌酸锂光学参量振荡器的超快纳米光子类脑处理器,利用非线性延迟动力学构建深度循环神经网络,在约10 GHz时钟频率下实现超93%准确率及亚纳秒级低延迟,避免了光电转换,为全光类脑计算铺路。

摘要:

在过去十年中,人工智能(AI)在基础科学和日常技术领域带来了颠覆性的进步。在各种机器学习算法中,深度神经网络已成为揭示大型数据集中复杂模式的关键工具,并在计算机视觉、自然语言处理和预测分析中有着重要应用。片上光子神经网络提供了一个极具前景的平台,它利用光信号的高带宽和低传播损耗特性来执行用于深度学习的模拟计算。然而,纳米光子电路尚未能在全光和超快模式下同时实现所需的线性和非线性操作。在此,我们报告了一种超快纳米光子类脑处理器,它采用了在薄膜铌酸锂(TFLN)上制造的光学参量振荡器(OPO)。输入数据被用于调制同步泵浦OPO的光脉冲。OPO产生的后续信号脉冲通过OPO的非线性延迟动力学相互耦合,从而形成深度循环神经网络的内部节点。我们将这种非线性耦合的OPO网络用于混沌时间序列预测、噪声通信信道中的非线性纠错以及噪声波形分类,并在约10 GHz的工作时钟频率下实现了超过93%的准确率。我们的OPO网络能够实现亚纳秒级延迟,这一时间尺度可与最先进的数字电子处理器中的单个时钟周期相媲美。通过避免光电光(OEO)转换的需求,我们的超快纳米光子神经网络为下一代具有超低延迟和高能效的紧凑型全光类脑处理器铺平了道路。

英文摘要:

Over the past decade, artificial intelligence (AI) has led to disruptive advancements in fundamental sciences and everyday technologies. Among various machine learning algorithms, deep neural networks have become instrumental in revealing complex patterns in large datasets with key applications in computer vision, natural language processing, and predictive analytics. On-chip photonic neural networks offer a promising platform that leverage high bandwidths and low propagation losses associated with optical signals to perform analog computations for deep learning. However, nanophotonic circuits are yet to achieve the required linear and nonlinear operations simultaneously in an all-optical and ultrafast fashion. Here, we report an ultrafast nanophotonic neuromorphic processor using an optical parametric oscillator (OPO) fabricated on thin-film lithium niobate (TFLN). The input data is used to modulate the optical pulses synchronously pumping the OPO. The consequent signal pulses generated by the OPO are coupled to one another via the nonlinear delayed dynamics of the OPO, thus forming the internal nodes of a deep recurrent neural network. We use such a nonlinearly coupled OPO network for chaotic time series prediction, nonlinear error correction in a noisy communication channel, as well as noisy waveform classification and achieve accuracies exceeding 93% at an operating clock rate of ~ 10 GHz. Our OPO network is capable of achieving sub-nanosecond latencies, a timescale comparable to a single clock cycle in state-of-the-art digital electronic processors. By circumventing the need for optical-electronic-optical (OEO) conversions, our ultrafast nanophotonic neural network paves the way for the next generation of compact all-optical neuromorphic processors with ultralow latencies and high energy efficiencies.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑