Optical implementations of neural networks (ONNs) herald the next-generation high-speed and energy-efficient deep learning computing by harnessing the technical advantages of large bandwidth and high parallelism of optics. However, because of limited hardware scale and incomplete numerical domain, the majority of existing ONNs are merely studied and benchmarked with basic classification tasks. Given that regression is a fundamental form of deep learning and accounts for a large part of current artificial intelligence applications, we demonstrate a silicon-based optical coherent dot-product chip (OCDC) capable of completing deep learning regression tasks. The OCDC adopts optical fields rather than intensities to represent values in the complete real-value domain. It conducts matrix multiplications and convolutions in neural networks of any complexity via reconfiguration and reusing. Hardware deviations are compensated via in-situ backpropagation control owing to the simplicity of chip architecture, thus enhancing the numerical accuracy of analog computing. Therefore, the OCDC meets the fundamental requirement for regression tasks and we successfully demonstrate a representative neural network, the AUTOMAP (a cutting-edge neural network model for image reconstruction). The quality of reconstructed images by the OCDC and a 32-bit digital computer is comparable. To best of our knowledge, there is no precedent of performing such state-of-the-art regression tasks on ONN chip. It is anticipated that the OCDC can promote novel accomplishment of ONNs in modern AI applications including autonomous driving, natural language processing, medical diagnosis, and scientific study. Moreover, the OCDC and auxiliary electronics have the potential to be monolithically fabricated with CMOS-compatible silicon photonic integration technologies.

中文翻译：

---

光学相干点积芯片，用于复杂的深度学习回归

神经网络（ONN）的光学实现通过利用光学大带宽和高并行性的技术优势，预示着下一代高速，高能效的深度学习计算。但是，由于有限的硬件规模和不完整的数值域，大多数现有的ONN只是通过基本的分类任务进行研究和基准测试。鉴于回归是深度学习的基本形式，并占了当前人工智能应用的很大一部分，我们演示了一种能够完成深度学习回归任务的基于硅的光学相干点积芯片（OCDC）。OCDC采用光场而不是强度来表示完整实值域中的值。它通过重新配置和重用在神经网络中进行任何形式的矩阵乘法和卷积。由于芯片架构的简单性，可通过原位反向传播控制来补偿硬件偏差，从而提高了模拟计算的数值精度。因此，OCDC满足了回归任务的基本要求，并且我们成功地演示了代表性的神经网络AUTOMAP（用于图像重建的尖端神经网络模型）。OCDC和32位数字计算机的重建图像质量是可比的。据我们所知，在ONN芯片上执行此类最新的回归任务尚无先例。预计OCDC可以在现代AI应用程序（包括自动驾驶，自然语言处理，医学诊断和科学研究。而且，OCDC和辅助电子器件具有使用CMOS兼容的硅光子集成技术进行整体制造的潜力。