Machine intelligence has become a driving factor in modern society. However, its demand outpaces the underlying electronic technology due to limitations given by fundamental physics, such as capacitive charging of wires, but also by system architecture of storing and handling data, both driving recent trends toward processor heterogeneity. Task-specific accelerators based on free-space optics bear fundamental homomorphism for massively parallel and real-time information processing given the wave nature of light. However, initial results are frustrated by data handling challenges and slow optical programmability. Here we introduce a novel amplitude-only Fourier-optical processor paradigm capable of processing large-scale ${\sim}({1000}\times{1000})$ matrices in a single time step and 100 µs-short latency. Conceptually, the information flow direction is orthogonal to the two-dimensional programmable network, which leverages ${{10}^6}$ parallel channels of display technology, and enables a prototype demonstration performing convolutions as pixelwise multiplications in the Fourier domain reaching peta operations per second throughputs. The required real-to-Fourier domain transformations are performed passively by optical lenses at zero-static power. We exemplary realize a convolutional neural network (CNN) performing classification tasks on 2 megapixel large matrices at 10 kHz rates, which latency-outperforms current graphic processing unit and phase-based display technology by 1 and 2 orders of magnitude, respectively. Training this optical convolutional layer on image classification tasks and utilizing it in a hybrid optical-electronic CNN, shows classification accuracy of 98% (Modified National Institute of Standards and Technology) and 54% (CIFAR-10). Interestingly, the amplitude-only CNN is inherently robust against coherence noise in contrast to phase-based paradigms and features a delay over 2 orders of magnitude lower than liquid-crystal-based systems. Such an amplitude-only massively parallel optical compute paradigm shows that the lack of phase information can be accounted for via training, thus opening opportunities for high-throughput accelerator technology for machine intelligence with applications in network-edge processing, in data centers, or in pre-processing information or filtering toward near-real-time decision making.

中文翻译：

---

大规模并行仅振幅傅立叶神经网络

机器智能已成为现代社会的驱动因素。但是，由于基本物理技术（例如导线的电容性充电）以及存储和处理数据的系统体系结构所带来的限制，其需求超过了基础电子技术，这两者都推动了处理器异质性的最新趋势。给定光的波特性，基于自由空间光学的特定于任务的加速器具有基本的同态性，可用于大规模并行和实时信息处理。但是，最初的结果因数据处理挑战和缓慢的光学可编程性而受挫。在这里，我们介绍了一种新颖的仅幅值傅立叶光学处理器范例，该范例能够处理大规模$ {\ sim}（{1000} \ times {1000}）$单个时间步长的矩阵和100 µs短延迟。从概念上讲，信息流的方向与二维可编程网络正交，该网络利用$ {{10} ^ 6} $并行显示技术通道，并使原型演示能够在傅立叶域中执行像素级乘法的卷积，达到每秒Peta操作的吞吐量。所需的实数到傅里叶域变换是由光学透镜在零静态功率下被动执行的。我们示例性地实现了一个卷积神经网络（CNN），它以10 kHz的速率在2兆像素的大型矩阵上执行分类任务，其延迟分别比当前的图形处理单元和基于相位的显示技术高1到2个数量级。在图像分类任务上训练此光学卷积层并将其用于混合光电CNN中，显示出分类精度为98％（美国国家标准技术研究院，修正版）和54％（CIFAR-10）。有趣的是 与基于相位的范例相反，纯振幅的CNN具有固有的抗相干噪声的功能，并且其延迟比基于液晶的系统低2个数量级。这种仅振幅的大规模并行光学计算范例表明，可以通过训练来解决相位信息的缺乏问题，从而为机器智能的高通量加速器技术提供了机会，这些技术可以应用于网络边缘处理，数据中心或网络中。预处理信息或进行过滤以实现近乎实时的决策。