All‐optical neural networks have advantages in higher throughput, higher speed as well as lower energy consumption compared to electrical neural networks. Optical neural networks have already shown great potential in various applications; however, the operation speed of the network is limited by the 2D detector as most optical neural networks rely on space to space projection. Here, a space to time projection approach to build diffractive deep neural network (D2NN) is proposed, which can project spatial intensity distribution into time‐domain intensity variation, thus bypassing the detection speed limit of 2D imaging device. Based on this scheme, high‐speed all‐optical logic gates are theoretically analyzed and experimentally realized. In this case, the network's operation speed is only limited by the photodetector (PD), which can reach GHz levels. Moreover, the method will show great advantage when it comes to wavelengths where 2D detectors are not achievable easily such as infrared, terahertz or microwaves.

中文翻译：

---

支持时空投影的超快全光衍射神经网络

与电神经网络相比，全光神经网络具有更高的吞吐量、更高的速度以及更低的能耗等优势。光神经网络已经在各种应用中显示出巨大的潜力；然而，网络的运行速度受到 2D 探测器的限制，因为大多数光学神经网络依赖于空间到空间的投影。这里，一种空间到时间投影方法来构建衍射深度神经网络（D2NN）的提出，可以将空间强度分布投影为时域强度变化，从而绕过二维成像设备的检测速度限制。基于该方案，对高速全光逻辑门进行了理论分析和实验实现。在这种情况下，网络的运行速度仅受到光电探测器（PD）的限制，光电探测器可以达到GHz级别。此外，当涉及到二维探测器不易实现的波长（例如红外线、太赫兹或微波）时，该方法将显示出巨大的优势。