Spatiotemporal all-optical diffractive neural networks empowered by spatiotemporal light field manipulation

Abstract All-optical neural networks (AONNs) offer significant potential for feature detection and pattern recognition by leveraging light-speed processing, high parallelism, and low energy consumption. However, traditional AONNs are limited to spatial light intensity distributions for both inputs and outputs, restricting network learning to the spatial-only dimension. To overcome this, we propose a novel spatiotemporal AONN architecture that integrates the time dimension with the spatial domain, enabling simultaneous modulation of light field information across both time and space. In this configuration, the output spatiotemporal light field is spatially filtered, and its temporal signals are detected using a single-pixel detector. This approach integrates the temporal dimension as a core computational resource into the network, driving a paradigm shift in AONNs from “imaging detection” to “temporal analysis”, effectively overcoming the speed bottleneck caused by the bandwidth limitations of traditional 2D array detectors. Moreover, experimental studies demonstrate that the system can accelerate the processing of real-world spatiotemporal lidar signals, facilitating the real-time reconstruction of multi-target moving scenes. Our innovative spatiotemporal AONN design offers enhanced flexibility, faster processing speeds, and superior performance, paving the way for more efficient and effective optical signal processing and computing applications.