Programmable photonic neural engine with all-optical nonlinear activation and 40,000 connections

Abstract

The rapid advancement of artificial neural networks (ANNs) demands computational platforms with higher speed, energy efficiency, and scalability. Optical computing offers an appealing solution owing to the high-speed processing, inherent parallelism, and low energy consumption. However, existing optical neural network architectures face fundamental trade-offs between scalability, reconfigurability, and processing latency, constrained by the need to rely on electronic nonlinearities or fixed optical interconnections. Here, we report an end-to-end photonic neuromorphic engine that integrates all-optical nonlinearity into a loop-based, time-multiplexed photonic architecture. This approach enables all optical neurons to share common hardware, supporting enhanced connectivity and full reconfigurability with reduced latency. Implementations of four distinct network topologies, with a maximum of 40,000 optical connections for the single-layer perceptron, achieve digital-level inference accuracy and computational latency over two orders of magnitude shorter than state-of-the-art electronic processors, potentially paving the way for scalable, reconfigurable, and ultrafast optical artificial intelligence systems.