Scalable optical neural networks based on temporal computing postprint

Abstract: The optical neural network (ONN) has been considered as a promising candidate for next-generation neurocomputing due to its high parallelism, low latency, and low energy consumption, with significant potential to release unprecedented computational capability. Large-scale ONNs could process more neural information and improve the prediction performance. However, previous ONN architectures based on matrix multiplication are difficult to scale up due to manufacturing limitations, resulting in limited scalability and small input data volumes. To address this challenge, we propose a compact and scalable photonic computing architecture based on temporal photoelectric multiplication and accumulate (MAC) operations, allowing direct processing of large-scale matrix computations in the time domain. By employing a temporal computing unit composed of cascaded modulators and time-integrator, we conduct a series of proof-of-principle experiments including image edge detection, optical neural networks-based recognition tasks, and sliding-window method-based multi-target detection. Thanks to its intrinsic scalability, the demonstrated photonic computing architecture could be easily integrated on a single chip toward large-scale photonic neural networks with ultrahigh computation throughputs.