End-to-end training and inference in integrated photonic deep neural networks

Photonic neural networks can significantly boost the performance of their digital electronic counterparts by benefiting from large optical bandwidth, low-loss signal transmission, and various multiplexing schemes. Hence, high accuracy and robust training and inference in such networks are critical for their large-scale deployment. Here, we demonstrate integrated end-to-end photonic deep neural networks that offer scalability, low-latency signal processing, and robust and repeatable end-to-end backpropagation training. Moreover, we use standard silicon photonic platforms which enable high-yield and low-cost fabrication of large-scale photonic neural networks for the next generation of deep learning systems.