Matrix Optimization on Universal Unitary Photonic Devices

Networks of tunable, integrated optical interferometers support quantum information processing and machine learning with much better energy efficiency than standard electronics. A network's gridlike structure and imperfections localize optical signals propagating through the device, which ultimately slows training by gradient-based optimization. Here this problem is solved by proper initialization, combined with redundant and remotely interacting interferometers. The authors' approach improves the convergence time of gradient-based optimization to random target operators by at least two orders of magnitude, at the scale of practical machine-learning applications (10⁴ to 10⁶ nodes).