Reinforcement learning in a large scale photonic networks

Neural network computational architectures have been inspired by the physical structure of the human brain. There, a large number of nonlinear nodes are connected into network, and learning is generally interpreted via modifications to the networks connectivity structure. Yet, implementations of neural networks are typically based on their emulation within serial and binary logic computers. Recently, Reservoir Computing (RC) has been demonstrated with various levels of hardware implementations inside nonlinear networks implemented in a delay system. I will demonstrate how the same scheme can be realized in a photonic networks of 1600 individual nonlinear oscillators. Furthermore, I will report on results of a full hardware‐based learning scheme for such networks. Our system therefore presents a significant step to a full physical implementation of the originally envisioned neuromorphic computing systems.