This website requires JavaScript.

Decision-making and control with metasurface-based diffractive neural networks

Jumin QiuTianbao YuLujun HuangAndrey MiroshnichenkoShuyuan Xiao
Dec 2022
The ultimate goal of artificial intelligence is to mimic the human brain toperform decision-making and control directly from high-dimensional sensoryinput. All-optical diffractive neural networks provide a promising solution forrealizing artificial intelligence with high-speed and low-power consumption. Todate, most of the reported diffractive neural networks focus on single ormultiple tasks that do not involve interaction with the environment, such asobject recognition and image classification, while the networks that canperform decision-making and control, to our knowledge, have not been developedyet. Here, we propose to use deep reinforcement learning to realize diffractiveneural networks that enable imitating the human-level capability ofdecision-making and control. Such networks allow for finding optimal controlpolicies through interaction with the environment and can be readily realizedwith the dielectric metasurfaces. The superior performances of these networksare verified by engaging three types of classic games, Tic-Tac-Toe, Super MarioBros., and Car Racing, and achieving the same or even higher levels comparableto human players. Our work represents a solid step of advancement indiffractive neural networks, which promises a fundamental shift from thetarget-driven control of a pre-designed state for simple recognition orclassification tasks to the high-level sensory capability of artificialintelligence. It may find exciting applications in autonomous driving,intelligent robots, and intelligent manufacturing.