Fast/ energy-efficient optical AI computer training using real measurement reinforcement learning instead of digital twin approximation

https://techxplore.com/news/2026-01-free-optical-ai.html

"overcomes failure due to difficulty modeling small hardware shifts in hardware, environmental interference of light paths, light propagation too complex for perfect simulation... optical processor learns on-site without need to understand underlying physics, prevents drastic/ unstable changes during training, algorithm reuses data reducing experiments needed to train... efficiently focuses light through random/ unknown diffuser, holography/ aberration correction, image classification, photonic accelerators"