New light-based chips enable robotic learning without electricity

Researchers have developed novel photonic computing chips that enable neural networks to learn entirely through light signals, eliminating the need for electronic components in key learning operations. This breakthrough overcomes a significant limitation in previous photonic AI systems, which relied on electronics for nonlinear computations essential to learning and decision-making. The new design uses a photonic spiking neural system that mimics biological neurons by transmitting rapid optical pulses, allowing both linear and nonlinear neural computations to be performed directly in the optical domain. The system comprises two chips: a 16-channel photonic neuromorphic processor with 272 trainable parameters and a laser array with a saturable absorber enabling low-threshold nonlinear optical spiking. The researchers demonstrated the system’s capabilities through reinforcement learning tasks, such as balancing a pole on a moving cart and stabilizing an inverted pendulum, achieving accuracy within 1.5-2% of software models. The photonic chips delivered impressive computational performance, with linear processing at 1.39 tera operations per second