Efficient Optical AI Network with Low Power Programming

Researchers in Switzerland have made a groundbreaking advancement in the field of artificial intelligence by developing a low-power non-linear neural network within an optical fiber. This innovative approach, spearheaded by a team at EPFL, offers a more energy-efficient method for performing nonlinear computations for optical AI.

The traditional digital neural networks rely on transistors to carry out nonlinear transformations, which often necessitate powerful lasers in optical systems. However, the new approach developed at EPFL, known as nonlinear processing with only linear optics (nPOLO), utilizes a low-power continuous-wave laser and diffractive layers to enable simultaneous linear and nonlinear operations within the optical domain.

Christophe Moser, head of the Laboratory of Applied Photonics Devices at EPFL, explains, “In order to classify data in a neural network, each node must make a decision based on weighted input data, leading to a nonlinear transformation. Our method is up to 1,000 times more power-efficient than state-of-the-art digital networks, making it a promising platform for optical neural networks.”

The team at EPFL has also been working on developing a compiler to convert digital data into a format suitable for the optical AI system. This compiler will play a crucial role in bridging the gap between electronic and optical systems, paving the way for more widespread adoption of optical neural networks.

Demetri Psaltis, director of the Optics Laboratory, highlights the scalability and energy efficiency of their approach, stating, “Our image classification experiments have shown that our method is scalable and significantly more power-efficient than existing digital networks. By encoding pixels spatially on a low-power laser beam and performing multiple transformations, we can achieve the non-linearity required for neural network calculations at a fraction of the energy cost.”