University of Pennsylvania engineers have developed a new chip that trains artificial intelligence (AI) using light instead of electricity. This chip is photonic, meaning it works with light beams. Unlike regular chips that use electricity, this one cuts energy use and speeds up AI training. A paper published in Nature Photonics describes this invention.

The chip handles nonlinear functions, which help AI learn complex tasks. Without nonlinear functions, AI cannot learn well because it stays too simple. Regular chips manage linear math, where inputs just add up. Until now, no chip could do nonlinear math with light alone. The Penn engineers claim they solved this challenge.

How the chip works and performs

The chip uses a special material that reacts to light. A signal light carries data through the material. Another light, called a pump beam, shines from above. The pump beam changes how the material works. This setup lets the chip control the signal light’s behavior.

By adjusting the pump beam, the chip performs different nonlinear math tasks. This makes the chip flexible and programmable. The chip learns by adjusting based on its results. It acts like a blank canvas where light draws instructions.

The engineers tested the chip on AI tasks. It scored over 97% accuracy on a simple decision task. On the Iris flower dataset, a common AI test, it scored over 96%. The chip matched or beat regular digital AI systems. It used fewer steps and less energy.

One test showed four light connections on the chip equaled 20 electronic ones. This shows the chip’s efficiency. The chip could change how AI works in the future. It might handle bigger tasks like training language models. By using light, it cuts energy use in AI data centers. This invention could lead to fully light-powered computers.

“This is a true proof-of-concept for a field-programmable photonic computer,” says research leader Lian Feng in a press release.

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.