News
As per the latest issue of Nature Photonics, a research team from the University of Pennsylvania has developed the first programmable chip capable of using light to perform nonlinear neural network training. This breakthrough could significantly accelerate artificial intelligence (AI) training while reducing energy consumption, which lays the foundation for fully light-driven computers.
Previously, although research teams had developed optically driven chips capable of handling linear mathematical operations, the challenge of representing nonlinear functions using purely optical methods had remained unsolved. Nonlinear functions are essential for training deep neural networks—without them, photonic chips cannot carry out deep learning or perform complex intelligent tasks.
The new advancement relies on a special semiconductor material that is sensitive to light. When a “signal” light carrying input data passes through the material, another beam of “pump” light is shone from above to modulate the material’s response characteristics. By adjusting the shape and intensity of the pump light, the team can control how the signal light is absorbed, transmitted, or amplified, based on its intensity and the material’s reaction—effectively programming the chip to execute various nonlinear functions.
Notably, this research does not alter the basic structure of the chip. Instead, it reshapes the way light travels through the material by forming patterns internally using the pump light. This creates a reconfigurable system capable of expressing multiple mathematical functions through pump light patterns, endowing it with real-time learning capabilities and the ability to adapt its behavior based on output feedback.
To validate the chip’s capabilities, the team used it to tackle several benchmark AI tasks. In a basic nonlinear decision boundary task, it achieved over 97% accuracy; in the well-known Iris dataset problem, it reached more than 96% accuracy. These results indicate that compared to traditional digital neural networks, the photonic chip not only performs on par or better but also consumes less energy by reducing reliance on power-hungry electronic components.
Additionally, the experiments showed that just four nonlinear optical connections could match the performance of 20 fixed nonlinear activation function-based electronic connections in traditional models, highlighting the immense potential of the technology. With further architectural scaling, efficiency is expected to improve even more.
Unlike previous photonic systems that were fixed after fabrication, this new chip offers a blank-slate platform. It allows programmable instructions to be “painted” onto it using pump light—providing a practical demonstration of the concept of field-programmable photonic computing and marking a crucial step toward training AI at the speed of light.
