The chip can use light to train nonlinear neural networks - a breakthrough that could significantly speed up AI training, reduce energy consumption and even pave the way for all-optical computers.
Current AI chips are electronic and rely on electricity to perform calculations, but this new chip is photonic. Described in the journal Nature Photonics, the chip modifies the behavior of light to perform the nonlinear math at the heart of modern AI.
Most current AI systems rely on neural networks, software that mimics biological neural tissue. Just as neurons connect to enable organisms to think, neural networks enable AI systems to perform complex tasks by connecting layers of simple units, or "nodes."
The team's breakthrough started with a special semiconductor material that responds to light. When a beam of "signal" light (carrying input data) passes through the material, a "pump" beam shining from above adjusts how the material reacts.
By changing the shape and intensity of the pump light, the team can control whether the signal light is absorbed, transmitted or amplified, depending on the intensity of the light and the behavior of the material. This process "programs" the chip to perform different nonlinear functions.
It is worth noting that this research did not change the basic structure of the chip, but used the patterns formed by light inside the material to reshape the way light travels through. This creates a reconfigurable system that can express a variety of mathematical functions based on the pumping pattern, giving it the ability to learn in real time and adjust its behavior based on output feedback.
To verify the capabilities of the chip, the team used it to solve a number of benchmark AI problems. In simple nonlinear decision boundary tasks, an accuracy of more than 97% was achieved; in the famous iris data set problem, an accuracy of more than 96% was achieved. This shows that compared with traditional digital neural networks, photonic chips not only have comparable or even better performance, but also lower energy consumption because they reduce reliance on power-consuming components.
In addition, the experiment also showed that only 4 nonlinear optical connections can achieve the effect of 20 fixed nonlinear activation function linear electronic connections in the traditional model, demonstrating the great potential of the technology. As the architecture is further expanded, the efficiency will be even more significant.
Unlike previous photonic systems that are fixed after manufacturing, this new chip provides a blank platform that can use pump light to draw programmable instructions like a paintbrush. It is a practical proof of the concept of field-programmable photonic computers and marks an important step towards training AI at the speed of light.