An artificial intelligence chip smaller than a grain of salt can perch at the end of an optical fibre, harnessing the physics of light to process the information passed through the fibre 鈥� while using far less energy and computational power than typical AI techniques.
Optical fibres can carry data at the speed of light, but decoding these light signals usually happens on a slower and more energy-intensive external computing device. The new AI chip could perform that data-processing task more quickly while using fewer resources, enabling things like simpler medical imaging devices or more efficient chips for a future quantum internet.
The device itself can be thought of as a passive, well-trained neural network that physically manipulates light to perform computations, says at the University of Shanghai for Science and Technology in China.
Advertisement
This type of light-based 鈥渄iffractive neural network鈥� was first developed by researchers at the University of California, Los Angeles, in 2018. Yu and his colleagues took that work a step further by making an extremely tiny version of such an AI chip and placing it on the end of an optical fibre the width of a human hair.
The chip can reconstruct images of numerals barely larger than a grain of pollen. Image reconstruction isn鈥檛 a new task for AI, but the size of this chip reduced the computational resources needed to a few thousandths of the amount AIs typically use.
Such an AI chip can process data flowing through fibre-optic cables as fast as light travels through its layers, doing computations within trillionths of a second, says at the University of California, Los Angeles. 鈥淥verall, I am excited and positive about this work 鈥� and it has the potential to scale up [for wider usage].鈥�
There are still some challenges. Slight imperfections could lead to changes in performance between each device, says Ozcan. That is due in part to the layered structure of the chip and to the way the researchers manufactured it. They etched light-manipulating patterns onto thin polymer layers, carefully aligning each layer and then integrating the chip with an optical fibre. Because of this multi-step process, it is difficult to achieve consistency and accuracy across every device.
The fixed physical design of the chips also means they are custom made for specific jobs. 鈥淎s the tasks change or the fibre-optic system changes, one would need to have a new design to be fabricated and integrated with the fibre,鈥� says Ozcan.
Still, he doesn鈥檛 consider any of those challenges to be major roadblocks to commercialisation. His own group鈥檚 work suggests this new development could lead to simplified and more capable designs for endoscopic imaging devices 鈥� tiny cameras in tube form that are inserted into the body during certain medical examinations.
The fibre-integrated chips could also be used within quantum photonic chips that are designed to control and manipulate light on a quantum level, says Yu. Such devices may eventually form the backbone of a quantum internet, a practically unhackable communication network.
Nature Photonics
