On March 3, the team led by Academician Gu Min from the Center for Artificial Intelligence Nanophotonics at the Future Optics Laboratory of Shanghai University of Science and Technology published a high-level paper entitled "Nanoprinted high-neuron-density optical linear perceptrons performing near-infrared inference on a CMOS chip" in the journal Light: Science and Applications. The paper proposes a research scheme for all-optical inference holographic nanostructures in the field of nanofabrication technology.
Today, unlocking smartphones with facial recognition is commonplace. However, few people realize the time-consuming and energy-intensive process behind it: sensors first collect optical information about the face and send it to a neural network in a computer. The visual information is then converted into electronic information by electronic hardware before being displayed. Gu Min's team has innovatively developed a new concept: a compact optical diffraction neural network. This network can perform all-optical inference and can be directly integrated with commercial CMOS sensors. In other words, this technology eliminates the traditional light-to-electricity conversion process, allowing optical information processing to be completed directly in the optical domain. This fast, energy-efficient, and functional optoelectronic device can be applied to fields such as security checks, medical imaging, autonomous driving, art appreciation, and satellite image processing. Compared to existing solutions, it occupies less space, consumes less energy, and is less expensive.
"Our nanoprinted inference perceptron in the visible and near-infrared bands has a computational power ceiling of 400 ExaFLOPS (10¹⁸ FLOPS of floating operations per second), which is 3 to 5 orders of magnitude higher than diffraction devices and integrated photonic hardware operating in the millimeter wave and microwave bands," said Dr. Elena Goi, the paper's first author and a core researcher in the team.
"By using super-resolution 3D nanofabrication technology, we can directly integrate AI optical devices into existing imaging sensors. This is equivalent to placing custom-made smart glasses on the imaging sensor, designed for specific tasks, which can process the incoming optical information before it is detected," Academician Gu Min explained.
This research was supported by a major project funded by the Special Development Fund of Shanghai Zhangjiang National Innovation Demonstration Zone.
Dr. Elena Goi conducts experiments
