Many environmentalists see electrical power as one of artificial intelligence’s major constraints, as processing data, running complex software models and cooling a data center requires a continuous, high-density flow of electricity.
Northeastern University engineer Yongmin Liu sees power and AI in a different light. Quite literally.
“We really want to develop new optical devices that use light to manipulate and to detect, transmit, and also to do some processing,” said Liu, a professor of mechanical and industrial engineering and electrical and computer engineering at Northeastern.
“By using light, we can do machine learning and deep learning with extremely low power consumption,” Liu said.
Liu develops optical, or photonic, metasurfaces that can control light. These surfaces include a human-made array of nanostructures, extremely miniscule structures measured in nanometers and that are not visible to the naked eye.
With a recent grant from the National Science Foundation, Liu wants to “establish a virtuous loop between photonics and AI.” In other words, using artificial intelligence to design new photonic structures, devices and systems, which can then be used to propel advancements in AI that can be applied once more to photonics and so on and on…
A major goal is to develop new metasurfaces that expand upon manipulating only one property of light, such as the angle by which light hits a surface, into simultaneously controlling multiple properties like color or frequency.
Liu sat down with Northeastern Global News to explain how light can both power and advance AI and more.
The interview has been edited and condensed for clarity.
What is photonics?
Photonics essentially is part of the understanding and the application of light. We try to understand the nature of light and also try to see how light can be used for different applications. Typically, we consider light as an electromagnetic wave, but from quantum mechanics, we can consider light as a stream of particles called photons. Photons are very similar to electrons, and just like electronic devices that control electrons, we have to understand how we can manipulate those photons.
What are some of the benefits of photons and using them to transmit information?
There are distinct advantages. Essentially, we can have a much higher speed and also much lower energy consumption compared with the traditional digital or electronic platform. Artificial intelligence is very popular nowadays, but if we just use a traditional electronic platform, it uses a significant amount of power. By using light, we can have extremely low power consumption to do similar functions like machine learning, deep learning, image generation…and at extremely high speed, like the speed of light.
What are some applications of this technology?
There is a surging interest in developing distinct photonic platforms – instead of the existing electronic or digital hardware – to realize optical neural networks, which use light instead of electricity to run artificial intelligence models.
This has a wide range of applications including image classification (the process of categorizing or classifying images into predefined categories) computing, and communications. Other applications include holograms, where patterns of stripes and whorls organize light that is moving in the same direction at the same wavelength (for example, light from a laser beam) into a three-dimensional representation of an object; nano-printing; and virtual reality.
Is this technology related to fiber optics?
For this new NSF project, we want to enhance the so-called ‘information capacity,’ or the maximum rate or amount of data that can safely travel through a communication channel without error, for optical fiber that transmits the data. To do this we can use different properties of light, for instance, color, which is related to wavelength. Consider a highway with a single lane corresponding to a single color or wavelength of light carrying data. Essentially, if you have different wavelengths or colors traveling down the road, you can add more lanes and transmit more data through the optical fiber or the photonic devices we try to build.
