When developing the concept, the researchers envisioned pieces of glass that look like translucent squares, within which are embedded tiny strategically placed bubbles and impurities that bend light in specific ways to differentiate among different images. That, say the researchers, is the artificial intelligence in action.
“This is completely different from the typical route to machine vision,” says UW-Madison electrical and computer engineering professor Zongfu Yu. “We’re using optics to condense the normal setup of cameras, sensors, and deep neural networks into a single piece of thin glass."
For their proof of concept, the engineers devised a method to make glass pieces that identified handwritten numbers. Light emanating from an image of a number enters at one end of the glass, and then focuses to one of nine specific spots on the other side, each corresponding to individual digits.
The glass was dynamic enough to detect, in real time, when a handwritten "3" was altered to become an "8."
“The fact that we were able to get this complex behavior with such a simple structure was really something,” says Erfan Khoram, a graduate student in Yu’s lab.
Designing the glass to recognize numbers was similar to a machine-learning training process, say the researchers, except that they "trained" an analog material instead of digital codes. Specifically, the engineers placed air bubbles of different sizes and shapes as well as small pieces of light-absorbing materials like graphene at specific locations inside the glass.
“We’re accustomed to digital computing, but this has broadened our view,” says Yu. “The wave dynamics of light propagation provide a new way to perform analog artificial neural computing.”
One advantage of the approach is that the computation is completely passive and intrinsic to the material - i.e., one piece of image-recognition glass could be used hundreds of thousands of times. Although the up-front training process could be time consuming and computationally demanding, the glass itself is easy and