A team of scientists has created a unique all-optical platform that can perform machine learning tasks at light speed.
The researchers, headed by Aydogan Ozcan, professor of electrical and computer engineering at UCLA, used a 3D printer to create an artificial neural network that uses light waves to process information and power the device. The result is the all-optical Diffractive Deep Neural Network (D2NN) architecture. The team published its findings in Science magazine.
The D2NN, which looks like the inside of a solid gold car battery, is made up of a series of 3D-printed translucent polymer panels. Each panel has an uneven surface made up of tens of thousands of artificial neuron “pixels” etched into it—resulting in what Ozcan describes as a complex maze of glass and mirrors.
The panel diffracts light coming from the target object in various directions and allows for light to travel between each layer—following a deep learning-based design of passive diffractive layers that work together. "The light enters a diffractive network and bounces around the maze until it exits. The system determines what the object is by where most of the light ends up exiting," said Ozcan.
The optical network is ideally suited for tasks such as image analysis, facial recognition and object classification. The device performs these tasks faster than conventional systems by using the very light bouncing off an object to identify the object—faster than the standard computer-based neural network.
During the testing process, the team trained the network to recognize and classify handwritten numbers from zero to nine, and in a more demanding test, to categorize clothing items including pants, dresses, coats, sandals, shirts, sneakers, bags and shoes.
"This work opens up fundamentally new opportunities to use an artificial intelligence-based passive device to instantaneously analyze data, images and classify objects," said Ozcan.
Amazingly, the device performs these tasks without using any electricity—it operates on the power of the light beam itself, using only passive components. And thanks to 3D printing, it only cost $50 to make the drive’s materials.
Researchers on the team envisage possibilities for D2NN architectures performing specialized tasks in microscopes and cameras. For example, your camera could identify the person you’re taking a picture of the instant you snap the photo, tag the photo with the subject’s name, and then post it to your social media time line.
Read more about machine learning at Computers Use Machine Learning to Detect Radiation Damage Better Than Humans Do.