With all the AI and cryptocurrency mining going on, it’s getting harder and harder to find a fast GPU. But even the most powerful GPUs aren’t as fast as light, which travels 186,282 miles per second. That’s why a neural network built by a group of UCLA researchers is so intriguing. Instead of the electrical impulses that send ones and zeros through computer circuit boards in traditional neural networks, UCLA’s diffractive deep neural network (D2NN) transmits light pulses between multiple 3D-printed plastic plates (which serve as layers). In lieu of transmitting alternating digits, the optical neural network either shines or reflects light to the next layer. When applied to a traditional image recognition task, the D2NN correctly identified images of different numbers with 91.75 percent accuracy, all at the speed of light. It’s just the beginning, but D2NNs could be a solution to the need for computing power and speed in our AI future, assuming we can make space for all those plastic plates.
sign up for our newsletter