“All-Optical Machine Learning Using Diffractive Deep Neural Networks,” by Xing Lin, Yair Rivenson, Nezih T. Yardimci, Muhammed Veli, Mona Jarrahi, and Aydogan Ozcan of UCLA, was posted to arXiv in April 2018 and published in Science in 2018. It demonstrated a neural network that computes with light passing through physical layers rather than with numbers shuffled through a processor.
The architecture, which the authors called a Diffractive Deep Neural Network or D2NN, is a stack of thin transmissive layers. Each point on each layer acts like a tiny tunable element that bends and delays the light passing through it. The pattern of these elements is found ahead of time by ordinary deep-learning training on a computer, but once the layers are manufactured they are passive: no power, no electronics, no clock. The team 3D-printed the layers and showed them working at terahertz frequencies, where the network learned to classify handwritten digits and even to act as an imaging lens.
The appeal is that the computation happens at the speed of light and consumes essentially no energy once the device is built, because the answer emerges from the physics of wave propagation rather than from billions of multiply-accumulate operations. The trade-offs are equally clear: the network is fixed at fabrication, it handles only the optical task it was designed for, and accuracy lagged digital networks. The work is a proof of concept for a broader idea - that some machine-learning inference could be offloaded onto optics.
Why business readers should care: the energy cost of AI inference is now a first-order constraint, and optical computing is one of the speculative routes to doing it far more cheaply. This paper is a foundational demonstration that “computing with light” is more than a metaphor, even if commercial use is still early.