A newly developed, 3-D printed optical deep learning network allows computational problems to be executed at the speed of light, a new study reports. The advancement offers a cheap, scalable and efficient way to create deep learning systems, which are accelerating the frontiers of science, for example in medical image analysis, language translation, image classification and more. The optical deep learning framework developed by Xing Lin and colleagues consists of layers of 3-D-printed, optically diffractive surfaces that work together to process information. The system, dubbed Diffractive Deep Neural Network (D2NN), works whereby each point on a given layer either transmits or reflects an incoming wave, which represents an artificial neuron that is connected to other neurons of the following layers through optical diffraction. By altering the phase and amplitude, each "neuron" is tunable. Putting D2NN to work, the researchers trained the system by exposing it to 55,000 images of handwritten digits, ranging from zero to nine. After training, D2NN could recognize these numbers with 95.08% accuracy, and the authors outline ways to further boost accuracy, such as adding additional "neural" layers. The authors note that this system could easily be scaled up by using different 3-D fabrications methods, optical components, and detection systems.
###
Science