A new research has reported that a newly created, 3D-printed, optical deep learning network enables computational problems to be performed at the speed of light.
The progress provides a low-cost, efficient, and scalable means to develop deep learning systems, which are speeding up the frontiers of science, for instance, in language translation, medical image analysis, image classification, and much more. The optical deep learning framework developed by Xing Lin and his team comprises of layers of 3D-printed, optically diffractive surfaces that function collectively to process information.
The system, named Diffractive Deep Neural Network (D2NN), works when each point on a given layer transmits or reflects an incoming wave, representing an artificial neuron connected to other neurons of the subsequent layers through optical diffraction. Each “neuron” can be tuned by modifying the amplitude and phase. The scientists put D2NN to work and the trained the system by exposing it to 55,000 images of handwritten digits, from zero to nine.
Following the training, D2NN was able to identify these numbers with an accuracy of 95.08%, and the researchers have defined ways to further enhance the accuracy, for instance, adding additional “neural” layers. The researchers stated that this system could easily be optimized with different 3D fabrications techniques, detection systems, and optical components.