Optical neural networks without electricity: how to pass light through a lens to instantly multiply matrices
University of California, Los Angeles
In 2018, a UCLA group 3D-printed a diffractive deep neural network (D²NN) from ordinary plastic. Five plastic plates perform matrix multiplication using light diffraction, achieving 91.75% accuracy on handwritten digit recognition with zero electricity after fabrication. Training is done on GPU, then weights are fixed in the plate geometry.
In 2018, a team led by Xin Lin and Aydogan Ozcan at UCLA 3D-printed an optical neural network called D²NN (Diffractive Deep Neural Network) from ordinary plastic. It consists of five plastic plates with microstructured surfaces placed in a terahertz beam (0.4 THz, wavelength 0.75 mm). Each plate acts as a neural network layer: the thickness of each microscopic pixel determines the phase shift (weight), diffraction between plates implements full connectivity, and the final interference pattern focuses light onto a 10-sector screen for digit classification. The network achieved 91.75% accuracy on MNIST digit recognition. During operation, no electricity is needed—only a light source. Training is performed on GPU via backpropagation, then the learned phase values are converted to physical thickness and 3D-printed. The system is linear, lacking nonlinear activation functions like ReLU; various approaches (electro-optical hybrids, two-photon absorption, electromagnetically induced transparency, acousto-optics) have been explored but none have reached commercial replacement.
- Abbreviations
- D²NN = Diffractive Deep Neural Network — Дифракционная глубокая нейронная сеть
- GPU = Graphics Processing Unit — Графический процессор
- MNIST = Modified National Institute of Standards and Technology database — База данных рукописных цифр
- ReLU = Rectified Linear Unit — Выпрямленная линейная единица
Source: Habr — хаб ML —
original