The research published in Optica under the title “All-optical neural network with non-linear activation functions” discloses a fully built All-Optical Neural Network (AONN) in which linear operations were programmed by spatial light modulators and Fourier lenses, while non-linear optical activation functions were realized based on electromagnetically induced transparency (EIT), a light-induced quantum interference effect among atomic transitions.
“This light-induced effect can be achieved with very weak laser power,” noted lead professor Shengwang Du. Because this effect is based on non-linear quantum interference, the researchers expect such a system could be extended to a quantum neural network that would be able to solve problems that could not be solved by classical methods.
In their paper, the researchers report a proof-of-concept two-layer, fully connected AONN with 16 inputs and two outputs. The team used its all-optical network to classify the order and disorder phases of the Ising model, a statistical model of magnetism. The results showed that the all-optical neural network was as accurate as a well-trained computer-based neural network.
“Our all-optical scheme could enable a neural network that performs optical parallel computation at the speed of light while consuming little energy,” co-author Junwei Liu said.
Moreover, the authors note that because the linear matrix elements and non-linear functions can be independently programmed, one system could easily be reconfigured to realize different AONN architectures, without modify its physical structure.
Next, the researchers plan to expand their all-optical approach to large-scale all-optical deep neural networks with complex architectures designed for specific applications such as image recognition.
Hong Kong University of Science and Technology – www.ust.hk