Training a Neural Network with Exciton-Polariton Optical Nonlinearity
Andrzej Opala, Riccardo Panico, Vincenzo Ardizzone, Barbara Piętka, Jacek Szczytko, D. Sanvitto, Michał Matuszewski, Dario Ballarini · Physical Review Applied · 2022
In contrast to software simulations of neural networks, hardware implementations have limited or no tunability. While such networks promise great improvements in terms of speed and energy efficiency, their performance is limited by the difficulty of applying efficient training. We propose and realize experimentally an optical system where highly efficient backpropagation training can be applied through an array of highly nonlinear, nontunable nodes. The system includes exciton-polariton nodes realizing nonlinear activation functions. We demonstrate high classification accuracy in the MNIST handwritten digit benchmark in a single-hidden-layer system.