Training a Neural Network with Exciton-Polariton Optical Nonlinearity

In contrast to software simulations of neural networks, hardware implementations have limited or no tunability. While such networks promise great improvements in terms speed and energy efficiency, their performance is by the difficulty applying efficient training. We propose realize experimentally an optical system where highly backpropagation training can be applied through array nonlinear, nontunable nodes. The includes exciton-polariton nodes realizing nonlinear activation functions. demonstrate high classification accuracy MNIST handwritten digit benchmark a single-hidden-layer system.