Training a neural network with exciton-polariton optical nonlinearity
This work addresses the challenge of efficient training in hardware neural networks for improved speed and energy efficiency, though it appears incremental as it builds on existing optical and nonlinear node concepts.
The authors tackled the problem of training hardware neural networks with limited tunability by proposing an optical system using exciton-polariton nodes for nonlinear activation, achieving high classification accuracy on the MNIST benchmark.
In contrast to software simulations of neural networks, hardware implementations have often limited or no tunability. While such networks promise great improvements in terms of speed and energy efficiency, their performance is limited by the difficulty to apply efficient training. We propose and realize experimentally an optical system where highly efficient backpropagation training can be applied through an array of highly nonlinear, non-tunable nodes. The system includes exciton-polariton nodes realizing nonlinear activation functions. We demonstrate a high classification accuracy in the MNIST handwritten digit benchmark in a single hidden layer system.