We’re pleased to share our latest findings: a theoretical investigation into the application of a bistable Fabry–Perot semiconductor laser under optical injection as an all-optical activation unit for multilayer perceptron optical neural networks.
In this study, we’ve examined the device’s ability to provide reconfigurable sigmoid-like activation functions with adjustable thresholds and saturation points. These functions were then benchmarked on machine learning image recognition problems.
The results? By adjusting a control parameter of the activation unit, we observed an increase in accuracy of up to 2% across various tasks. Notably, with a simple two-layer perceptron neural network, we achieved inference accuracies of up to 95% for the MNIST dataset and 85% for the Fashion-MNIST dataset.
This research sheds light on the potential of optical technology in enhancing machine learning capabilities. Stay tuned for further updates from our team.
Paper: link
Authors: Jasna Crnjanski, Isidora Teofilović, Marko Krstić, Dejan Gvozdić
