Class-specific differential detection improves the inference accuracy of diffractive optical neural networks

Abstract

We report new design strategies to increase the inference accuracy of Diffractive Deep Neural Networks (D2NNs). Using a differential detection scheme that is combined with the joint-training of multiple D2NNs, each specialized on a single object-class, D2NN-based all-optical classification systems numerically achieve blind-testing accuracies of 98.52%, 91.48% and 50.82% for MNIST, Fashion-MNIST and grayscale CIFAR-10 datasets, respectively. Furthermore, using three independently-trained D2NNs that project their light onto a common output plane enables the system to achieve 98.59%, 91.06% and 51.44%, respectively. Through these systematic improvements, the reported blind-inference performance sets the state-of-the-art for an all-optical neural network design.