Ensemble learning enhances the inference accuracy of diffractive deep neural networks

Diffractive deep neural networks (D²NNs) form an optical computing framework, which utilize deep learning-based optimization to design diffractive surfaces that collectively execute a desired optical mapping or statistical inference between an input and output plane. Here, we demonstrate the use of ensemble learning and feature engineering to significantly improve the inference performance of diffractive optical systems for object recognition. We trained an initial collection of 1252 unique D²NNs, where miscellaneous object plane and Fourier plane filters were utilized to engineer and diversify the spatial and spectral features of the input object wavefront. Then, to reduce the size and complexity of the final D²NN ensemble, we designed a pruning algorithm, the basis of which is iterative elimination of D²NNs based on optimized weights assigned to them. This algorithm resulted in diffractive ensembles of 14 and 30 D²NNs, achieving a blind testing accuracy of >61% and >62%, respectively, on CIFAR-10 dataset, which constitute the highest inference accuracies achieved to date by any diffractive optical system on the same dataset.