Table I.

Blind testing accuracies (reported in percentage) for all-optical (D²NN only), D²NN and perfect imager-based hybrid systems used in this work for MNIST dataset. In the D²NN-based hybrid networks reported here, 5 different digital neural networks spanning from a single fully-connected layer to ResNet-50 were co-trained with a D²NN design, placed before the electronic neural network. All the electronic neural networks used ReLU as the nonlinear activation function, and all the D²NN designs were based on spatially and temporally coherent illumination and linear optical materials, with 5 diffractive layers. For a discussion on methods to incorporate optical nonlinearities in a diffractive neural network, refer to [15]. Yellow and blue colors refer to Δ_z = 40×λ and Δ_z = 4×λ, respectively.