Table A1.

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. The 2-stage hybrid system training discussed in the Methods section was not used here. Instead, D²NN and 5 different digital neural networks were jointly-trained at the same time from scratch. 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 materials, with 5 diffractive layers. Yellow and blue colors refer to Δ _z = 40×λ and Δ _z = 4×λ, respectively.