Table II.

Blind testing accuracies (reported in percentage) for all-optical (D²NN only), D²NN and perfect imager-based hybrid systems used in this work for Fashion-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 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. For the results reported in the all-optical part of this table, Fashion-MNIST objects were encoded in the amplitude channel of the input plane. When they are encoded in the phase channel (as in [15]), blind testing accuracies for a 5-Layer, phase-only (complex) D²NN classifier become 89.13% (89.32%) with Δ_z = 40×λ and 85.98% (88.54%) with Δ_z = 4×λ as reported in Table A2, as part of Appendix A.