Fig. 4: Diffractive all-optical transformation results for an arbitrary complex-valued nonunitary and invertible transform.
From: All-optical synthesis of an arbitrary linear transformation using diffractive surfaces
a Schematic of a K-layer diffractive network, that all-optically performs a linear transformation between the input and output fields-of-views that have Ni and No pixels, respectively. The all-optical transformation matrix due to the diffractive layer(s) is given by A′. b. The magnitude and phase of the ground truth (target) input-output transformation matrix, which is an arbitrarily generated complex-valued nonunitary and invertible transform. c All-optical transformation errors (see Eq. 5). The x-axis of the figure shows the total number of neurons (N) in a K-layered diffractive network, where each diffractive layer includes N/K neurons. Therefore, for each point on the x-axis, the comparison among different diffractive designs (colored curves) is fair as each diffractive design has the same total number of neurons available. The simulation data points are shown with dots and the space between the dots are linearly interpolated. d Cosine similarity between the vectorized form of the target transformation matrix in (b) and the resulting all-optical transforms (see Eq. 16). e Output MSE between the ground-truth output fields and the estimated output fields by the diffractive network (see Eq. 18). f The diffraction efficiency of the designed diffractive networks (see Eq. 19)
