Diffractive deep neural network adjoint assist or (DNA)2: a fast and efficient nonlinear diffractive neural network implementation

Abstract

The recent advent of diffractive deep neural networks or D2NNs has opened new avenues for the design and optimization of multi-functional optical materials; despite the effectiveness of the D2NN approach, there is a need for making these networks as well as the design algorithms more general and computationally efficient. The work demonstrated in this paper brings significant improvements to both these areas by introducing an algorithm that performs inverse design on fully nonlinear diffractive deep neural network - assisted by an adjoint sensitivity analysis which we term (DNA)2. As implied by the name, the procedure optimizes the parameters associated with the diffractive elements including both linear and nonlinear amplitude and phase contributions as well as the spacing between planes via adjoint sensitivity analysis. The computation of all gradients can be obtained in a single GPU compatible step. We demonstrate the capability of this approach by designing several types of three layered D2NN to classify 8800 handwritten digits taken from the MNIST database. In all cases, the D2NN was able to achieve a minimum 94.64% classification accuracy with 192 minutes or less of training.