Optical micro-phase-shift dropvolume in a diffractive deep neural network

Abstract

To provide a desirable number of parallel subnetworks as required to reach a robust inference in an active modulation diffractive deep neural network, a random micro-phase-shift dropvolume that involves five-layer statistically independent dropconnect arrays is monolithically embedded into the unitary backpropagation, which does not require any mathematical derivations with respect to the multilayer arbitrary phase-only modulation masks, even maintaining the nonlinear nested characteristic of neural networks, and generating an opportunity to realize a structured-phase encoding within the dropvolume. Further, a drop-block strategy is introduced into the structured-phase patterns designed to flexibly configure a credible macro–micro phase dropvolume allowing for convergence. Concretely, macro-phase dropconnects concerning fringe griddles that encapsulate sparse micro-phase are implemented. We numerically validate that macro–micro phase encoding is a good plan to the types of encoding within a dropvolume.