Space-efficient optical computing with an integrated chip diffractive neural network (Nat Commun 13, 1044)

Optical neural networks (ONNs) are a rapidly developing field of technology that uses light to perform complex computing tasks. Until now, most ONNs have been large and power-hungry, limiting their practical use. Traditional approaches rely on units called Mach-Zehnder interferometers (MZIs) for their calculations. For large-scale computations, this approach can require an excessive amount of units and power.

Here we introduce an integrated diffractive optical network (IDNN) that can achieve the similar performance using significantly fewer MZIs and two very compact diffractive cells, thus saving space and power. We tested the IDNN on two widely-used datasets and found that it could achieve the same level of accuracy as previous MZI-based ONNs, but with about a tenth of the footprint and power usage.