Diffractive Optical Neural Networks

Replacing the conventional imaging optics in machine vision systems with diffractive optical neural networks (DONNs) that leverage spatial light modulation and optical diffraction have been promising to enable new machine learning intelligence and functionality in optical domain and reduce computing energy and resource requirements in electrical domain. In this chapter, the fundamental models to describe and design DONNs are first reviewed. Passive DONNs systems operating in the terahertz and short wavelength ranges are then introduced. Moreover, the advanced architectures that are resilient to hardware imperfections, that demonstrate improved performance, and that are implemented in photonic integrated circuits, are discussed. Furthermore, the implementations of system reconfigurability through hybrid optoelectronic approaches are described. In addition, the effects from the physical model inaccuracy and how physics-aware training is used to correct deployment errors from both models and hardware are discussed. Finally, an all-optical reconfigurable DONNs system based on cascaded liquid-crystal spatial light modulators is demonstrated.