Taming Optical Computing Architectures: From Extreme Learning Machines to Diffractive Neural Networks

Analog processing has recently gained new traction as a result of developments in alternative computing paradigms to the von Neumann one. In particular, concerning optical computing, neuromorphic inspired frameworks strongly simplified the approach to data processing in the optical domain, now opening interesting perspectives for faster and energy-efficient solutions. Yet, while interesting and versatile architectures such as diffractive neural networks [1] and optical extreme learning machines[2]–[4] have been materializing all of this potential, an extensive comparison between distinct architectures is still missing, mostly due to the lack of effective performance metrics and comparable implementations.