Physics-aware Differentiable Discrete Codesign for Diffractive Optical Neural Networks

  • 2022-09-28 18:13:28
  • Yingjie Li, Ruiyang Chen, Weilu Gao, Cunxi Yu
  • 1

Abstract

Diffractive optical neural networks (DONNs) have attracted lots of attentionas they bring significant advantages in terms of power efficiency, parallelism,and computational speed compared with conventional deep neural networks (DNNs),which have intrinsic limitations when implemented on digital platforms.However, inversely mapping algorithm-trained physical model parameters ontoreal-world optical devices with discrete values is a non-trivial task asexisting optical devices have non-unified discrete levels and non-monotonicproperties. This work proposes a novel device-to-system hardware-softwarecodesign framework, which enables efficient physics-aware training of DONNsw.r.t arbitrary experimental measured optical devices across layers.Specifically, Gumbel-Softmax is employed to enable differentiable discretemapping from real-world device parameters into the forward function of DONNs,where the physical parameters in DONNs can be trained by simply minimizing theloss function of the ML task. The results have demonstrated that our proposedframework offers significant advantages over conventional quantization-basedmethods, especially with low-precision optical devices. Finally, the proposedalgorithm is fully verified with physical experimental optical systems inlow-precision settings.

 

Quick Read (beta)

loading the full paper ...