Scattering-based Optical Computing

​While scattering is traditionally considered as an obstacle in optical imaging systems, recent

studies showed it can rather be exploited as an analog substrate of computing systems. In this

work, we demonstrate an optical computing scheme where optical patterns are transmitted

through strongly scattering layers, producing complex optical fields that encode nonlinear

transformations of the inputs. A generative deep learning model such as variational autoencoder (VAE)

trained on these speckle patterns reconstructs the original patterns and learns an efficient latent representation

of the scattering process. The results indicate that multiple scattering acts as a physical random projection

mechanism, effectively mixing features and thus enhancing separability and generalization.

Compared to direct imaging, the scattering system achieves improved noise tolerance and data

augmentation with less computational resources. This work highlights how optical disorder can

perform useful physical computation when combined with modern deep generative models.