Optical Scattering Enables Simulation Trained Classification of Images
Davis Garwood, Chloe Widman, Wolfgang Losert, Corey Hart and Joseph T. Schick
Optical scattering is a powerful approach to convolve a signal with a physical filter in a manner analogous with a reservoir computing model. Here we investigate the ability of such filters to encode and decode light that is spatially shaped into handwritten digits from the MNIST dataset and then scattered through a solid medium with embedded PMMA (polymethyl methacrylate) microspheres. Using the collected scattering patterns, a model based on event triggered averaging was constructed capable of decoding presented digits from a scattering pattern. This approach is verified using training and validation on a simulation of scattering. Most significantly, we show that input classification and decoding from optical scattering can be achieved with training on simulations of scattering, yielding accurate classification of experimentally measured scattering patterns.
Keywords: Optical computing, neuromorphic computing, optical scattering, machine learning, artificial intelligence, Non-Von Neumann computation
DOI: 10.32908/ijuc.v20.210126