Since their conception as Diffractive Deep Neural Networks (D2NNs) by Lin in 2018, the field of optical information processing with sub-wavelength patterned thin scatterers has raised significant attention, as it not only allows for leveraging the inherent advantages of optical signals, such as parallelism and high speed, but also allows for convenient processing of optical information in its native domain. The nanoprinting of such diffractive networks with two-photon polymerization methods is of interest, as it allows for fabrication of optical inference systems with record-high neurons densities, targeted for the VIS/NIR wavelength regime and fabricated directly on commercial CMOS imaging sensors. Current methods for in-silicon training of these Diffractive Neural Networks employ numerical models of ‘masked’ detectors, where only certain areas of the networks output plane are compared for performance evaluation, while the remainder of the detector area is being ignored. While such training methods have the advantage of straightforward numerical implementation, for fabricated devices where typically the whole output plane in captured for data acquisition, this can lead to significantly reduced performance due to strong background signals in the areas that were ‘masked off’ during training and hence susceptibility to errors in aligning the distinct detector areas for readout. In this contribution we critically discuss methods for training of diffractive neural networks which consider the whole output plane of the network in order to achieve low background noise and hence high detector-to-background contrast for 3D nanoprinted diffractive neural networks with increased experimental robustness due to reduced susceptibility for errors in alignment of detector areas.
|