The authors have tried to draw analogy between diffractive optics and neural nets. Diffractive optics, which is commonly used in beam shaping with lasers, is used with a terahertz source to “classify” masks based on MNIST by making beams going through the mask to be focused in different bucket detectors based on the number. It’s still not entirely clear to me why they don’t need the nonlinearity like what you need in regular NNs. This paper was published in Science.
Paper discussion: https://youtu.be/64vebcBGiUU