348 — All-Optical Machine Learning Using Diffractive Deep Neural Networks
This research reminds me a lot of a SIGGRAPH video paper I watched recently, in which the caustics of a refractive surface are leveraged in order to form a high-contrast image on a focal plane.
Using a similar technology of “choreographed caustics,” these researchers engineered surfaces that redirected light in highly specific ways, with each surface redirecting light to a set of downstream targets such that consecutive layers of a neural network are represented by the consecutive diffractions of light along each surface.
To demonstrate this technique, the authors design an MNIST classifier that accepts as input a “vector” set of 2D beams of light in the pixel formation of a digit, and outputs a classification, where the classification is represented as the output region with the maximum light energy. Light is scattered through the highly structured lens-like plates of 3D printed “network” at the speed of light.
This technology reduces the runtime of a network greatly because layer-to-layer transmission occurs at the speed of light; though it greatly increases the ‘training’ time since the training process must now include the diffractive optics mathematics to construct the diffractor planes.
This is really interesting technology for use in cases where the task does not change over time but the runtime of each operation is crucial. Coincidentally, MNIST digit classification in a post-office setting is a perfect example of this use-case.