All-optical inference engine based on a shape-optimized multimode optical fiber

The increasing demand for high-speed and power-efficient computation, along with progress in photonics, led to extensive investigation of all-optical computation methods in recent years; this is because they have parallel computation ability and high power efficiency compared with their electrically implemented counterparts. As a step toward fiber-based all-optical computing, we present here a proof-of-concept all-optical inference engine implemented using an optomechanical system based on optimization of the shape of a multimode optical fiber. The positions of four points along the fiber are optimized using a genetic algorithm to concentrate light at specific points at the output of the fiber, enabling an optical classification of the pattern of light input to the fiber. We demonstrate recognition of images of a single digit among images of all digits input to the fiber with an average accuracy of $\ensuremath{\sim}80\mathrm{%}$.