Diffractive deep neural networks with multimode fibers

Optical computing is gaining traction as a viable alternative to digital neural networks, especially in machine learning applications, due to its potential for significantly reduced power consumption. Our work presents a multimode-fiber-based diffractive deep neural network operating in a linear regime without employing high peak power lasers. By mechanically perturbing the multimode fiber, we optimize the linear mode coupling for enhanced feature extraction within the optical processor. This method demonstrates substantial accuracies on real-world datasets, including breast cancer classification and satellite-based scene classification, reaching performance levels comparable to modern neural networks while training far fewer parameters. We demonstrate the efficiency of linear programmable photonic neural networks for low-power optical computing applications.