Spatiotemporal chaos-based photonic neural networks

Abstract

In advanced machine learning tasks, artificial neural networks are frequently utilized; yet, as a result of von Neumann bottleneck-limited hardware, they require large data sets and significant power consumption. Chaotic dynamical systems are effective tools for reservoir computing applications, and optics provide a platform for high-speed computation. Here, we present a chaotic optical neural network that harnesses the complex modal energy flow dynamics of a multimode fiber to perform energy-efficient machine learning. The proposed architecture addresses the energy and time problems facing today's systems by benefiting from the butterfly effect. In biomedical and satellite-based scene classification tasks, our photonic neural network performs exceptionally well. Our novel methodology illustrates how chaotic dynamics can be utilized in machine learning and optical computing.