Performance Trends of Optical and Digital Neural Networks

Optical approaches to neuromorphic computing are gaining prominence due to their inherent parallelism, scalability, and reduced energy consumption compared to their digital counterparts [1]. Neuromorphic computing through wave dynamics, especially in optics, harnesses the intrinsic complexity of nonlinear optical phenomena, providing a fundamentally innovative approach to computation within physical substrates. However, these systems are often treated as black boxes, with limited insight into their underlying computational capabilities. In this work, we focus on parametrizing the inference capacity of an optical fiber-based Extreme Learning Machine (ELM), which leverages nonlinear frequency generation [2], providing a nonlinear mapping between input and output signals.