Photonic next-generation reservoir computer based on distributed feedback in optical fiber

Area of Science:

• Optics and Photonics
• Machine Learning
• Dynamical Systems Analysis

Background:

• Reservoir computing (RC) is a powerful machine learning technique for analyzing dynamical systems.
• Photonic RCs offer low-latency predictions but are limited by physical cavities, hindering memory control and requiring long warm-up times.
• Existing photonic RCs face challenges in memory structure control and transient elimination due to reliance on nonlinear physical cavities.

Purpose of the Study:

• To develop a photonic next-generation reservoir computer (NG-RC) that overcomes the limitations of cavity-based photonic RCs.
• To demonstrate a fiber optic platform for photonic NG-RC that eliminates the need for a physical cavity.
• To achieve high-performance prediction tasks on complex dynamical systems using the novel photonic NG-RC.

Main Methods:

• A fiber optic platform was utilized to construct the photonic NG-RC, avoiding the need for a nonlinear physical cavity.
• Feature vectors were generated through nonlinear combinations of input data with varying delays.
• Rayleigh backscattering and coherent, interferometric mixing followed by quadratic readout were employed to produce output feature vectors.

Main Results:

• The photonic NG-RC demonstrated state-of-the-art performance on the observer (cross-prediction) task for Rössler, Lorenz, and Kuramoto-Sivashinsky systems.
• The system successfully generated feature vectors without a physical cavity, utilizing unconventional nonlinearity.
• The approach showed scalability to high-dimensional systems while maintaining low latency and low power consumption, outperforming digital implementations.

Conclusions:

• The developed photonic NG-RC on a fiber optic platform effectively addresses the limitations of traditional cavity-based systems.
• This novel approach offers a promising solution for low-latency, low-power, and high-dimensional dynamical systems analysis using optics.
• The study highlights the potential of fiber optics and unconventional nonlinearities for advanced machine learning applications.