Optical next generation reservoir computing

Area of Science:

• Physics
• Computer Science
• Nonlinear Dynamics

Background:

• Artificial neural networks with internal dynamics excel at information processing.
• Reservoir computing (RC) offers rich computational expressivity and physical implementation advantages.
• Next-generation reservoir computing (NGRC) enhances expressivity but faces challenges in physical realization.

Purpose of the Study:

• To demonstrate a physically open optical NGRC system for information processing.
• To leverage light scattering in disordered media for NGRC computations.
• To address the limitations of conventional optical RC and digital NGRC.

Main Methods:

• Utilizing light scattering through disordered media as the optical reservoir.
• Directly driving the optical reservoir with time-delayed inputs.
• Implicitly generating polynomial features of delayed inputs for functionality, mirroring digital NGRC.

Main Results:

• Successfully predicted short-term dynamics of Lorenz63 and Kuramoto-Sivashinsky chaotic time series.
• Replicated long-term ergodic properties of chaotic systems.
• Demonstrated superior performance over conventional optical RC in training length, hyperparameters, and forecasting accuracy.

Conclusions:

• The optical NGRC framework offers a viable physical realization for NGRC.
• This approach inspires NGRC in other physical systems and new applications.
• Potential for developing deep and parallel architectures in physical computing systems.