03

2025

-

09

Sylvain Gigan, Qiang Liu, and others Light | Optical "Next-Generation Reservoir Computing Neural Network"

Author:


In the era of artificial intelligence, recurrent neural networks act like intelligent agents with "dynamic memory," demonstrating powerful information processing capabilities. Among them, reservoir computing (RC) networks, with their rich model expressiveness and unique compatibility with simulated physical systems, have become one of the important architectures in the field of optical computing. Recently, the new generation algorithm "Next Generation Reservoir Computing" (NGRC) has further enhanced the ability to process temporal data but has created a "compatibility gap" with physical systems.

Recently, the École Normale Supérieure in Paris collaborated with teams from Tsinghua University and the University of Hong Kong to propose an experimental scheme to implement the NGRC algorithm in optical multiple scattering systems. While reducing the reservoir size, training data volume, and hyperparameters, it achieved a leap in the prediction capability of chaotic systems, bridging a new connection between optical computing and nonlinear dynamical systems. This research was published under the title " Optical Next Generation Reservoir Computing " in Light: Science & Applications with Hao Wang, a joint PhD student of Tsinghua University and École Normale Supérieure, as the first author; Assistant Professor Jianqi Hu as co-first author and corresponding author; and Professors Qiang Liu and Sylvain Gigan as co-corresponding authors.

Opportunities and Challenges of NGRC

In a narrow sense, reservoir computing networks are a special type of recurrent neural network that can fully utilize the intrinsic dynamic evolution characteristics of the network to achieve functions such as memory, prediction, and classification [1] RC, as a "physics-friendly" model, has been widely deployed on optical, electrical, mechanical, and other platforms due to its simple training and compatibility with complex nonlinear physical computing systems. Recently, an algorithm called "Next Generation Reservoir Computing" (NGRC) integrates nonlinear vector autoregression with core neural network ideas, enabling insight into system evolution rules with only a small amount of historical data, thus pushing temporal prediction capabilities to new heights [2] However, this algorithm upgrade has led to a sharp decline in compatibility with physical systems, making scalable large-scale optical NGRC implementation a challenge.

Algorithmic Rebirth in Multiple Scattering Systems

The core of NGRC lies in synthesizing reservoir feature vectors from raw data, i.e., directly and solely using feature sets of multiple adjacent (or equally spaced) time points under a certain function basis (such as monomial sets of polynomial basis functions) [3] The research team took a novel approach by reconstructing the algorithm implementation method based on the above algorithm characteristics and the physical nature of multiple scattering optical computing systems. They encoded temporal data at different moments into the phase distribution of the input light field and used the physical properties of linear scattering media to autonomously generate polynomial data features from the output light field intensity signals—this perfectly maps to the mathematical core of NGRC.

Experiments show that this optical NGRC system can not only accurately predict the short-term evolution of low-dimensional chaotic systems (Lorenz63) and large-scale spatiotemporal chaotic sequences (Kuramoto-Sivashinsky) but also fully reproduce their long-term statistical properties. Additionally, the team explored an optical NGRC observer, opening new avenues for real-time monitoring of complex dynamical systems. Compared to traditional optical reservoir computing networks [4] this optical system not only inherits the core advantages of the NGRC algorithm—smaller training sets, fewer hyperparameters, better prediction performance, and improved interpretability—but also embraces the potential scalability advantages of optical computing.

This research follows the paradigm of "co-design of optical physics and software algorithms," adding a new member to the optical reservoir computing network architecture [5,6] and is expected to bring new research opportunities to physical reservoir systems beyond optics. In the future, the research team will further explore how to enhance its interpretability and develop deeply parallel architectures to unleash the natural advantages of optical computing.