In Situ Optimization of an Optoelectronic Reservoir Computer with Digital Delayed Feedback
2025
Read In Situ Optimization of an Optoelectronic Reservoir Computer with Digital Delayed Feedback (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-11126.txt.
Introduction
Fyodor Morozko and co-authors report, in a 2025 paper, a way to tune a light-based reservoir computer directly on the hardware, without a software twin. Reservoir systems keep their inner wiring fixed and random, which suits physical builds but makes tuning awkward. Their device uses an optoelectronic loop with digital delayed feedback.
Five settings were tuned together on the device itself. The tuned system scored well on three standard tests: waveform sorting, time series prediction, and speech recognition. It beat simulation-tuned settings on two of the three. The authors call in-place tuning a step toward practical physical computing that skips costly modeling.
Catalog record
| YEAR | 2025 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-11126
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
In Situ Optimization of an Optoelectronic Reservoir Computer with Digital Delayed Feedback Abstract Abstract Reservoir computing (RC) is a powerful computational framework that addresses the need for efficient, low-power, and high-speed processing of time-dependent data. While RC has demonstrated strong capabilities in signal processing and pattern recognition, its practical deployment in physical hardware is hindered by a critical challenge: the lack of efficient, scalable parameter optimization methods for real-world implementations. Traditionally, RC optimization has relied on software-based modeling, which limits the adaptability and efficiency of hardware-based systems, particularly in high-speed and energy-efficient computing applications. Herein, an in situ optimization approach was employed to demonstrate an optoelectronic delay-based RC system with digital delayed feedback, enabling direct, real-time tuning of system parameters without reliance on external computational resources. By simultaneously optimizing five parameters, normalized mean squared error (NMSE) of 0.028, 0.561, and 0.271 is achieved in three benchmark tasks: waveform classification, time series prediction, and speech recognition outperforming simulation-based optimization with NMSE 0.054, 0.543, and 0.329, respectively, in the two of the three tasks. This method enhances the feasibility of physical reservoir computing by bridging the gap between theoretical models and practical hardware implementation. keywords reservoir computing, neuromorphic computing, physical computing, optoelectronic oscillator, in situ optimization Fyodor Morozko 1 Shadad Watad 1 Amir Naser 1 Andrey Novitsky Alina Karabchevsky 1,3 * F. Morozko, S. Watad, A. Naser, A. Novitsky, Karabchevsky 1 School of Electrical and Computer Engineering Ben-Gurion University of the Negev Beer-Sheva 8410501, Israel E-mail: [email protected] A. Novitsky 2 Belarusian State University Minsk, 220030, Belarus A. Karabchevsky 3 Department of Physics Lancaster University LA1 4YB, United Kingdom 1 Introduction Reservoir computing (RC) has emerged as a powerful neuromorphic computing framework for processing time-dependent data, offering advantages in speed, energy efficiency, and ease of training compared to conventional machine learning approaches. By leveraging a fixed, high-dimensional nonlinear transformation of input signals, RC circumvents the need for extensive weight optimization, making it particularly attractive for hardware implementations in optical, electronic, and mechanical systems. However, despite its potential, a key limitation remains: the lack of efficient, scalable optimization strategies for tuning system parameters in physical RC implementations. Traditional approaches rely heavily on software-based modeling, which is impractical for real-time, adaptive hardware systems. Addressing this challenge is crucial to unlocking the full potential of RC for applications in ultrafast signal processing, real-time decision-making, and low-power embedded AI. 10 Figure 1: a Generic reservoir computing principle. The depicted layout consists of distinct components: an input layer (bronze spheres) responsible for receiving external data, a reservoir (ruby spheres) featuring randomized fixed connections, and a linear readout layer (green spheres). b Delay reservoir computing. f(x) is the activation function performing the nonlinear transformation exhibited by the element, and h(t) is the impulse response. W , W^{I} and W^{R} are, respectively, reservoir, input, and readout connectivity matrices. Traditional RC optimization relies on software simulations, which may not accurately capture the complexities of physical systems, such as noise, non-idealities, and hardware variability [ 15 ] [ 18 , 19 , 20 , 21 , 22 , 23 , 24 ] . In situ optimization directly accounts for these real-world factors, leading to more robust and reliable performance. Physical RC systems, especially optoelectronic ones, operate in dynamic environments where temperature fluctuations, component aging, and external disturbances can affect performance. An in situ approach, however, can enable real-time parameter tuning, allowing the system to adapt and maintain optimal functionality without requiring offline recalibration, enhancing the computational efficiency. Software-based optimization often involves time-consuming simulations and iterative tuning, which may not be feasible for high-speed, real-time applications. In situ optimization allows direct tuning within the hardware, reducing computational overhead and enabling faster convergence to optimal operating conditions. By enabling direct, real-time parameter tuning, in situ optimization may transform optoelectronic delay-based RC from a theoretical concept into a practical, high-performance computing tool with broad applications in AI, signal processing, and beyond. Herein, we address the challenge of optimizing an optoelectronic delay-based reservoir computing (RC) system through an in situ approach. We optimize five key hyperparameters simultaneously: the delay-to-clock cycle ratio, input scaling, phase bias, nonlinearity gain, and the regularization parameter used in readout training. The effectiveness of this in situ optimization is demonstrated across three benchmark tasks: classification of sinusoidal versus rectangular waveforms, NARMA10 time-series prediction, and Japanese vowel classification. Our method achieves state-of-the-art normalized mean squared error (NMSE) values of 0.028, 0.561, and 0.271, respectively. Furthermore, we experimentally validate theoretical predictions from Refs. [ 27 , 28 ] which were never realized, and show that resonances between the delay and clock cycle timescales adversely affect RC accuracy, particularly in the NARMA10 task. This is evidenced by significant variations in NMSE, ranging from 0.561 to over 1, depending on the delay-to-clock cycle ratio. These findings provide valuable insights into optimizing the performance of delay-based RC systems for various computational tasks. 2 Results 2.1 Device performance We implemented a delay system in an optoelectronic oscillator, a system consisting of a laser light source, an electro-optic modulator, a photodetector whose output is connected to the modulation port of the modulator forming a closed loop. Sinusoidal transmission characteristic of a Mach-Zehnder electro-optic modulator renders the system nonlinear while introducing a delay into the feedback loop leads to the emergence of multiple-valued stationary states, causing the system to exhibit complex nonlinear behavior [ 3 , 1 ] . Particularly, optoelectronic oscillator, upon the tuning of laser power, photodetector gain, and phase bias of the modulator, can exhibit transitions between stable, multi-stable, periodic, and chaotic dynamics known as Hopf bifurcations [ 5 , 36 , 4 ] . Dynamics of an optoelectronic oscillator can be described by Ikeda model [ 3 , 1 , 36 , 4 ] according to which voltage V at the modulation port of MZM follows the delay-differential equation V(t)+T_{R}\frac{\mathrm{d}V}{\mathrm{d}t}(t)=G^{*}P[V(t-\tau)], (1) where P[V] is the power transmitted by the MZM and received by the photodetector, \tau is the delay time, G^{*} is the voltage gain of the photodetector, and, T_{R} is the response time of the system. The transmission characteristic P[V] of the Mach-Zehnder modulator is given by [ 5 , 36 , 4 ] P[V]=\frac{1}{2}P_{\mathrm{max}}\left(1+M\sin(\pi(V+V_B)/V_\pi+\phi)\right), (2) where P_{\mathrm{max}} is the total optical power in the system, M , V_{\pi} , and \phi are the modulation depth, half-wave voltage, and intrinsic phase of the MZM, respectively, and V_{B} is the bias voltage at the bias port of the MZM. If the response time is much shorter than the delay time T_{R}\ll\tau , the derivative in Eq. ( 1 ) can be neglected, and we can describe the dynamics V(t) by a difference equation [ 5 ] V(t)=G^{*}P[V(t-\tau)], (3) while analyzing V(t) at discrete time steps t=t_{0}+n\tau . Introducing the dimensionless state variable x=V/V_{\pi} , performing time multiplexing according to Eq. ( 11 ), and adding external signal we rewrite Eq. 3 in terms of a generic RC evolution Eq. ( 7 ) as \mathbf{x}(n+1)=\frac{G}{2}\left(1+M\sin(\beta W\vb{x}(n)+\rho W^I\vb{u}(n+1)+\Phi_0)\right), (4) where \Phi_{0}=\phi+\pi V_{B}/V_{\pi} is the phase bias, \beta and \rho are the feedback and input scaling, respectively, and G=G^{*}/V_{\pi}P_{\mathrm{max}} is the net gain. In delay RC the matrix W is generally a circulant matrix whose entries depend on the ratio of the delay time \tau to the clock cycle