A new publication from Opto-Electronic Sciences; DOI 10.29026/oea.2025.240135, discusses how quasi-convolution coding empowers a streamlined reservoir computer.

As the pace of digitalization accelerates, traditional software-based artificial neural networks are increasingly approaching the physical limits, as dictated by Moore's Law, struggling to keep up with the exponential growth in data processing and computational demands. Silicon-based chip networks are facing urgent challenges, especially processing speed and energy efficiency. In contrast, photonic computing offers promising opportunities to meet the ever-boost demand for computational power, benefitting from its superior processing speed, low energy consumption, and highly efficient parallel processing capabilities. Among numerous photonic neural networks, photonic reservoir computing (RC) stands out due to its excellent physical compatibility and simplified training processes. Particularly noteworthy is the time-delay RC (TDRC), which constructs a complex system with high-dimensional mapping capabilities, based on a single nonlinear node combined with the feedback loop.

Compared to traditional spatial architectures, TDRC introduces a novel computational paradigm that greatly simplifies the network design. By incorporating the feedback loop, this structure significantly enhances the memory capabilities of the system, which are crucial for processing time-dependent tasks. However, this approach also brings challenges in terms of the hardware flexibility, computational speed, and accuracy. Extending the delay line to achieve the better performance in the TDRC will also result in a larger footprint and reduced information-processing rates. To overcome the close reliance on the feedback loop, researchers are gradually motivated to explore feedback-free reservoir structures. These approaches strive to compensate for the required memory through advanced pre/post-processing algorithms or by virtue of the intrinsic properties of physical devices. While these methods have made some progress in simplifying network architectures, they may still have their own deficiencies in terms of biological interpretability and coding efficiency. Additionally, the dependence on specific devices also limits the adaptability of the entire network. Therefore, designing reasonable memory mechanisms along with flexible and efficient coding strategies will be crucial to advancing highly integrated photonic computing systems.

The authors of this article propose an efficient encoding method termed quasi-convolution coding (QC), based on the operational criterion of convolutional encoding. By leveraging the dual polarization modes of the commercial vertical-cavity surface-emitting laser (VCSEL), they successfully simulated the memory mechanisms of the human brain and constructed an easily integrated optical platform for artificial neural networks. The researchers first developed a mathematical model to validate the critical role of encoding signals in enhancing the memory capacity of the feedback-free RC. Through the well-designed QC, they showed that the feedback loop in the traditional TDRC can be effectively compensated, substantially reducing the network complexity. This approach offers a feasible strategy for constructing integrated deep RC systems, paving the way for the advancement of the large-scale, high-throughput RC while maximizing the inherent simplicity of the RC architecture.

Underpinned by this foundational research, the team subtly harnessed the crosstalk effect between the dual polarization modes of the VCSEL to simulate the complex interactions between received and historical signals within the human brain. This innovative approach markedly enhances the memory capacity of the streamlined, feedback-free model. The proposed QRC exhibits exceptional superiority, excelling both in the precise classification of discrete tasks and in time series predictions that demand memory capabilities. Furthermore, the dual-channel architecture of the VCSEL accelerates the processing speed, allowing the simultaneous operation of different data streams in diverse polarization modes, thereby effectively reducing the processing latency by half.

Analogous to the convolutional coding, the kernel size and step coefficient are pivotal in the design of the QC process. Through the meticulous selection of the parameters, the critical temporal features within the input signals will be captured, maintaining the necessary correlations between adjacent sampling periods. Additionally, the researchers investigated the introduction of activation functions during the training process of the output layer. This strategy expands the number of virtual nodes, enriching neuronal representation without increasing the network complexity or sacrificing the processing speed. This work also conducted a comprehensive comparison of multiple neural network models, distinctly emphasizing the effect of the QC and underscoring the vast potential of the QRC in advancing the development of integrated photonic neuromorphic computing.

Keywords: photonic reservoir computing / machine learning / vertical-cavity surface-emitting laser / quasi-convolution coding / augmented memory capabilities

# # # # # #
Prof. Xiaofeng Li is now the vice president of Soochow University as well as the director of key laboratories from Jiangsu Province and the Ministry of Education of China. His research interests include nanophotovoltaics, surface-plasmon physics and devices, hot-electron photodetection, optical sensing, photonic reservoir computing, etc. He has already authored around 250 papers in journals including Nature Energy, Advanced Materials, Light-Science & Applications, Opto-Electronic Advances, Nano Letters, etc. He had been serving as an associate editor for OSA Applied Optics and IEEE Photonics journal for six years and is now the editorial member of PhotoniX and several other journals.
Prof. Xiaofeng Li's homepage: https://web.suda.edu.cn/xfli/

Prof. Nianqiang Li is now the vice dean of the School of Optoelectronic Science and Engineering, Soochow University, as well as the vice director of key laboratories from Jiangsu Province and the Ministry of Education of China. His current research mainly includes the generation and application of laser chaos, photonic spiking neural networks, reservoir computing, microwave signal generation and application, etc. He has already authored around 100 papers in journals including Opto-Electronic Advances, Photonics Research, Neural Networks, Chaos Solitons & Fractals, Journal of Lightwave Technology, Optics Letters, Optics Express, etc.
Prof. Nianqiang Li's homepage: https://web.suda.edu.cn/nli/
# # # # # #
Opto-Electronic Advances (OEA) is a rapidly growing high-impact, open access, peer reviewed monthly SCI journal with an impact factor of 15.3 (Journal Citation Reports for IF2023). Since its launch in March 2018, OEA has been indexed in SCI, EI, DOAJ, Scopus, CA and ICI databases over the time, and expanded its Editorial Board to 34 members from 17 countries.
# # # # # #


More information: http://www.oejournal.org/oea
Editorial Board: http://www.oejournal.org/oea/editorialboard/list
All issues available in the online archive (http://www.oejournal.org/oea/archive).
Submissions to OEA may be made using ScholarOne (https://mc03.manuscriptcentral.com/oea).
ISSN: 2096-4579
CN: 51-1781/TN
Contact Us: oea@ioe.ac.cn
Twitter: @OptoElectronAdv (https://twitter.com/OptoElectronAdv?lang=en)
WeChat: OE_Journal
# # # # # #

Zhou CD, Huang Y, Yang YG et al. Streamlined photonic reservoir computer with augmented memory capabilities. Opto-Electron Adv 8, 240135 (2025). doi: 10.29026/oea.2025.240135