Optoelectronic recurrent neural network using optical-electrical-optical converters with RC delay

Masaya Arahata, Shota Kita, Kazuo Aoyama, Akihiko Shinya, Hiroshi Sawada, Masaya Notomi

Abstract

Optical neural networks (ONNs) have been attracting intense attention owing to their low latency and low-power consumption. Among the ONNs, optical recurrent neural networks (RNNs) enable low-power and high-speed time-series data processing using a compact loop structure. The loop losses need to be efficiently compensated so that the time-series information is maintained in the RNN operation. For this purpose, we focus on the optoelectronic RNN (OE-RNN) with optical-electrical-optical (OEO) converters to compensate for the loop losses. However, the effect of resistive-capacitive (RC) delay of OEO converters on the RNN performance is unclear. Here, we study in simulation an OE-RNN equipped with OEO converters with RC delay. We confirm that our modeled OE-RNN achieves high training accuracy for time-series data classification even when RC delay is comparably large to the time interval of time-series data. Our analyses reveal that the accumulation of time-series data by RC delay does not degrade the RNN performance but rather can compensate for the degraded RNN performance due to loop losses. From the theoretical analysis referring to the gradient explosion and vanishing problems, we find the region related to loss and RC delay where high training accuracy can be achieved. In simulation, we confirm this compensation effect in the large OE-RNN circuit up to 32x32 scale. Our proposed scheme opens a new way of time-series data processing by utilizing RC delay for optical computing and optical communication.

Introduction

The demand for high-speed, energy-efficient processing of time-series data has led to a surge of interest in optical computing. Optical Recurrent Neural Networks (RNNs) are particularly promising because their feedback loops allow for a compact hardware footprint while processing sequential information. However, a persistent challenge in these systems is "loop loss"—the gradual attenuation of optical signals as they circulate through the feedback path.

To mitigate these losses, researchers often integrate Optical-Electrical-Optical (OEO) converters into the loop. These components provide the necessary gain and non-linear activation functions. A side effect of OEO conversion is the inherent resistive-capacitive (RC) delay, caused by the physical properties of photodetectors and modulators. Traditionally, this delay has been viewed as a performance bottleneck that smears signals and limits bandwidth. This study challenges that view, demonstrating that RC delay can serve as a form of temporal memory that actually improves network stability and performance.

System Architecture and Modeling

The proposed architecture is an Elman-type optoelectronic RNN (OE-RNN) designed for time-series classification. The architecture consists of an input unit, an optical add unit, an OEO conversion unit, and a hidden unit implemented via a "Clements circuit." This circuit uses a mesh of Mach-Zehnder interferometers (MZIs) to perform unitary matrix multiplications, ensuring that the norm of the hidden state is preserved and gradients remain stable during training.

The OEO unit is the core of the proposed model. It is characterized by two main parameters: loop gain ($\beta$) and RC delay ($\alpha$). The RC delay is modeled as a linear time-invariant system where the output signal at any given time step is a weighted sum of the current input and an exponentially decaying "memory" of previous inputs.

Schematic of the OE-RNN architecture and the OEO unit with RC delay modeling.

Theoretical Framework: The Stability Region

A critical contribution of this work is the derivation of a stability condition for the OE-RNN. By applying the chain rule to the network's loss function, the authors identified that the effective transition matrix of the network is $S \equiv \beta W + \alpha I$, where $I$ is the identity matrix. To avoid the common RNN pitfalls of gradient vanishing or explosion, the eigenvalues of $S$ should ideally have an absolute value near 1.

Mathematically, the RC delay ($\alpha$) adds to the diagonal of the weight matrix. This shift can "rescue" a network suffering from high loop loss (where $\beta < 1$) by pushing the eigenvalues back into the stable zone. The authors defined a specific design rule: high training accuracy is achievable when loop gain and delay satisfy the inequality: $- \alpha + 1 \leq \beta \leq \alpha + 1$.

Experimental Results and Scalability

The model was benchmarked using MNIST (handwritten digits) and Fashion-MNIST datasets. To adapt these 2D images for an RNN, the team applied a Fast Fourier Transform (FFT) and used the resulting spectral components as sequential inputs.

The simulations confirmed a powerful "compensation effect." In a 4x4 OE-RNN, performance dropped significantly when loop gain was low and delay was absent. However, by increasing the RC delay, the network's accuracy was restored to levels comparable to a lossless system. This effect was not limited to small-scale models; the researchers successfully scaled the simulation to 8x8, 16x16, and 32x32 circuits.

Scalability results showing consistent performance across different circuit sizes and datasets, with red lines indicating the theoretical stability zone.

As shown in the scalability results, the "compensation effect" remains robust across different tasks and circuit sizes. The network achieved high accuracy even when the RC delay was large relative to the input time interval, proving that high-speed operation is possible even with "imperfect" hardware components.

Conclusion and Future Outlook

This research demonstrates a paradigm shift in how we view physical limitations in optical hardware. By treating RC delay as a computational resource—a form of hardware-native temporal memory—researchers can build more robust and efficient optical RNNs. While these results are currently based on numerical simulations, they provide a clear roadmap for physical implementation on silicon photonic chips. Future work will explore the network's performance on more complex, natural time-series data and investigate the impact of real-world noise and pulse-shaping effects.

Summary updated

Abstract

Optical neural network (ONN) has been attracting intense attention owing to their low latency and low-power consumption. Among the ONNs, optical recurrent neural network (RNN) enables low-power and high-speed time-series data processing using a compact loop structure. The loop losses need to be efficiently compensated so that the time-series information is maintained in the RNN operation. For this purpose, we focus on the optoelectronic RNN (OE-RNN) with optical-electrical-optical (OEO) converters to compensate for the loop losses. However, the effect of resistive-capacitive (RC) delay of OEO converters on the RNN performance is unclear. Here, we study in simulation an OE-RNN equipped with OEO converters with RC delay. We confirm that our modeled OE-RNN achieves the high training accuracy of time-series data classification even when RC delay is comparably large to the time interval of time-series data. Our analyses reveal that the accumulation of time-series data by RC delay does not degrade the RNN performance but rather can compensate for the degraded RNN performance due to loop losses. From the theoretical analysis referring to the gradient explosion and vanishing problems, we find the region related to loss and RC delay where the high training accuracy can be achieved. In simulation, we confirm this compensation effect in the large OE-RNN circuit up to 32$\times$32 scale. Our proposed scheme opens a new way of time-series data processing by utilizing RC delay for the optical computing and optical communication.
BibTeX
@article{arahata2025optoelectronic,
  author = {Masaya Arahata and Shota Kita and Kazuo Aoyama and Akihiko Shinya and Hiroshi Sawada and Masaya Notomi},
  title = {{Optoelectronic recurrent neural network using optical-electrical-optical converters with RC delay}},
  journal = {Journal of Lightwave Technology},
  year = {2025},
  doi = {10.1109/jlt.2025.3629691},
}
Figure