Reservoir Computing via Quantum Recurrent Neural Networks

Samuel Yen-Chi Chen, Daniel Fry, Amol Deshmukh, Vladimir Rastunkov, Charlee StefanskiarXiv 2022 · DOI · arXiv

Abstract

Sequential modeling is a cornerstone of modern data science, yet training Quantum Recurrent Neural Networks (QRNNs) remains computationally prohibitive on current hardware. This paper presents a Reservoir Computing (RC) approach to QRNNs, where quantum parameters are fixed as a dynamical reservoir and only a final classical layer is trained. Our results demonstrate that this framework matches the performance of fully trained models while offering superior training efficiency and robustness to quantum noise.

Introduction

Predicting temporal sequences is critical for applications ranging from financial forecasting to physical system modeling. While Quantum Machine Learning (QML) offers high expressive power, the optimization of Variational Quantum Circuits (VQCs) on Near-term Intermediate-Scale Quantum (NISQ) devices is hampered by the high cost of gradient estimation. Specifically, the "parameter-shift" rule requires multiple circuit evaluations per parameter, making recurrent quantum models difficult to scale. This research investigates whether the Reservoir Computing paradigm can provide the benefits of quantum memory without the associated training overhead.

Methods: The QRNN-RC Framework

The researchers proposed the QRNN-RC (Quantum Recurrent Neural Network - Reservoir Computing) framework, applying it to three standard architectures: the basic QRNN, the Quantum Long Short-Term Memory (QLSTM), and the Quantum Gated Recurrent Unit (QGRU).

The Reservoir Computing Paradigm

In a traditional reservoir computing setup, a fixed, non-linear dynamical system (the reservoir) maps inputs into a high-dimensional feature space. In this work, the reservoir is composed of randomly initialized and frozen VQCs.

Schematic of the Reservoir Computing framework where an input sequence is processed by a fixed dynamical reservoir, followed by a trainable classical output layer.

By replacing classical neurons with VQCs, the model leverages quantum states to capture complex temporal dependencies. For instance, the QLSTM architecture utilizes multiple VQCs to manage cell states and gates, as shown below.

The architecture of a Quantum Long Short-Term Memory (QLSTM) cell, featuring multiple Variational Quantum Circuits (VQCs) and classical activation functions.

Only the final classical linear layer is optimized using standard techniques, bypassing the need for quantum gradient calculations during training.

Experimental Evaluation

The models were tested on function approximation tasks (Damped Simple Harmonic Motion and Bessel functions) and complex time-series benchmarks (NARMA5 and NARMA10).

Key Findings

  • Efficiency: QRNN-RC models achieved convergence in significantly fewer epochs (15–30) compared to fully optimized counterparts.
  • Noise Resilience: Under simulated NISQ noise (e.g., IBM Peekskill models), the RC approach proved more robust. By treating noise as a fixed part of the reservoir's dynamics rather than a perturbation to be optimized away, the models produced smoother and more reliable predictions.
  • Performance: In NARMA10 tasks, QLSTM-RC and QGRU-RC outperformed classical reservoirs of similar parameter counts, suggesting that quantum reservoirs provide a richer feature space for sequential data.

Conclusion

Reservoir Computing offers a viable "shortcut" for implementing recurrent quantum models on contemporary hardware. By eliminating the need for quantum backpropagation, QRNN-RC reduces computational costs while maintaining high accuracy and improving noise tolerance. Future research will focus on scaling these reservoirs to larger qubit counts and exploring their application in real-world domains such as speech processing and financial fraud detection.

Summary updated

Abstract

Recent developments in quantum computing and machine learning have propelled the interdisciplinary study of quantum machine learning. Sequential modeling is an important task with high scientific and commercial value. Existing VQC or QNN-based methods require significant computational resources to perform the gradient-based optimization of a larger number of quantum circuit parameters. The major drawback is that such quantum gradient calculation requires a large amount of circuit evaluation, posing challenges in current near-term quantum hardware and simulation software. In this work, we approach sequential modeling by applying a reservoir computing (RC) framework to quantum recurrent neural networks (QRNN-RC) that are based on classical RNN, LSTM and GRU. The main idea to this RC approach is that the QRNN with randomly initialized weights is treated as a dynamical system and only the final classical linear layer is trained. Our numerical simulations show that the QRNN-RC can reach results comparable to fully trained QRNN models for several function approximation and time series prediction tasks. Since the QRNN training complexity is significantly reduced, the proposed model trains notably faster. In this work we also compare to corresponding classical RNN-based RC implementations and show that the quantum version learns faster by requiring fewer training epochs in most cases. Our results demonstrate a new possibility to utilize quantum neural network for sequential modeling with greater quantum hardware efficiency, an important design consideration for noisy intermediate-scale quantum (NISQ) computers.
BibTeX
@misc{chen2022reservoir,
  author = {Samuel Yen-Chi Chen and Daniel Fry and Amol Deshmukh and Vladimir Rastunkov and Charlee Stefanski},
  title = {{Reservoir Computing via Quantum Recurrent Neural Networks}},
  year = {2022},
  eprint = {2211.02612},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.02612},
}

Loading related content...

Figure