Federated Quantum-Train Long Short-Term Memory for Gravitational Wave Signal

Chen-Yu Liu, Samuel Yen-Chi Chen, Kuan‐Cheng Chen, Wei‐Jia Huang, Yen-Jui ChangINFOCOM WKSHPS 2025

TL;DR

This paper introduces Fed-QT-LSTM, a novel federated learning framework that uses quantum neural networks (QNNs) to generate classical LSTM weights. This approach significantly reduces the number of trainable parameters and eliminates the need for quantum devices during inference, making it practical for real-world applications like gravitational wave signal prediction.

What Problem Does This Paper Address?

The paper tackles the challenge of applying quantum machine learning (QML) to sensitive, distributed data, such as that found in edge computing environments. Traditional QML often requires transferring large datasets to centralized quantum cloud services, which raises privacy concerns and can be computationally inefficient. Furthermore, current QML methods face limitations in encoding large classical datasets into quantum systems due to hardware constraints (e.g., limited qubits and short coherence times). Relying on quantum devices for inference also introduces latency. The authors aim to bridge this gap by developing a federated learning approach that leverages the computational power of QNNs for efficient model training while preserving data locality and enabling classical inference.

Methods

The core innovation is the Federated Quantum-Train Long Short-Term Memory (Fed-QT-LSTM) framework. It combines three key concepts:

  • Federated Learning (FL): This allows multiple client nodes to train a global model collaboratively without sharing their raw data. Each client trains a local model and sends only the model updates (parameters) to a central server for aggregation.
  • Quantum-Train (QT): This framework uses a Quantum Neural Network (QNN) to generate the weights for a classical neural network. Instead of encoding data directly into a QNN, the QNN's output probabilities are mapped to the classical model's parameters. This significantly reduces the number of parameters needed for the classical model.
  • Long Short-Term Memory (LSTM): This is a type of recurrent neural network well-suited for sequential data, like time-series signals.

In Fed-QT-LSTM, each client node uses a QNN to generate the weights for a classical LSTM model. These QNNs are trained locally on the client's data. The generated classical LSTM weights are then aggregated at the central server to form a global LSTM model. A key advantage is that once trained, the Fed-QT-LSTM model functions entirely classically, eliminating the need for quantum hardware during inference. The paper also explores the impact of the "depth" of the QNN (number of QT layers) on performance.

Datasets

The paper uses simulated gravitational wave (GW) signal datasets for training and evaluation. These datasets are generated using inspiral-merger-ringdown models consistent with General Relativity, sourced from the Kaggle page for the Riroriro GW simulation package. The goal is to reconstruct these GW signals.

Key Insights and Ideas

  • Quantum-inspired parameter compression: QT's ability to use QNNs to generate classical model weights offers a powerful way to compress models, drastically reducing the number of trainable parameters compared to traditional methods.
  • Practical quantum advantage for edge AI: By decoupling quantum computation to the training phase and enabling classical inference, Fed-QT-LSTM makes QML practical for edge devices and privacy-sensitive applications.
  • Federated learning enhances scalability and privacy: Integrating QT with FL allows for large-scale distributed learning while preserving data privacy, a critical aspect for many real-world scenarios.
  • QNN depth matters for task complexity: The number of QT layers (QNN depth) is a crucial hyperparameter. For simpler tasks, a shallower QNN suffices, but for more complex tasks (like certain GW signal settings), deeper QNNs are necessary to achieve better performance.

Key Results

  • Fed-QT-LSTM consistently achieved superior performance compared to traditional LSTM and QLSTM models on simulated GW datasets.
  • It demonstrated lower training and testing losses across various simulated GW settings.
  • The framework achieved this with a significant reduction in trainable parameters: Fed-QT-LSTM uses only 123 parameters to generate a classical LSTM model with 1,781 parameters (for L=10 QT layers), representing a ~93% reduction compared to a standard federated LSTM.
  • While QLSTM showed potential, Fed-QT-LSTM's performance was comparable to or better than larger classical LSTM models, while being significantly more parameter-efficient.

Limitations and Future Work

  • QNN depth optimization: While the paper shows the impact of QT layers, finding the optimal depth for different tasks requires careful tuning.
  • Generalizability: The current evaluation is limited to simulated GW data.
  • Future Work: The authors plan to extend Fed-QT-LSTM to a broader range of datasets and application domains, including complex real-world scenarios like weather prediction and large-scale natural simulations. They also aim to explore its application in other complex time-series prediction and signal reconstruction tasks.

Summary updated

Abstract

We present Federated QT-LSTM, a novel framework that combines the Quantum-Train (QT) methodology with Long Short-Term Memory (LSTM) networks in a federated learning setup. By leveraging quantum neural networks (QNNs) to generate classical LSTM model parameters during training, the framework effectively addresses challenges in model compression, scalability, and computational efficiency. Importantly, Federated QT-LSTM eliminates the reliance on quantum devices during inference, making it practical for real-world applications. Experiments on simulated gravitational wave (GW) signal datasets demonstrate the framework's superior performance compared to baseline models, including LSTM and QLSTM, achieving lower training and testing losses while significantly reducing the number of trainable parameters. The results also reveal that deeper QT layers enhance model expressiveness for complex tasks, highlighting the adaptability of the framework. Federated QT-LSTM provides a scalable and efficient solution for privacy-preserving distributed learning, showcasing the potential of quantum-inspired techniques in advancing time-series prediction and signal reconstruction tasks.
BibTeX
@inproceedings{liu2025federated,
  author = {Chen-Yu Liu and Samuel Yen-Chi Chen and Kuan‐Cheng Chen and Wei‐Jia Huang and Yen-Jui Chang},
  title = {{Federated Quantum-Train Long Short-Term Memory for Gravitational Wave Signal}},
  booktitle = {INFOCOM WKSHPS},
  year = {2025},
}
Figure