Quantum-Train with Tensor Network Mapping Model and Distributed Circuit Ansatz

Chen-Yu Liu, Chu-Hsuan Abraham Lin, Kuan-Cheng ChenICASSP 2025 · DOI

TL;DR

This paper introduces a novel Quantum-Train (QT) framework that enhances scalability and efficiency in hybrid quantum-classical machine learning. By replacing the traditional multi-layer perceptron (MLP) with a tensor network (TN) for mapping quantum measurements to classical model weights and employing a distributed quantum circuit ansatz, the authors achieve competitive performance with significantly fewer parameters. This approach makes QT models more practical for near-term quantum hardware.

What Problem Does This Paper Address?

Hybrid quantum-classical machine learning (QML) models hold great promise, but face significant challenges. One major hurdle is efficiently mapping the results of quantum computations (specifically, measurement probabilities from quantum circuits) to the weights of classical machine learning models. Traditional Quantum-Train (QT) frameworks use Multi-Layer Perceptrons (MLPs) for this mapping. However, MLPs struggle with scalability and interpretability when dealing with the high-dimensional data generated by quantum circuits. Furthermore, the qubit requirements for QML models can become prohibitively large, limiting their application on current quantum hardware and increasing simulation costs. This paper aims to address these limitations by proposing a more efficient and scalable mapping mechanism and a distributed approach to quantum circuit design.

Methods

The authors propose two key innovations to the Quantum-Train framework:

  • Tensor Network (TN) Mapping Model: Instead of an MLP, they use a Tensor Network, specifically a Matrix Product State (MPS) decomposition, to map quantum measurement probabilities to classical model weights. TNs are known for their ability to efficiently represent high-dimensional data and complex correlations, making them a suitable alternative for capturing the intricate relationships between quantum states and classical weights. This approach aims to improve expressive power while reducing the number of tunable parameters.
  • Distributed Circuit Ansatz: To tackle the issue of large qubit requirements, they introduce a distributed circuit ansatz. This approach decomposes a large quantum circuit into smaller, independent sub-circuits that can be processed in parallel. The results from these sub-circuits are then combined via a tensor product. This not only reduces the qubit requirements for individual components but also allows for more efficient simulation and potentially deployment on distributed quantum hardware.

Datasets

The paper evaluates its proposed framework on the CIFAR-10 image classification dataset. This is a standard benchmark dataset in machine learning, commonly used for evaluating image classification models. The authors compare their results against a baseline established in a previous QT paper [12] which also used CIFAR-10.

Key Insights and Ideas

  • Tensor Networks Offer Superior Parameter Efficiency: The TN mapping model significantly reduces the number of parameters needed to achieve comparable or better accuracy compared to the MLP-based QT approach. This is crucial for model compression and efficient deployment.
  • Distributed Circuits Alleviate Qubit Constraints: Decomposing the QNN into smaller, distributed circuits effectively bypasses the limitations imposed by the number of available qubits on current quantum hardware, making larger QML models more feasible.
  • Classical vs. Quantum Parameter Impact: The study highlights a trade-off: increasing quantum parameters (e.g., QNN layers) can lead to greater accuracy gains per parameter, but at the cost of increased coherence time requirements and simulation complexity. Conversely, increasing classical parameters (e.g., TN bond dimension) offers a more scalable and practical path for NISQ-era devices, even if the accuracy gains per parameter are less pronounced.
  • QT Framework's Adaptability: The QT framework, with these enhancements, proves to be a flexible and adaptable architecture that can leverage advancements in both quantum and classical computing techniques.

Key Results

The paper demonstrates that the proposed Quantum-Train with Tensor Network (QT-TN) mapping model and distributed circuit ansatz achieves competitive performance on the CIFAR-10 dataset. Specifically:

  • At a testing accuracy of approximately 60%, the QT-TN model requires slightly over 10,000 parameters, whereas the baseline QT-MLP model requires approximately 23,000 parameters. This represents a significant reduction in parameter count.
  • The QT-TN approach offers superior parameter efficiency compared to the QT-MLP baseline, requiring fewer parameters to reach comparable accuracy levels.
  • The authors note that the QT-TN results outperform those of the MLP mapping model, indicating improved generalization and efficiency.

Limitations and Future Work

The authors acknowledge several limitations and outline promising future directions:

  • Scalability of TNs: While TNs offer parameter efficiency, increasing the bond dimension (a classical parameter) leads to a steeper increase in parameter count compared to the QT-MLP's increase in quantum parameters (QNN layers). However, they argue this is a more cost-effective approach in the current NISQ era.
  • Future Work:
    • Exploring more advanced tensor decomposition techniques like Tree Tensor Networks (TTN) and Projected Entangled Pair States (PEPS) for even better compression and accuracy.
  • Investigating alternative tensor decompositions to further reduce qubit counts without compromising model performance.
  • Optimizing the distributed circuit ansatz for real quantum hardware.
  • Testing the framework on larger datasets and more complex tasks, such as reinforcement learning and natural language processing.
  • Integrating error mitigation techniques and optimizing the TN mapping model for practical, real-world quantum hardware applications.

Summary updated

BibTeX
@inproceedings{liu2025quantum,
  author = {Chen-Yu Liu and Chu-Hsuan Abraham Lin and Kuan-Cheng Chen},
  title = {{Quantum-Train with Tensor Network Mapping Model and Distributed Circuit Ansatz}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10889597},
}
Figure