Scholarlink

<p>We propose Quantum-Train (QT), a hybrid quantum-classical framework for training neural networks that reduces trainable parameter complexity while maintaining competitive performance. QT addresses/avoids three critical challenges in quantum machine learning (QML): (1) the cost of data encoding into quantum circuits, (2) the large parameter footprint of classical models, and (3) the quantum resource demands during inference. The method leverages a parameterized quantum neural network (QNN) to generate classical model weights via a learnable mapping model, achieving compression from <i>M</i> parameters to <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(O(\text {polylog}(M))\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>O</mi> <mo stretchy="false">(</mo> <mtext>polylog</mtext> <mo stretchy="false">(</mo> <mi>M</mi> <mo stretchy="false">)</mo> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> under polynomial-depth QNNs. We provide a theoretical approximation bound quantifying QT’s expressivity and empirically validate the framework on classification tasks, showing that QT achieves competitive accuracy using significantly fewer training parameters. Beyond MLPs, we introduce tensor network-based mappings for enhanced parameter efficiency and extend QT to generate LoRA modules (QT-LoRA), demonstrating improved performance over standard LoRA under low-resource constraints. Furthermore, QT models retain predictive accuracy under hardware noise and finite-shot settings, with gradient analyses showing robustness against barren plateau effects. These findings illustrate QT’s potential as a flexible, scalable, and hardware-friendly paradigm for integrating quantum computation into modern machine learning workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quantum-Train: rethinking hybrid quantum-classical machine learning in the model compression perspective

  • 期刊论文

  • 发表日期：

    2025年8月5日

查看全文
  • Chen-Yu Liu,
  • En-Jui Kuo,
  • Chu-Hsuan Abraham Lin,
  • Jason Gemsun Young,
  • Yeong-Jar Chang,
  • Min-Hsiu Hsieh,
  • Hsi-Sheng Goan

摘要

We propose Quantum-Train (QT), a hybrid quantum-classical framework for training neural networks that reduces trainable parameter complexity while maintaining competitive performance. QT addresses/avoids three critical challenges in quantum machine learning (QML): (1) the cost of data encoding into quantum circuits, (2) the large parameter footprint of classical models, and (3) the quantum resource demands during inference. The method leverages a parameterized quantum neural network (QNN) to generate classical model weights via a learnable mapping model, achieving compression from M parameters to \(O(\text {polylog}(M))\) O ( polylog ( M ) ) under polynomial-depth QNNs. We provide a theoretical approximation bound quantifying QT’s expressivity and empirically validate the framework on classification tasks, showing that QT achieves competitive accuracy using significantly fewer training parameters. Beyond MLPs, we introduce tensor network-based mappings for enhanced parameter efficiency and extend QT to generate LoRA modules (QT-LoRA), demonstrating improved performance over standard LoRA under low-resource constraints. Furthermore, QT models retain predictive accuracy under hardware noise and finite-shot settings, with gradient analyses showing robustness against barren plateau effects. These findings illustrate QT’s potential as a flexible, scalable, and hardware-friendly paradigm for integrating quantum computation into modern machine learning workflows.