Scholarlink

<p>Quantum long short-term memory (QLSTM) networks have gained attention for sequential modeling tasks but suffer from inconsistent implementations, particularly regarding the number of variational quantum circuits (VQCs) used (4, 5, or 6) without a well-defined rationale. This lack of standardization limits their efficiency, especially on near-term quantum devices (NISQ). We introduce the Single Quantum-Output LSTM (SQO-LSTM), an optimized QLSTM architecture that consolidates multiple VQCs into a single quantum module, reducing complexity while maintaining performance. This design minimizes quantum resource usage, making it more suitable for current NISQ devices. SQO-LSTM was evaluated on 20 diverse text classification datasets (binary and multi-class). These datasets span linguistic diversity (English and Spanish), various levels of structural complexity, and different sequence lengths. Our experimental findings demonstrate a clear quantum advantage: SQO-LSTM showing consistent improvements over classical and QLSTM baselines under identical experimental settings, where it surpasses the 4-VQC QLSTM by 10–20% on challenging datasets such as RP and MILK, while significantly reducing training times. Moreover, we explored different classical-quantum layer configurations and ansatz choices, finding that the Bi-Ansatz structure optimally improves accuracy (e.g., RP dataset: 76% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\rightarrow\)</EquationSource> <EquationSource Format="MATHML"><math> <mo stretchy="false">→</mo> </math></EquationSource> </InlineEquation> 86%). These results demonstrate that our architecture is not only simpler and faster but also significantly more efficient and SQO-LSTM as a strong candidate, achieving the best accuracy on nearly balanced datasets, although it remains suboptimal for imbalanced data. By optimizing architecture and reducing complexity, our approach paves the way for practical quantum machine learning applications in natural language processing and sequential modeling tasks.</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

SQO-LSTM: a single quantum-output long short-term memory for classification tasks

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Yousra Bouakba,
  • Hacene Belhadef,
  • Abdelhalim Saadi

摘要

Quantum long short-term memory (QLSTM) networks have gained attention for sequential modeling tasks but suffer from inconsistent implementations, particularly regarding the number of variational quantum circuits (VQCs) used (4, 5, or 6) without a well-defined rationale. This lack of standardization limits their efficiency, especially on near-term quantum devices (NISQ). We introduce the Single Quantum-Output LSTM (SQO-LSTM), an optimized QLSTM architecture that consolidates multiple VQCs into a single quantum module, reducing complexity while maintaining performance. This design minimizes quantum resource usage, making it more suitable for current NISQ devices. SQO-LSTM was evaluated on 20 diverse text classification datasets (binary and multi-class). These datasets span linguistic diversity (English and Spanish), various levels of structural complexity, and different sequence lengths. Our experimental findings demonstrate a clear quantum advantage: SQO-LSTM showing consistent improvements over classical and QLSTM baselines under identical experimental settings, where it surpasses the 4-VQC QLSTM by 10–20% on challenging datasets such as RP and MILK, while significantly reducing training times. Moreover, we explored different classical-quantum layer configurations and ansatz choices, finding that the Bi-Ansatz structure optimally improves accuracy (e.g., RP dataset: 76% \(\rightarrow\) → 86%). These results demonstrate that our architecture is not only simpler and faster but also significantly more efficient and SQO-LSTM as a strong candidate, achieving the best accuracy on nearly balanced datasets, although it remains suboptimal for imbalanced data. By optimizing architecture and reducing complexity, our approach paves the way for practical quantum machine learning applications in natural language processing and sequential modeling tasks.