QSAN: A Near-Term Achievable Quantum Self-Attention Network.

Area of Science:

• Quantum Machine Learning (QML)
• Artificial Intelligence
• Computer Vision

Background:

• Classical self-attention mechanisms (SAM) excel at feature connection but are absent in many quantum machine learning (QML) models.
• This limitation restricts QML's scalability for high-dimensional quantum data.
• Existing QML models face challenges in efficiently processing complex datasets.

Purpose of the Study:

• To introduce a quantum self-attention mechanism (QSAM) to enhance QML models.
• To develop a quantum self-attention network (QSAN) for efficient processing of quantum data.
• To improve the learning capability and efficiency of QML models for large-scale applications.

Main Methods:

• Developed a quantum self-attention mechanism (QSAM) using a quantum logic similarity (QLS)-based quantum bit self-attention score matrix (QBSASM).
• Designed a quantum self-attention network (QSAN) with optimized quantum circuits for measurement time compression.
• Utilized a prototype of quantum coordinates to define mathematical relationships for programming.

Main Results:

• QSAN demonstrated significantly faster convergence rates (1.7x and 2.3x) compared to hardware-efficient and QAOA ansatz on MNIST binary classification.
• Achieved 100% prediction accuracy on MNIST, indicating superior learning capability.
• Attained high prediction accuracy on CIFAR-10 classification with a smaller scale than classical models.

Conclusions:

• QSAN effectively enhances QML model efficiency and learning capability.
• The developed QSAM and QSAN provide a foundation for future quantum machine learning on massive datasets.
• This work promotes advancements in quantum computer vision and other QML-driven fields.