Post-variational classical quantum transfer learning for binary classification.

Area of Science:

• Quantum Computing
• Machine Learning
• Artificial Intelligence

Background:

• Variational Quantum Circuits (VQCs) face limitations in hybrid transfer learning, including high training overhead and optimization challenges.
• Existing hybrid classical-quantum models often struggle to achieve optimal performance and efficiency.

Purpose of the Study:

• To introduce and evaluate Post Variational Classical Quantum Transfer Learning (PVCQTL) strategies for overcoming VQC limitations.
• To enhance the performance and efficiency of hybrid classical-quantum transfer learning models.

Main Methods:

• Developed PVCQTL with three designs: modified observable construction, a hybrid approach, and a variational-post-variational combination.
• Evaluated PVCQTL on pre-trained models (VGG19, ResNet50, ResNet18, MobileNet) using 4 and 8 qubits.
• Compared PVCQTL against classical and quantum baselines, including MLP, ResNet50, HQCNN, and CQTL.

Main Results:

• PVCQTL consistently outperformed classical and quantum baselines in accuracy across various datasets.
• The modified observable variant achieved 85% accuracy on the Deepfake dataset with reduced computational cost.
• PVCQTL demonstrated improved accuracy on three additional binary classification datasets, indicating generalizability.

Conclusions:

• PVCQTL offers a significant improvement over traditional hybrid classical-quantum transfer learning approaches.
• The proposed strategies effectively mitigate training overhead and optimization issues in VQCs.
• PVCQTL shows promise for robust and accurate performance in diverse machine learning tasks.