An Interpretable Quantum Adjoint Convolutional Layer for Image Classification.

Area of Science:

• Quantum Computing
• Machine Learning
• Artificial Intelligence

Background:

• Quantum machine learning (QML) models often lack interpretability due to closed-box quantum convolutional layers (QCLs).
• Existing QML interpretability methods primarily focus on post-hoc analysis, neglecting intrinsic causes.
• This opacity hinders the reliability and optimal mapping of classical data in QML.

Purpose of the Study:

• To introduce an intrinsically interpretable quantum machine learning scheme.
• To address the opacity and suboptimal data mapping issues in current QML models.
• To enhance the reliability and generalizability of QML through inherent interpretability.

Main Methods:

• Introduction of the quantum adjoint convolution operation (QACO) as an intrinsic interpretability scheme.
• QACO's quantum mapping correlates with image position and pixel values, equivalent to the Frobenius inner product (FIP).
• Integration of the quantum phase estimation (QPE) algorithm into the quantum adjoint convolutional layer (QACL) for parallel FIP computation.

Main Results:

• Achieved higher average test accuracy: 6.3% on Fashion MNIST, 3.4% on MNIST, and 2.9% on DermaMNIST compared to classical and uninterpretable quantum models.
• Demonstrated 73.3% noise-robust accuracy under Gaussian noise.
• Experimental validation on PennyLane and TensorFlow platforms.

Conclusions:

• The proposed QACO and QACL offer superior generalizability and resilience in practical QML applications.
• Intrinsic interpretability enhances the reliability of QML models by clarifying their decision-making processes.
• This approach provides a foundation for more transparent and trustworthy quantum machine learning systems.