Title: Enhancing AI Interpretability via Quantum Annealing for Feature Selection

Abstract: In high-stakes applications, the deployment of deep learning models necessitates a rigorous understanding of their predictive mechanisms, as errors can lead to severe repercussions. Grasping the rationale behind model outputs is essential for verifying that algorithms are learning appropriate patterns, identifying data biases, refining architectural designs, and fostering trust in the system. This study introduces a novel technique for interpreting Convolutional Neural Networks (CNNs) within the context of image classification. The proposed method functions by isolating the most salient feature maps that drive individual predictions. To address the inherent complexity of this selection process, we formulate the issue as a constrained optimization problem suitable for quantum annealing. Our evaluation benchmarks this approach against leading explainable AI methods, namely GradCAM and GradCAM++. The results indicate superior class disentanglement, characterized by sharper decision boundaries and clearer reasoning processes. These findings suggest that our method elevates the quality of explanations, thereby facilitating a better understanding of the specific features influencing particular predictions. Furthermore, we investigate the computational dynamics of the quantum annealing algorithm, focusing on the system’s minimum energy gap during execution and the likelihood of locating the optimal solution. These analytical insights offer a theoretical foundation for the practical efficacy of the proposed method.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC