Quantum Conflict Measurement in Decision Fusion for Out-of-Distribution Detection.

Area of Science:

• Quantum Information Science
• Artificial Intelligence
• Decision Theory

Background:

• Quantum Dempster-Shafer theory (QDST) uses quantum mass functions (QMFs) for uncertain information but struggles with conflicting QMFs.
• Existing methods for QMF fusion and decision-making often rely on predicted labels, limiting their ability to handle unseen data or out-of-distribution samples.

Purpose of the Study:

• To develop a novel quantum conflict indicator (QCI) to measure conflicts between QMFs.
• To introduce QCI-Fusion for fusing highly conflicting QMFs.
• To propose QCI-Decision, an unsupervised decision architecture for out-of-distribution (OOD) detection and in-distribution (ID) classification.

Main Methods:

• Developed a Quantum Conflict Indicator (QCI) with ideal conflict measurement properties.
• Introduced QCI-Fusion for fusing conflicting QMFs.
• Designed QCI-Decision for unsupervised OOD detection and ID classification, rejecting OOD samples.

Main Results:

• QCI is the first metric with desirable properties for conflict measurement.
• QCI-Decision shows minimal deviation (≤1.49%) from original model predictions.
• QCI-Decision improves Area Under the ROC Curve (AUC) by up to 0.6% and reduces False Positive Rate (FPR) by up to 1.63% compared to state-of-the-art OOD methods.
• QCI-Decision offers threefold faster fusion speed than QCI-Fusion with comparable performance.

Conclusions:

• The proposed QCI and QCI-Decision provide effective solutions for managing conflicting QMFs and enabling robust open-world quantum information decision-making.
• QCI-Decision demonstrates superior performance in OOD detection and classification accuracy while significantly improving efficiency.