GUARDIAN: Generating Understanding and Accountability in Robust Detection with Directed Energy Insights and Analysis of Neural Collapse

The GUARDIAN project addresses significant challenges in modern Directed Energy (DE) systems, which are crucial for defense applications targeting complex threats such as small unmanned aircraft, cruise missiles, and ground-based targets. Current machine learning techniques in these systems often function as black boxes, lacking explainability and trustworthiness, posing substantial challenges for deployment in critical defense scenarios. By leveraging the Neural Collapse phenomenon, GUARDIAN aims to overcome these obstacles by developing theoretical foundations for explainable and trustworthy AI within DE applications. This approach will transform current black-box models into fully explainable and mathematically understood systems, crucial for achieving the trust required by defense strategies and directives. Addressing this issue will enhance trust in AI systems, align with the DoD's Responsible AI Strategy and Implementation Pathway, and comply with DoD Directive 3000.09 on Autonomy in Weapons Systems. Furthermore, it will improve operational effectiveness by providing better decision-making insights, significantly enhancing DE systems' capabilities. Phase I will lay the foundational groundwork by:Dataset Selection and Preparation: Select and preprocess relevant Defense/Directed Energy and public datasets to ensure compatibility with Neural Collapse experiments.Network Training and Preliminary Experiments: Train convolutional neural networks (CNN) to achieve Neural Collapse, using David Donoho's methodology to drive cross-entropy loss to zero; validate Neural Collapse (NC1 to NC4) on both Directed Energy and public datasets.Task-Specific Information (TSI) Analysis: Evaluate performance improvements and optimize neural network architectures using TSI.Patch-Based Insights for Uncertainty Quantification: Quantify uncertainties and integrate these metrics with Neural Collapse findings through patch-based analysis.Innovative Architectures and Preliminary Research: Identify promising tasks for Phase II based on preliminary results. The project promises a significant impact on the Air Force Research Laboratory (AFRL) and other DE organizations by addressing black-box incompatibility for DE-specific target acquisition and tracking problems. Additionally, the GUARDIAN project has strong potential for generalization to other DoD weapons systems and commercial applications, such as medical diagnostic systems. The goal is to advance neural network technology beyond traditional black-box models towards fully explainable, mathematically understood models, meeting the need for trustworthy AI across various domains.