Explainable linguistic characterizations of black boxes
Brendan J. Alvey · 2024
[EMBARGOED UNTIL 08/01/2025] Recent breakthroughs in Artificial Intelligence (AI) have led to incredible new products and systems built off of Large Language Models (LLMs), such as ChatGPT. Linguistic models have the potential to communicate results much more effectively than graphical visualizations alone. Despite their undeniable power and usefulness, LLMs often hallucinate answers which makes them unsuitable for critical applications. This dissertation first presents work showing how advances in simulation and AI are used to improve explosive hazard detection (EHD) models. It then shows how to characterize those models through graphical visualizations. Next, a method for constructing accurate and deterministic linguistic summaries of black box (BB) models is developed and demonstrated on the EHD problem. Lastly, this foundation is extended to compare multiple BB systems. This opens the door for automated and explainable comparisons of BB, which could be used to make self-improving, closed-loop AI systems.