Explainable AI via Linguistic Summarization of Black Box Computer Vision Models

There is an ever-growing demand to characterize and understand AI as it is integrated into everyday life. Linguistic summaries have been previously used to provide natural language descriptions of data and models. However, the number of possible summaries increases rapidly with the number of data attributes. To make sense of the vast number of possible linguistic statements for a system, we introduce a hierarchical approach for generating and ranking linguistic statements. Each description of the model is assigned a value based on user criteria, allowing summaries to be tailored to specific users. We provide visualizations of the generation of summaries for a popular computer vision detector on a synthetically generated dataset.