Sangeet Khemlani (Naval Research Laboratory - Navy Center for Applied Research in Artificial Intelligence)
Title: “The Process of Explanation”
Abstract: Explanations are a hallmark of human thinking; they shape people's thoughts, beliefs, and social interactions. Recent work investigates how people compare competing explanations, but few studies examine how people generate explanations in the first place. Artificial intelligence systems, such as deep networks and classifiers, have difficulty producing even rudimentary explanations of their own behavior. I argue that the generative process will remain opaque to researchers until it can be implemented in an algorithm. To develop the algorithm, at least three issues need to be resolved: how such an algorithm halts; the purpose of its output; and how it can operate rapidly. My talk will present recent studies that describe progress in resolving each of these issues. Humans consider an explanation complete -- they halt their search -- when it refers to a single mental simulation. They explain in order to resolve conflicts. And they rapidly construct explanations by privileging certain conceptual connections over others. I will conclude by describing a computer model of human reasoning developed in my lab that makes it possible to implement a theory of how people generate explanations.