Visual mental imagery: A view from artificial intelligence.

This article investigates whether, and how, an artificial intelligence (AI) system can be said to use visual, imagery-based representations in a way that is analogous to the use of visual mental imagery by people. In particular, this article aims to answer two fundamental questions about imagery-based AI systems. First, what might visual imagery look like in an AI system, in terms of the internal representations used by the system to store and reason about knowledge? Second, what kinds of intelligent tasks would an imagery-based AI system be able to accomplish? The first question is answered by providing a working definition of what constitutes an imagery-based knowledge representation, and the second question is answered through a literature survey of imagery-based AI systems that have been developed over the past several decades of AI research, spanning task domains of: 1) template-based visual search; 2) spatial and diagrammatic reasoning; 3) geometric analogies and matrix reasoning; 4) naive physics; and 5) commonsense reasoning for question answering. This article concludes by discussing three important open research questions in the study of visual-imagery-based AI systems-on evaluating system performance, learning imagery operators, and representing abstract concepts-and their implications for understanding human visual mental imagery.