ABSTRACT
This paper addresses an old and fundamental problem, the role of visual imagery in cognition. While the problem has a long history in philosophy and psychology, it has had less attention explicitly directed toward it in artificial intelligence research on reasoning (as opposed to machine vision and graphics). This paper addresses the question of what it means for a cognitive agent to think directly in visual images (depictions), and how such abilities might be formalized and accomplished with computer hardware. It is argued that the identification of imagery with non-propositional or with non-digital representations is incorrect; rather, the quality of imagery that gives it a special character is that it employs non-deductive inference, and this may well be achieved by descriptive, digital representations. Furthermore, research on knowledge representation within AI suggests approaches to the classical problems of imagistic thinking. A program is described for translating propositionally stated geometric assertions into diagrammatic representations and employing a constraint-propagating procedure to manipulate the representations, thereby making inferences and testing conjectures.
