Computational Intelligence, Volume 9, Number zyxw 4, zyxwv 1993 SOME CHALLENGES FOR A COMPUTATIONAL THEORY OF IMAGERY BARBARA TVERSKY Department zyxwvu of Psychology, Stanford University Palo Alto, zyxwvu CA 94305-2130, U.S.A. zyxwv Imagery is an endlessly fascinating topic. Most, but not all, people report having imagery, and some seem to enjoy reporting theirs in great detail. This activity, however, seemed relegated to the domain of literature until experimental techniques were developed that showed that perception-like mental representations seemed to underlie behavior in a number of cognitive tasks. Time to retrieve information or to make judgments reflected pictorial properties such as shape and feature overlap (Posner 1969; Tversky 1969) and also apparently analog properties such as size, distance, and angle of disparity (Kosslyn 1980; Shepard and Cooper 1982). This, combined with evidence that imaging interfered with perceiving (Segal and Fusella 1970) and with later evidence that people with brain damage disturbing perception had analogous disturbances in imagery (Farah 1988), sug- gested that some of the same brain structures involved in seeing and perceiving are also involved in imaging. Other work demonstrated consensus on some aspects of the subjective experience of imagery (Kosslyn 1980). Like many early views in emerging areas of research, one popular early view of imagery -as reflecting analog visual features in the world and as accompanied by the subjective experience of having an image-has proved to be too restrictive. The same techniques used to objectify visual imagery have also been used to demonstrate that mental represen- tations can be spatial and even kinesthetic, as well as visual, and that mental representa- tions of the visual-spatial-kinesthetic world can be categorical rather than analog. Congen- itally blind people can perform a variety of classic imagery tasks, such as mental rotation (Marmor and Zaback 1976), using imagery mnemonics (Kerr 1983), and navigating in space (Loomis et al. in press), presumably because these tasks have spatial as well as visual components that can be accessed by modalities other than vision. Left-right judgments of depicted hands and feet seem to require kinesthetic mental rotation of hand or foot, rather than visual mental rotation (Parsons 1987). Mental reorientation in described environments may exhibit a categorical and not an analog pattern (Franklin 1991; Franklin and Tversky 1990). Systematic distortions in memory for maps and graphs suggest that mental repre- sentations for them are constructed by using perceptual and conceptual categories (Schiano and Tversky 1992; Tversky 1992). Finally, although performance of some tasks is accom- panied by the subjective experience of inspecting a visual image, many, especially spatial and kinesthetic ones, are not. In contrast to previous A1 implementations of imagery, Glasgow’s array representa- tions seem to be able to account for many of these newer findings as well as the classic ones. Of course, accounting for the psychological data is not Glasgow’s primary aim, but that is the aim of most interest to me. First, arrays are spatial, so they can account for phenomena that are spatial, but not visual, including metaphoric ones, such as transitive inferences on age relations. Second, arrays vary in degree of refinement, approaching analog representations at one extreme and categorical representations at the other. Thus, they can account for many of the well-known errors in cognitive maps, for example, that most people erroneously indicate that San Diego is west of Reno (Stevens and Coupe 1978) and Rome is north of Philadelphia (Tversky 1981). zyxw 0 zyxwvutsrq 195’3, Blackwell Publishers, 238 Main Street, Cambridge, MA 02142, USA, and 108 Cowley Road, Oxford, OX4 IJF, UK.