Visual routines and attention

The human visual system solves an amazing range of problems in the course of everyday activities. Without conscious effort, the human visual system finds a place on the table to put down a cup, selects the shortest checkout queue in a grocery store, looks for moving vehicles before we cross a road, and checks to see if the stoplight has turned green. Inspired by the human visual system, I have developed a model of vision, with special emphasis on visual attention. In this thesis, I explain that model and exhibit programs based on that model that: (1) Extract a wide variety of spatial relations on demand. (2) Learn visuospatial patterns of activity from experience. For example, one program determines what object a human is pointing to. Another learns a particular pattern of visual activity evoked whenever an object falls off a table. The program that extracts spatial relations on demand uses sequences of primitive operations called visual routines. The primitive operations in the visual routines fall into one of three families: operations for moving the focus of attention; operations for establishing certain properties at the focus of attention; and operations for selecting locations. The three families of primitive operations constitute a powerful language of attention. That language supports the construction of visual routines for a wide variety of visuospatial tasks. The program that learns visuospatial patterns of activity rests on the idea that visual routines can be viewed as repeating patterns of attentional state. I show how my language of attention enables learning by supporting the extraction, from experience, of such patterns of repeating attentional state. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)