MIT Department of Electrical Engineering & Computer Science

DATE: THURSDAY, MARCH 16, 1995

TIME: TALK AT 4:15
Refreshments at 4:00

PLACE: NE43-518

Visual Architecture and Cognitive Architecture

Ian Horswill
MIT Artificial Intelligence Laboratory

The subfields of AI have drifted apart in recent years. An example is the lack of work in bridging vision and core AI (reasoning, language, etc.). I will present an implemented system that answers questions about simple scenes using a biologically plausible vision system based on Ullman's theory of visual routines. The system's cognitive architecture is unusual in that it exhibits compositional semantics without building trees or assuming a global pointer mechanism. The visual system is novel in that (1) it runs in real-time on low-cost hardware, (2) it efficiently performs a range of tasks in real time using a fixed collection of components, and (3) its interface is a useful intermediate point between forcing vision to speak in logic and forcing reasoning to speak in CAD models.

I will show how the visual system can be used to emulate the visuo-motor capabilities of the Polly robot, to perform visual searches to find satisfying variable assignments of Horn clauses, and to find referents of simple noun phrases. I will also discuss a number of formal limitations that suggest interesting psychophysical experiments.

HOST: Prof. Patrick Winston