Describing, Using "Recognition Cones"

Abstract

Parallel-serial perceptual "recognition cones" (RE-CO-DERS) are being developed to handle scenes of interacting objects, by successively transforming and coalescing information. Recognition cones apply the same mechanisms to pre-process, characterize, find parts of objects, find qualities, name, and describe. They attempt to model living perceptual systems, in their generality of functions and homogeneity of processes, as well as in their overall structure. A recognition cone program for scenes of interacting objects is presented, described and discussed. The concept of a "description" is discussed, and a variety of types of descriptive information is surveyed. The recognition cone is then extended to handle many of them. Several further problems in describing are discussed, including the importance of culling information from the overly large - indeed potentially infinite - exhaustive description that is possible, and using an understanding (whether built-in, or through commands, conversation, or learning), of the recipient of the description - the hearer - to help determine relevance.