BIOLOGICAL VS ARTIFICIAL REPRESENTATIONS OF VARIABLES IN FINE-GRAIN PARALLEL SYSTEMS.

A framework called Variable Representation Space (VRS) for the classification of connectionist representations of variables has been developed. The VRS facilitates a formal analysis of the classes of representations possible in connectionist networks. The VRS is used to suggest that the types of variable representations used in artificial connectionist vision systems are different from the types used in biological vision systems. A study of the biological types of representations suggests their advantages for encoding variables in computer vision systems. The detection of variable values is discussed in terms of the often ignored distinction between the image signals and the conceptual features used to describe the image, such as 'edges'. Since an edge is a conceptual entity, it is not possible to design an image operator which can detect edges, as image operators act only upon the image signal. The solution presented here is to determine the response function of a given operator to image signals which systematically vary along the relevant dimensions. Then the response of the operator to an arbitrary image can be correctly interpreted in terms of the response function.