Comments on: ";Expert" vision systems: Some issues

operations at each of an N by N-based pyramid’s log N layers, e.g., to assess (moving inward from the raw image) gradients, short edges and colors, curves and longer edges, angles and textures, simple enclosures and other shapes, and simple objects, more complex shapes and objects-and continuing into more and more complex shapes, objects, and groups of objects. Such a software/hardware system resembles-and differs from-“expert production systems” in several important ways: It uses explicit information represented in micro-modular transforms; these are reminiscent of simple productions, but they are extended so that they can handle information about spatial relations, thresholds, weights, strengths, certainties, etc. It applies its productions in parallel, combining their results as appropriate-and therefore (when realized on appropriate parallel hardware) can be made extremely fast and flexible, and capable of handling large amounts of variation and distortion. It imposes a structure over its transforming operations that gives them a shallow (logarithmic) serial depth that allows them to examine successively more global information in a simple, natural progression. Potentially, such a system should be able to combine “symbolic, semantic” and “iconic, pictorial” information, using each where appropriate, and to process images fast enough to handle scenes input in real time from the real world.