dlm@allegra.att.COM (11/11/88)
Why AI may need Connectionism? A Representation and Reasoning Perspective Lokendra Shastri Computer and Information Science Department University of Pennsylvania Tuesday, November 15 at 10:00 am. AT&T Bell Laboratories - Murray Hill Room 3D436 Any generalized notion of inference is intractable, yet we are capable of drawing a variety of inferences with remarkable efficiency. These inferences are by no means trivial and support a broad range of cognitive activity such as classifying and recognizing objects, understanding spoken and written language, and performing commonsense reasoning. Any serious attempt at understanding intelligence must provide a detailed computational account of how such inferences may be drawn with requisite efficiency. In this talk we describe some work within the connectionist framework that attempts to offer such an account. We focus on two connectionist knowledge representation and reasoning systems: 1) A connectionist system that represents knowledge in terms of multi-place relations (n-ary predicates), and draws a limited class of inferences based on this knowledge with extreme efficiency. The time taken by the system to draw conclusions is proportional to the @i(length) of the proof, and hence, optimal. The system incorporates a solution to the "variable binding" problem and uses the temporal dimension to establish and maintain bindings. 2) A connectionist semantic memory that computes optimal solutions to an interesting class of @i(inheritance) and @i(recognition) problems extremely fast - in time proportional to the @i(depth) of the conceptual hierarchy. In addition to being efficient, the connectionist realization is based on an evidential formulation and provides a principled treatment of @i(exceptions), @i(conflicting multiple inheritance), as well as the @i(best-match) or @i(partial-match) computation. We conclude that working within the connectionist framework is well motivated as it helps in identifying interesting classes of limited inference that can be performed with extreme efficiently, and aids in discovering constraints that must be placed on the conceptual structure in order to achieve extreme efficiency. Sponsor: Mark Jones - jones@allegra.att.com