REU SUPPLEMENT: Parallel Processing Projects in Artificial Intelligence

点击翻译按钮获取中文摘要

This award provides support to develop computing infrastructure for a wide variety of research with the primary areas being: plausible ascription of beliefs by an AI system; complex symbolic information processing in connectionist systems; associative network databases for computer vision; the Model Generative Reasoning system; and parallel algorithms for scene model generation. The new infrastructure includes the upgrading of a parallel processor from a 4 processor system to a 22 processor system, the acquisition of a number of workstations, and the provision of hardware and software technicians. In the Beliefs/Metaphor project on natural language discourse understanding, two complementary approaches to integrating the processing of metaphors and the processing of information about participants'belief are being explored. In one approach, metaphors are processed by a simple extension of the belief-reasoning machinery in an existing program, whereas the other approach bases the representation of belief states on common-sense metaphorical models of mind that are often exploited in discourse. In the Connectionism project, connectionist information-processing techniques are being developed that are useful in high-level cognitive tasks, such as natural language understanding and common-sense reasoning, as opposed to the perceptual and adaptation tasks to which most connectionist work has been directed. Part of the work is exploiting signal timing effects that occur in networks inspired by real biological neural networks. The purpose of the vision database project is to design, implement, and test a database (appropriate for robotics applications) which supports associative organization and retrieval, clustering by physical properties, higher levels of abstraction, and efficient search for matching entities. The paradigm used is an extension of Pathfinder networks (originally intended to model human semantic memory) called monotonic search networks, which is a type of proximity graph. Model Generative Reasoning is a general automated problem solving system developed by Coombs and Hartley. It makes expert problem solvers robust in the light of conflicting evidence through the maintenance of alternative hypotheses in a proper abduction/deduction cycle. In the Parallel Algorithms for Scene Model Generation project, time- varying imagery based on natural scenes is being analyzed with the goal of developing dynamic three-dimensional models of the scenes. Important features of the work include a hypothesis-and-test paradigm for construction of the models, merging of multiple cues through the use of local consistency constraints, and maintenance of model consistency in a parallel environment through data locking protocols developed automatically from graph grammar productions.