Collaborative Research: The Generalized A* Architecture for Perceptual Systems

提案0534820“合作研究：感知系统的广义A*架构”PI: Pedro FelzenswalbUniversity of chicago摘要该研究探讨了解决感知和概率推理中出现的广泛问题的方法。研究的重点是最小成本结构问题，其目标是在一组组合规则下找到可推导的最小成本结构。这种成本最低的结构可能是解析树、可变形模型与图像的匹配，或者为马尔可夫随机场中的变量赋值。所研究的算法是基于a *搜索的推广和对任意最小代价结构问题的抽象。寻找成本最低的抽象结构通常比寻找成本最低的具体结构要有效得多。此外，最小代价抽象结构在A*搜索意义上为解决具体优化问题提供了启发式指导。广义A*利用抽象结构跨越感知系统的所有处理层次，指导在抽象空间的层次中寻找更具体的结构。该项目预计将通过为视觉系统、语言处理系统和概率推理系统的设计和实现提供新技术和软件工具，产生广泛的影响。

Proposal 0534820"Collaborative research: The Generalized A* Architecture for Perceptual Systems"PI: Pedro FelzenswalbUniversity of ChicagoABSTRACTThis research investigates methods for solving a broad class of problems that arise in perception and probabilistic reasoning. The work focuses on least-cost structure problems where the goal is to find a least-cost structure derivable under a set of compositional rules. This least-cost structure might be a parse tree, a match of a deformable model to an image, or an assignment of values to variables in a Markov random field. The algorithms investigated are based on a generalization of A* search and abstractions to arbitrary least-cost structure problems. Finding the least-cost abstract structure is typically vastly more efficient than finding the least-cost concrete structure. Furthermore, the least-cost abstract structure provides heuristic guidance, in the sense of A* search, in solving the concrete optimization problem. Generalized A* uses abstract structures spanning all levels of processing of a perceptual systems in guiding the search for more concrete structures in a hierarchy of abstraction spaces. This project is expected to have broad impact by providing new techniques and software tools for the design and implementation of vision systems, language processing systems and probabilistic reasoning systems in general.