A homogeneous computational model for spatial inference on massively-parallel architectures

A computational model for two-dimensional spatial inference on massively parallel single-instruction multiple-data (SIMD) architectures is described. In the model, spatial information is represented by three basic types of parallel variable or pvar: label maps which assign unique numbers to sets of related processors (e.g. the largest cube address of the set of processors representing a connected region), feature maps which contain the property values of related sets of processors, and hypothesis maps which indicate the probability, membership, belief, etc. that a processor sets belongs to a particular class. Spatial inference involves the application of parallel operators to pvars, e.g. labeling operators to assign unique labels to related groups of processors that belong to the same class, spatial operators to compute features of connected regions, and inference operators to assign classes to regions based on their properties. The application of the model to a geographic information retrieval problem is described