Pushing Up Hypotheses Using Context-Dependent Links

Perceptual understanding systems often involve building hierarchical groupings. The elements of such a hierarchy range from primitive elements at the lowest levels (such as pixels, in the case of image-understanding systems) to representations of complete real-world objects at higher levels. At each level, data are grouped together because they are identified as component parts of a single feature at the next level; the domain-specific work lies in correctly identifying these groups through all the levels. The construction of such a hierarchy generally proceeds from raw data input up to top-level explanations, however, it usually cannot operate in a strictly bottom-up fashion. Often a number of hypotheses must be made as to the correct grouping, one of which is eventually selected according to a higher-level constraint; that is, there is also a top-down aspect to the process. The number of primitive elements is typically quite large, and if several levels of hypothesizing are necessary, the number of hypotheses can increase exponentially; efficient means of representing and managing hypotheses are therefore desirable. This paper describes such a method, called context-dependent links. Context-dependent links allow the system to conserve memory (and for many operations, time) by exploiting the redundancies between different grouping hypotheses. At the same time, they make explicit certain natural relationships among the hypotheses themselves.