Recognition by directed attention to recursively partitioned images

A learning/recognition model (and instantiating program) is described which recursively combines the learning paradigms of conceptual clustering (Michalski, 1980) and learning-from-examples to resolve the ambiguities of real-world recognition. The model is based on neuropsychological and psychological evidence that the visual system is analytic, hierarchical, and composed of a parallel/serial dichotomy (many, see conclusions by Crick, 1984). Emulating the experimental evidence, parallel processes in the model decompose the image into components and cluster the constituents in much the same way as the image processing technique known as moment analysis (Alt, 1962). Serial, attentive mechanisms then reassemble the decompositions by investigating spatial relationships between components. The use of attentive mechanisms extends the moment analysis technique to handle alterations in structure and solves the contention problem created by combining the two learning paradigms. The contention results from a disagreement between the teacher and the model on what constitutes the salient features at the highest level of the symbol. There are four cases ZBT must handle, two of which result from the disagreement with the teacher. The parallel/serial dichotomy represents a vertical/horizontal tradeoff between the invariant and variant features of a domain. The resultant learned hierarchy allows ZBT to recognize structural differences while avoiding problems of exponential growth