Integrating numerical and syntactic learning models for pattern recognition

Abstract

In this paper we consider how the recognition, interpretation of image structures, patterns, objects can be posed in terms of "Inductive Bayesian Networks" (IBN) which combine syntactic domain models with the numerical/statistical characteristics of what is sensed. The net result of this formulation is the production of contextual and relational rules which can be used to summarize, generalize structural descriptions from examples in ways which are consistent with domain knowledge. In this approach the associated algorithms are also constrained by principles of Minimum Description Length (MDL) which endeavor to produce structural descriptions which generalize over numerical data attribute while specializing over symbolic description length. Examples in pattern and object recognition are discussed.