A comparison of Neural and Graphical Models for Syntactic and Structural Pattern Recognition

Recent developments in the theory and uses of Bayesian Networks in pattern recognition and image understanding (PRIU) raise questions about the relationships between Bayesian compared to non-Bayesian approaches. In this paper we compare Neural-based verses Bayesian-based methods for PRIU. We conclude with the view that a singular PRIU architecture that models “from pixels to predicates” in one explicit system model, is most desirable from an optimization perspective and that hierarchical hidden Markov random fields are one example of such an approach but where algorithms from neural computing also apply.