ABSTRACT
This paper outlines a statistical approach to unifying certain symbolic and neural net architectures, by deriving them from a stochastic domain model with sufficient structure. The goal is to derive neural networks which retain some of the expressive power of a semantic network and also some of the pattern recognition and learning capabilities of more conventional neural networks. The domain model is a stochastic L-system grammar, whose rules for generating objects and their parts each include a Boltzmann probability distribution. Using such a domain model in high-level vision, it is possible to formulate object recognition and visual learning problems as constrained optimization problems (Mjolsness 1991) of a restricted form which can be locally optimized by suitable neural network architectures. In ths way, one can semi-automatically produce neural nets for object recognition with some of the representational properties of semantic nets.
