Dynamical Neural Networks Construction for Processing of Labeled Structures

We show how Labeling RAAM (LRAAM) can be exploited to generate `on the fly' neural networks for associative access of labeled structures. The topology of these networks, that we call Generalized Hopfield Networks (GHN), depends on the topology of the query used to retrieve information, and the weights on the networks' connections are the weights of the LRAAM encoding the structures. A method for incremental discovering of multiple solutions to a given query is presented. This method is based on terminal repellers, which are used to `delete' known solutions from the set of admissible solutions to a query. Terminal repellers are also used to implement exceptions at query level, i.e., when a solution to a query must satisfy some negative constraints on the labels and/or substructures. Besides, the proposed model solves very naturally the connectionist variable binding problem at query level. Some results for a tree-like query are presented. Finally, we define a parallel mode of execution, exploiting terminal repellers, for the GHN, and we propose to use terminal attractors for implementing shared variables and graph queries.