Connectionist realization of semantic networks using an adaptive associative memory, AAM

A new connectionist semantic network that can perform a class of inheritance and recognition inference problems based on conceptual hierarchies extremely fast is presented. The computation time for inference tends not to increase with the size of conceptual hierarchies and to increase sublinearly with the input complexity. All of the inference problems are performed by transforming them into lower level processing: pattern recovery and pattern segmentation. The network is very simple in its architecture and has high flexibility: (a) the existence of exceptions in property inheritance does not affect the network´s performance; (b) modifications (addition and deletion) of knowledge can be easily performed by one-shot relearning; (c) generalization is a basic system property in the sense that concepts automatically acquire any effects possessed by their subordinate concepts