Systematicity in Hebbian cell assemblies

The results of a practical investigation into the application of the Hebbian connectionist paradigm to the problem of systematicity are presented. It is demonstrated how a connectionist system based purely on biologically-plausible assumptions can perform structure-sensitive tasks on a temporal stream of input data. The system is based on the concepts of Hebbian Cell Assemblies and Adaptive Resonance Theory. By testing it against Elman's word prediction task (1991) it is shown that it can successfully and rapidly learn to process structure-sensitive data. The model is then extended to include a variable-binding mechanism based on transient weight modication. The addition allows it to quickly learn and apply ad hoc propositional rules in a systematic way.
The ability to process syntactic structures is uncommon among Hebbian networks given the well-known limitations of the Hebbian learning rule. This thesis demonstrates their power when enhanced with the tools of the Adaptive Resonance Theory. However, the ability to rapidly learn and process compositional structures at a level requiring variable binding is unique among connectionist systems in general. Thus, the work presented here not only vindicates Hebbian-based approaches to structure-sensitive processing, but it also presents an unparallelled method for rapid acquisition of structural knowledge from a serially presented input stream.