ABSTRACT
Mixture of knowledge held in weights and short-term information held as unit activation has been the method of dealing with short- and long-term information in most connectionist cognitive modelling. This chapter introduces recently developed dual-store connectionist architecture and speculates on the variety of cognitive phenomena that might prove suitable for modelling by networks with weights that interact on different timescales. It is helpful to introduce a simple taxonomy of connectionist architectures with short- and long-term components. The third main category of systems where short- and long-term capabilities interact is one where a single connection between a pair of units can carry more than one weight. These are systems with dual- or multi-weight connections. Glasspool describes an extension to Burgess & Hitch's connectionist model of the phonological loop that accounts for lexicality effects. Dual-component and dual-weight connectionist architectures offer the opportunity to build explicit computational models. The chapter discusses possible ways in dual-store model can be used to simulate hippocampal function.
