A hybrid connectionist model of serial order behaviour

Abstract

Structure-sensitive symbolic representations are traditionally favoured over connectionist models using distributed representations to explain systematicity in cognition. It has been argued that connectionist models can encode structural information, but their ability in the systematic and free transformation of component ideas in conscious mental processes is in question. A two-layer connectionist network employing a structured representation was constructed to investigate its capacity to locally transform a serial order. Simple weight exchange is sufficient for the transformation, which is difficult with a simple recurrent network employing distributed representations. This finding confirms the necessity of a structured representation in displaying systematicity.