State-Trace Analysis of Sequence Learning by Simple Recurrent Networks

Abstract

This study investigated the use of state-trace analysis (Bamber, 1979) when applied to computational models of human learning. We aimed to investigate the performance of simple recurrent networks (SRNs) on a sequence learning task. Elman’s (1990) SRN and Cleeremans & McClelland’s (1991) Augmented SRN are both benchmark models of human sequence learning. The differences between these models, comprising of an additional learning parameter and the use of response units activated by output units constituted our main manipulation. The results are presented as a state-trace analysis, which demonstrates that the addition of an additional type of weight component, and response units to a SRN produces multi-dimensional state-trace plots. However, varying the learning rate parameter of the SRN also produced two functions on a state-trace plot, suggesting that state trace analysis may be sensitive to variation within a single process.