The influence of prior knowledge and related experience on generalisation performance in connectionist networks

Abstract

The work outlined in this paper explores the influence of prior knowledge and related experience (held in the form of weights) on the generalisation performance of connectionist models. Networks were trained on simple classification and associated tasks. Results regarding the transfer of related experience between networks trained using back-propagation and recurrent networks performing sequence production, are reported. In terms of prior knowledge, results demonstrate that experienced networks produced their most pronounced generalisation performance advantage over naïve networks when a specific point of difficulty during learning was identified and an incremental training strategy applied at this point. Interestingly, the second set of results showed that knowledge learnt about in one task could be used to facilitate learning of a different but related task. However, in the third experiment, when the network architecture was changed, prior knowledge did not provide any advantage and indeed when learning was expanded, even found to deteriorated.