Abstract
If neuronal models are successful, they will account for specifically human complex behaviours. Most of these behaviours can be described as governed by rules. In recent years, much effort has been spent to elucidate the neuronal basis of rules, and many researchers have focussed on modelling rules and regularities underlying language, particularly those relevant for past-tense formation. After introducing problems posed by past-tense formation, important aspects of recent controversies between connectionists and linguists concerning the nature of rules will be reviewed and analysed on the basis of elementary simulations. It is argued that modular networks with varying connection probabilities between their layers would be ideal for modelling learning and processing of past-tense formation. The motivation for postulating such networks comes from neurobiological models of language and from neuroanatomical data about cortico-cortical connectivity. Furthermore, such modular networks may explain double dissociations of regular and irregular past-tense formation in neurological patients, as reported in recent neuropsychological publications. It is concluded that past-tense formation does not pose problems to pattern associators, given that some structure is built into the network which approximates wirings in the human cortex.