A constructivist neural network model is presented that learns the past tense of English verbs. The model builds its architecture in response to the learning task in a way consistent with neurobiological and psychological evidence. The model outperforms existing connectionist and symbolic past tense models in terms of learning and generalization behavior, and it displays a U-shaped learning curve for many irregular verbs. The trained model develops a modular architecture with dissociations between regular and irregular verbs, and lesioning the different pathways leads to results comparable with neurological disorders. It is argued that the success of the model is due to its constructivist nature, and that the distinction between fixed-architecture and constructivist models is fundamental. Given this distinction, constructivist systems provide more realistic models of cognitive development.