Modelling Cognitive Development with Constructivist Neural Networks Gert Westermann Sony Computer Science Laboratory, 6 rue Amyot, 75005 Paris, France In: Proceedings of the Sixth Neural Computation and Psychology Workshop, 2000. Springer. Abstract Based on recent evidence from cognitive developmental neuroscience, I argue for the importance of constructivist models of cognitive developmental phenomena. This point is empirically investigated with a constructivist neural network model of the acquisi- tion of past tense/participle inflections. The model dynamically adapts its architecture to the learning task by growing units and connections in a task-specific way during learning. In contrast to other, fixed-architecture models, the constructivist network dis- plays a realistic, U-shaped learning behaviour. In the trained network, realistic “adult” representations emerge that lead to aphasia-like dissociations between regular and ir- regular forms when the model is lesioned. These results show that constructivist neural networks form valid models of cognitive developmental processes and that they avoid many of the problems of fixed-architecture models. 1. Introduction The computational modelling of psychological processes necessarily involves abstrac- tions. Every modeller makes often implicit assumptions about which aspects of the original system can be abstracted away without compromising the model’s value for explaining the observed phenomena. For example, much has been written about the relationship between biological and artificial neural networks, and it seems clear that they only share the most basic principles of operation, abstracting away most of the details of neural processing that are not considered essential. Nevertheless, the use of artificial networks in psychological modelling is valuable because they are thought to capture the characteristic principles of biological neurons, such as learning based on complex nonlinear associations. However, in designing a model it is important that none of those aspects of the original system are abstracted away that are essential for the functioning of the mod- elled process. For example, a model of bird flight that does not take the properties of air into account will most likely not be useful. In this paper, I argue that in the case of cognitive development, constructivist learn- ing, that is, the dynamic adaptation of the learning system’s architecture to the learning