School of ECM University of Surrey Guildford, Surrey GU2 5XH, UK

Tel: +44 (0)1483 259823 Fax: +44 (0)1483 876051

The motives for simulating language comprehension or production, and for implementing connectionist networks for the purpose of simulating aphasia, are varied. For some, it is a way of cataloguing experimental data and analysing possible patterns within them, while for others the networks are in themselves potential models of aspects of human cognition.

We are currently investigating the utility of a connectionist modelling approach based on the notion of the modular neural network architecture. Such architectures consist of interconnected collections of individual artificial neural networks, where each component may differ in terms of network topology or learning algorithm employed.

Modular connectionist architectures may demonstrate both supervised and unsupervised learning, and consequently show an ability to both generalise and take into account aspects of the environment. By modelling each proposed division of the human language-processing system using an individual neural network which has been chosen on the basis of its suitability for modelling that division, we obtain a greater degree of freedom for exploring the effects of both global and local lesioning on the behaviour of the model. Our intention here is to provide the cognitive neuropsychology community with data structures and algorithms which have proven mathematical properties and which may be better suited to cope with the intricacies of real-world patient data, which must necessarily form the foundation of any well-grounded connectionist modelling work.