Methodology and validity in the construction of computational models of cognitive deficits following brain damage

Over recent years, neural network models of several cognitive neuropsychological disorders have been developed. These include word recognition difficulties, face recognition difficulties, attentional deficits, visual processing impairments, semantic deficits, and aphasia. These models are useful in various ways. Firstly, they require detailed specifications of theories, and can focus attention on critical assumptions. Secondly, they can query alternative theories, and provide predictions which can be verified by testing patients. In this paper, issues relating both to the methodology and validity of attempts to model cognitive deficits using neural networks will be discussed, providing examples from several studies. Issues discussed will include the requirement for models to perform normally prior to damage, and to show potential effects of rehabilitation or partial recovery following damage. A single model should be able to incorporate multiple symptoms of a deficit and ideally also multiple syndromes when different lesions are introduced. The model must also be able to handle variability between patients with the same syndrome, and even with the same patient at different test sessions.