Modelling Brain Lesions in Neural Networks

The observation of patients with lesions in the brain and specific cognitive functions impaired has led during the last 130 years to stress modular organization and physical localizations of different functions. If this point of view is pushed to the extreme, one would have to conclude as a consequence that there is no place left for disordered models as the ones that derive from neural networks. We argue that theoretical estimates of the burden so imposed on the organization capabilities of neural development and/or natural evolution make this extreme position untenable. We then show in one particular example (the syndrome called “prosopagnosia”) that an alternative explanation is possible. Assuming a completely amorphous neural model that stores categorized ultrametric patterns we use the probability measure on synaptic interactions introduced by Gardner to analyze the effect of a random destruction of synapses. At an intermediate level of destruction the retrieval of the pattern is impaired particularly in those items that permit the distinction among exemplars in a class.