Multimodular networks and semantic memory impairments.

We present a multi-modular approach to neural modeling of associa-tive memory. By segregating between intra-modular and inter-modular synaptic transmission, and subjecting the latter to non-linear dendritic processing, we can successfully store memories encoded on diierent numbers of modules. This model has a striking capability of memory retrieval from partial inputs when the appropriate neurons in only few modules, or even a single module, are activated by aaerent connections. Hence, if modular aaerents are lesioned, memories that are encoded in more modules are more resilient to damage. Assuming that memories of concrete objects are encoded in a larger number of modules than abstract ones, our results provide a novel explanation of the phenomenon of category speciic semantic impairments in patients with deep acquired dyslexia.