Stochastic double dissociations in distributed models of semantic memory

Abstract

We present a preliminary set of connectionist models of impairments to semantic memory, exploring the conditions under which double dissociations between knowledge of living and non-living entities occur. Small et al [1] argue that category specific impairments are a consequence of semantic feature based representations in a fully distributed memory system. Farah and McClelland [2] argue that category specific impairments arise due to modular structure in semantic memory, albeit structure that is specific to modality; they hypothesise that living and non-living entities have a differential reliance on perceptual and functional features. We evaluated these respective claims by lesioning a simple autoassociative model of semantic memory, using a 2x2 design: fully distributed architecture versus partially modular architecture with modality specific channels; Small et al’s training set versus a training set constructed according to Farah and McClelland’s perceptual/functional scheme. One thousand stochastic lesions were applied to each network ‘subject.’ The results supported Farah and McClelland: on average, double dissociations required modular structure and differential reliance on modalities. However, by choosing select (i.e. rare) lesions from each set of 1000, double dissociations of living versus non-living knowledge were found in both networks using both training sets. We discuss the idea that statistical distributions of impairments in patients with similar lesions are necessary to compare against the predictions of functional models, and thus that single case studies may be insufficient to distinguish distributed and modular architectures.