Exploring the influence of architecture and learning algorithm on the breakdown of the "hub" model of semantic cognition

Guest, O. , Cooper, R. & Davelaar, E.

Birkbeck, University of London

The hub model of "semantic cognition" was shown in Rogers et al. (2004) to reproduce the behaviour of neurological patients who perform poorly on a variety of tests of semantic knowledge. The model aims to provide a comprehensive explanation for semantic deficits as found in patients with semantic dementia (SD) and, as extended in Lambon Ralph et al. (2007), individuals with herpes simplex virus encephalitis (HSVE). The authors appeal to the underlying neural network's properties, mainly the emergence of attractors, which provide the network with a mechanism for relating perceptual input to an internal amodal semantic state. Not only does this model emulate SD and HSVE semantic impairments, but it also underpins a theoretical account of such semantic disturbances. As such, this model provides a useful starting point for further work to encompass even more deficits. However our attempts to reimplement the Rogers et al. model in compatible neural network types, including two recurrent network models and one Boltzmann machine model, have been largely unfruitful. On one hand we were able to successfully teach the models the training patterns and thus recreate "normal" behaviour, but on the other hand, we did not manage to fully replicate the scores of semantically impaired patients by lesioning the models. Our results suggest that while semantic impairments reminiscent of patients may arise when the Rogers et al. model is lesioned, such impairments are not a necessary consequence of the model. We discuss the implications of these apparently negative results for the Rogers et al. account of semantic cognition, focusing on both reconciling the model's behaviour with patient scores and on the model's limitations.