Modelling Semantic Memory & Its Impairments Following Neural Damage

This project consists of a computational model of semantic memory; the underlying architecture of which, is a Boltzmann machine, a kind of attractor neural network. The general specification for the model and the set of training patterns are derived from Rogers et al. (2004). The framework contains a semantic store, which has access to streams of high-level perceptual and linguistic input features and equivalent forms of output. The semantic layer is equipped with the ability to internalise the structure found in the input, thus creating amodal semantic concepts. To evaluate the sophistication of the model a comparison is made with semantically impaired patients. For this reason, damage is imposed, in a range of severity values, paralleling temporal lobe lesions. The overall extent, and category- or modality- specificity, of this degradation is determined using a battery of semantic tests, as given to patients with semantic impairments. Results from these tests are shown to follow a qualitatively analogous trend to the scores achieved by semantic patients. The damaged model provides an attractor-based explanation for the patient data. Nonetheless, it is believed that this model, has too narrow a scope rendering it unable to capture all the details of semantic dementia, and therefore, the semantic system in general. Thus, some extensions are proposed: a neuromodulatory system (Gotts & Plaut, 2002); and a trainable executive, decision making, semantic module (Lambon Ralph et al., 2007).