A semantic dementia-based testbed for assessing how well lesioned neural networks replicate patient behavioural data

Abstract

We introduce a way to compare neural networks on their mechanistic similarity to the brain by investigating how they perform when impaired. These networks are lesioned such that they simulate semantic dementia (SD). Two models are used for comparison: one is trained with a category objective and another with a semantics objective. Both models are used to perform the word-picture matching task, a well known task in the research and clinical assessment of SD (Rogers et al., 2004) (Rogers, Ralph, Patter- son, & Jefferies, 2015). The results show that both models produce sensible results and that we have created a proof of concept for a semantic dementia based testbed. However, there is no clear answer on which of the two models performs more like SD patients: both models produce behavioural features similar to those of SD patients.