The joint learning of both mappings in a single-pathway appears to be difficult or impossible. The corollary of these computational insights is that the double dissociations between certain types of aphasia (e.g., conduction aphasia—impaired repetition versus semantic dementia—impaired

Selleck JAK inhibitor comprehension and speaking/naming) reflect these same divisions of labor in the human brain. The simulations also suggest that the division of labor between the two pathways is not absolute or mutually exclusive. The two pathways work together to deliver adult language performance (and aphasic naming and repetition abilities; see Nozari et al. [2010]). This division of labor represents one solution for an intact, fully-resourced computational model. The solution is not fixed, however, and following damage, processing can be reoptimized both within and across the two pathways, thereby mimicking spontaneous recovery observed post stroke (Lambon Ralph, 2010, Leff et al., 2002, Sharp et al., 2010 and Welbourne and Lambon Ralph, 2007). These simulations suggest that this recovery sometimes comes at the cost of Autophagy inhibitor screening library other functions

(e.g., more of the computation underpinning repetition can be taken up by the ventral pathway but this is only possible for words and not nonwords). Analysis of each layer in the model demonstrated that the internal similarity structure changed gradually across successive regions. In line with found recent neuroimaging results (Scott

et al., 2000 and Visser and Lambon Ralph, 2011), the ventral pathway shifted from coding predominantly acoustic/phonological to predominantly semantic structure. Additional control simulations (comparing this multilayer pathway with a single, larger intermediate layer; see Figure S3) indicated that this gradual shift led to much better performance when extracting the modality-invariant meaning from the time-varying auditory input. Finally, a second key finding from these analyses is that the structure of the representations can change across tasks even within the same region. For example, the aSTG is much more sensitive to semantic similarity during speaking/naming than in comprehension, a fact that might explain recent VSLM data (Schwartz et al., 2009) (see Results). If correct, then this result has clear implications for the limits of the subtraction assumption (Price et al., 1997), commonly utilized in functional neuroimaging. When implementing any cognitive or neural hypothesis in a computational model various assumptions have to be made explicit. In this section we outline our working assumptions and the rationale underlying them. We then provide a summary of implementational details. Copies of the model files are available from the authors upon request.