Computationally, separate but interacting networks for semantic control and representation resolve a long-standing puzzle. On one hand, concepts must generalize across contexts (e.g., canaries, like other birds, lay eggs). On the other hand, different situations often require us to retrieve diverse conceptual properties to complete tasks (e.g., vs. when either spotting or catching a canary). As shown in the Figure (Panel A), both challenges are solved by implemented computational models which have separate but interacting networks for control and representation10. As per the hub-and-spoke framework, modality-specific information interacts with a transmodal hub (purple square). A separate region represents the current task context (yellow square). These two components interact with an integrative system (black circle), which dynamically and transiently reshapes the multidimensional similarity structure arising at the context-independent hub in order to generate task-, time- and context-relevant behavioural responses. Thus, as shown in Panel B, if two contrastive tasks (e.g., spotting vs. catching the canary) require focus on colour vs. movement properties, then the conceptual∩task) integrative layer generates context-relevant internal representations congruent with the target behaviour. Because the hub is not directly connected to task/context information, it learns representations which capture structure that is independent of the various and idiosyncratic contexts encountered, providing the core computational basis for cross-context conceptual generalization (see ‘The hub-and-spoke’ Section in the main text). The same computational characteristic does not arise in models that blend task/context and perceptual inputs, immediately, within a single intermediating hub10. Thus, while this model was initially advanced solely to account for cognitive phenomena, it leads to the conclusion that semantic cognition requires separate and interacting neural networks for representation and control (as implicated in the clinical and cognitive neuroscience literatures; see Main Text).
