The neural representation of visual relational and item-specific information

Relational representations sit at the core of models of higher cognition and they enable distinctively human flexible behavior, yet their neural basis has been little studied. In order to advance our understanding of how the brain instantiates relational representations and whether it does so in a manner distinct from other sorts of information, in this work we employed fMRI and multivoxel pattern analysis to examine the neural representation of visual relational in comparison with item-specific information in the spatial and object property domains. Item-specific spatial information is necessary for action and for feature-binding. Spatial relational information is necessary for view-independent recognition and for spatial reasoning. Item-specific information about object properties such as size and luminance is necessary for such tasks as object or scene recognition. Information about relative feature values between objects is necessary for recognition under differing viewing conditions and for inference and generalization. Subjects performed a delayed recognition working memory task on simple visual scenes that differed across conditions only in the type of information that was to be encoded. In the spatial domain, we found a double dissociation between items and relations: whereas item-specific processing implicated a frontoparietal attention network including the superior frontal sulcus and intraparietal sulcus, relational processing preferentially recruited a cognitive control network, particularly lateral prefrontal cortex and inferior parietal lobule. Moreover, pattern classification revealed that the actual meaning of the relation could be decoded within these same regions, most clearly in rostrolateral prefrontal cortex. These results suggest that spatial relational information is maintained as a distinct type of information from object locations, and support a hierarchical, representational account of prefrontal cortex organization. In the nonspatial, object domain, in contrast, we found that both item and relation tasks activated visual and parietal cortex to similar degrees. However, multivariate methods could identify distinguishable neural populations representing item and relational information within these areas. In addition, modeling results suggest that item-specific and relative magnitude representations share a common representation scheme within parietal cortex. Together, these findings demonstrate that relational and item-specific processing is often closely related, but also reveal that there are at least certain types of relations whose representations are maintained and used in a qualitatively distinct manner in prefrontal cortex when explicit, symbolic processing is required, thus informing models of human cognition and prefrontal cortical organization.