Fig. 7: Representational relationships between visual areas adaptively vary around a typical network-level configuration.

a, MDS embeddings of the in silico univariate fMRI responses of early-, mid- and high-level visual areas for all 73,000 NSD images, indicating the typical representational relationship configuration of the visual cortical network. b, MDS embeddings of the in silico multivariate fMRI responses for all 73,000 NSD images, indicating the typical representational relationship configuration of the visual cortical network. c, MDS embeddings of the in silico multivariate fMRI responses for the controlling images from two multivariate RNC control conditions: images aligning EBA and RSC (purple condition) or disentangling PPA and RSC (orange condition). d, MDS embeddings of the in silico univariate fMRI responses for the controlling images from four univariate RNC control conditions: images driving both FFA and RSC (green condition), suppressing both V1 and PPA (blue condition), driving EBA while suppressing FFA (yellow condition) or suppressing V1 while driving V2 (red condition). In a–d, the opaqueness of the lines connecting each area reflects the proximity of these areas in two-dimensional embedding space (more opaque lines indicate higher proximity; we computed the opaqueness independently for each MDS plot, by normalizing the corresponding pairwise area distances between 0.1 (almost transparent) and 1 (fully opaque)). A higher proximity between two areas indicates a stronger resemblance of their representational content. Six representative images are shown for each MDS analysis. Photos from the COCO image dataset/Flickr⁸⁵.