Extended Data Fig. 4: Univariate RNC algorithm.

Univariate RNC searches for images leading to aligned or disentangled in silico univariate fMRI responses of two visual areas. The 73,000 NSD images are fed to the trained encoding models of two areas, and the resulting in silico fMRI responses averaged across voxels, obtaining a one-dimensional univariate response vector of length 73,000, for each area. The univariate response vectors of the two areas are either summed (alignment) or subtracted (disentanglement), the sum or difference scores ranked, and the controlling images leading to highest and lowest scores are kept. This results in four sets of controlling images, each set corresponding to a different neural control condition. The controlling images from the sum vector lead to two neural control conditions in which both areas have aligned univariate responses (that is, images that either drive or suppress the responses of both areas), whereas the controlling images from the difference vector lead to two neural control conditions in which both areas have disentangled univariate responses (that is images that drive the responses of one area while suppressing the responses of the other area, and vice versa). Photos from the COCO image dataset/Flickr⁸⁵.