Fig. 6. Structure of OTC tuning as captured by PCA on CSVA model feature weights.

a Mapping PC scores from the group-level PCA of CSVA model feature weights, across OTC voxels, reveals considerable spatial structure in tuning to the top three PCs. Maps for two representative subjects (S1 and S2) are shown (see Fig. S10 for maps from all subjects). Voxel-wise PC scores were calculated as the product of CSVA model feature weights for a given voxel by feature loadings for each PC. A RGB color space is used; red=scores on PC1, green=scores on PC2, blue=scores on PC3. PC scores are thresholded at 6 standard deviations above and below 0, with values beyond the threshold given the maximal (or minimal) color channel value. Areas where MRI data were not acquired are shown in black. Both voxels where the CSVA model did not fit significantly and those where the CSVA model fit significantly but did not outperform the semantic only model were excluded from the PCA (these voxels are shown in gray). PCA maps using CSVA model feature weights from all voxels where the CSVA model fit significantly are given in Fig. S11. b Individual features are projected into the 3-dimensional space defined by the top three PCs, focusing on those parts of the space to which there is a strong response as shown in Fig. 6a. The first column ‘color’ shows the location in PC space using the same RGB color space as in Fig. 6a. The second column ‘features’ shows the features with loadings on the top three PCs that correspond to that location in PC space.