Fig. 6.

Layer-specific reconstructions predict eye movement patterns. We see a correspondence between the layer hierarchy in the CNN and the visual hierarchy in the brain, such that reconstructions of early/late layers from early/late visual regions, respectively, best predict eye movement patterns. Spatial reconstructions for each layer were calculated by averaging across the feature-based channel dimension in CNN-aligned fMRI activity. Results are shown for unsmoothed and smoothed reconstructions. All layers for all shown ROIs are significant at P < 1 × 10^–3 (determined using permutation testing). Error bars represent standard error of the mean within the current data set (n = 11, Within-individual and Internal validation) or within the external validation data set (n = 22, External Validation)