Fig. 4.

Model-based pipeline to reconstruct spatial priority maps. a fMRI was used to measure evoked activity for each image in the experiment. Each image was parameterized into a CNN feature-space. We learned the transformation between fMRI activity and CNN activity separately for each CNN layer using PLSR in a LORO cross-validated fashion. These learned transformations were applied to left-out data to align fMRI activity to CNN activity from each layer. b To reconstruct spatial priority maps from CNN-aligned fMRI activity, we applied the spatial attention model developed on computational CNN activity. CNN-aligned fMRI activity from each layer was averaged across the feature-based dimension to produce reconstructed layer-specific activity maps, which were then averaged together to produce a reconstructed spatial priority map for each image (base reconstruction). In a separate analysis, the reconstructions were smoothed with a 2D Gaussian kernel (SD = 24 px) and pointwise multiplied by a centered 2D Gaussian kernel (600 × 600 px, SD = 600 px, resized to image resolution of 600 × 800 px) to account for center-bias (smoothed and center-bias corrected reconstruction)