An investigation of the visual coding of faces using kernel canonical correlation analysis

METHODS Unlike other multivariate techniques, kCCA associates the fMRI voxel data with information in stimulus images (Hardoon, et al., 2007). Specifically, kCCA seeks maximally correlated pairs of weights over fMRI voxels and stimulus features. This yields a set of orthogonal pairs of linear combinations of fMRI voxels and stimulus features. Consequently, fMRI voxels are assigned weights which collectively may be viewed as an image, representing the distributed patterns of brain responses which relate statistically to stimulus information. When stimuli have common image features which overlap spatially (as with faces), the stimulus “weight maps” may also be viewed as interpretable images, showing the spatial distribution of pixels relating to the corresponding brain response pattern. Using high resolution fMRI, we scanned ventral extra-striate brain regions while participants viewed blocks of faces comprising …