Abstract
Individual differences in face recognition ability are positively correlated with the blood-oxygen-level-dependent (BOLD) response in ventral face-selective areas, but the aspects of tuning to faces that differ between people in these areas remain unclear. Using a multi-feature voxelwise encoding model framework, we investigated cortical tuning to both structural and semantic face features in individuals who varied in face recognition ability as assessed by the Cambridge Face Memory Test (CFMT). Functional MRI data were acquired while 37 participants (21 women) viewed 936 natural faces varying in perceived race, age, gender and emotional expression. We fit both structural and semantic feature models to estimation set data using regularized linear regression. The estimated beta weights were applied to image features for novel faces viewed during validation runs; Pearson correlations were calculated between predicted and actual BOLD time-series for each voxel. The number of significantly predicted voxels (p<0.05) within each participant-defined region of interest was used as a measure of model performance. There was no significant relationship between CFMT scores and the fit of structural face models. In contrast, semantic model fit did vary as a function of CFMT scores. Specifically, higher CFMT scores were associated with stronger representation of identity features in occipital and fusiform face areas. This was not observed for emotion features. Decomposition of the identity feature model into high-level (e.g. gender, age) and low-level features (e.g. short hair) revealed that higher CFMT scores were uniquely associated with stronger representation of high-level identity features in both OFA and FFA. These findings suggest that individual differences in face recognition ability reflect differential encoding of more abstract, higher-level aspects of identity in ventral face-selective areas.