Pruning or tuning? Maturational profiles of face specialization during typical development

Abstract Introduction Face processing undergoes significant developmental change with age. Two kinds of developmental changes in face specialization were examined in this study: specialized maturation, or the continued tuning of a region to faces but little change in the tuning to other categories; and competitive interactions, or the continued tuning to faces accompanied by decreased tuning to nonfaces (i.e., pruning). Methods Using fMRI, in regions where adults showed a face preference, a face‐ and object‐specialization index were computed for younger children (5–8 years), older children (9–12 years) and adults (18–45 years). The specialization index was scaled to each subject's maximum activation magnitude in each region to control for overall age differences in the activation level. Results Although no regions showed significant face specialization in the younger age group, regions strongly associated with social cognition (e.g., right posterior superior temporal sulcus, right inferior orbital cortex) showed specialized maturation, in which tuning to faces increased with age but there was no pruning of nonface responses. Conversely, regions that are associated with more basic perceptual processing or motor mirroring (right middle temporal cortex, right inferior occipital cortex, right inferior frontal opercular cortex) showed competitive interactions in which tuning to faces was accompanied by pruning of object responses with age. Conclusions The overall findings suggest that cortical maturation for face processing is regional‐specific and involves both increased tuning to faces and diminished response to nonfaces. Regions that show competitive interactions likely support a more generalized function that is co‐opted for face processing with development, whereas regions that show specialized maturation increase their tuning to faces, potentially in an activity‐dependent, experience‐driven manner.

studies, the definition of specialization depends on a set of statistical comparisons of the fMRI signal across different experimental conditions (Joseph et al., 2002, Joseph and Gathers, 2003, Gathers et al., 2004, Joseph et al., 2006a, Joseph et al., 2011 whereas others have combined fMRI signal from different experimental conditions into a "specialization index" (Golarai et al., 2007, Simmons et al., 2007, Joseph et al., 2011. With respect to the statistical comparison approach, Joseph and colleagues offered several definitions of specialization that vary in degree of stringency. For example, the most stringent definition of specialization requires that the category of interest must produce a statistically greater signal than all other experimental conditions (including the baseline condition) and none of the experimental conditions can be different from each other. This is referred to as a "selective" response because only the category of interest activates a particular region or voxel. The least stringent definition of specialization only requires that the category of interest yields a statistically greater signal than the baseline condition and at least one other condition. This is referred to as a "preferential" response because the category of interest is preferred over some other categories and conditions, but the response is not exclusive to the category of interest. Other definitions of specialization yield an intermediate degree of selectivity. Joseph and Gathers (2002) reported that regions like the FFA do not yield a selective response, but instead, are associated with a less stringent profile of specialization. However, we use the term "specialization" in this paper.
We examined the distributional properties of different indices that describe the degree of specialization for faces or objects. For each index described below, we extracted fMRI signal change within regions of interest (ROIs) that were reported by Collins et al. (2012) as facepreferential regions. We used ROIs defined by Collins et al. so that the ROIs could be defined in an independent manner from the way they are defined in the present study. Percent signal change for faces, objects and textures (or raw intensity values for these conditions, as used for one of the indices below) were then extracted within these ROIs and used to create different specialization indices using the present dataset because this dataset has more subjects than the Collins et al.
dataset and includes children so that the distributional properties could be examined in adults and children.
With respect to the specialization index approach, Golarai et al. (2007) used the following formula to define degree of face specialization (FSI, face specialization index): Where F pc and O pc are percent signal change for faces or objects, respectively, relative to baseline.
But Simmons et al. (2007) noted that this particular formula can yield extremely high values if negative values are in the formula so they suggested a correction for negative values as:  In the next FSI, percent signal change relative to fixation baseline is used rather than signal intensity values and the denominator is the range of percent signal change values in a given region for a given subject across all three conditions: We also explored another formula that uses a modification of the Simmons et al.
adjustment and scales the differential response only to the maximum value (rather than the range): An illustrative case is shown in Figure A3. Percent signal change for all three conditions is shown for a single subject in one brain region (a) and in another brain region (b). Inspection of the graphs indicates that face specialization should be greater in region (b) than in region (a).
Although faces yield a greater signal than either objects or textures in region (a), the difference is somewhat modest compared to the profile in (b) where the face signal is even greater than the average of the object and texture signal. However, FSI A indicates the same degree of    With respect to the main effect of category (Table B1), the different measures yield fairly consistent results, with FSI B and FSI A having nearly identical findings. The other two specialization indices detected all of the same category effects detected by FSI B and FSI A except for the right FFA. In addition, these two indices detected category effects in the right IFGorb and medPFC whereas FSI B and FSI A did not.
None of the indices yielded main effects of age in any ROI.
For the Category x Age interactions (Table B2), FT/OT and FSI JGB yield identical results; with FSI B and FSI A detecting interactions in all of the same regions as the other two indices.
However, FSI B and FSI A detected interactions in more ROIs, such as medPFC and IFGorb.
We note that since ROIs were defined in half of the subjects, but the ANOVA was conducted in all subjects, there is a concern of "double dipping" in which analyses are conducted using the same dataset used to define ROIs (Kriegeskorte et al 2009). The reason for including all adult subjects into the primary ROI analysis was to increase statistical power. However, we also conducted the same ANOVAs in the half of the adult sample that was not used to define ROIs. As seen in Tables B1 and B2,