Abstract
The fusiform face area (FFA) has been associated with stimulus (i.e., face)-specific coding rather than with a specific mode of processing (e.g., featural or configural) (Yovel & Kanwisher, 2004) using standard univariate fMRI analysis. Here, we used multivoxel pattern analysis (MVPA) to investigate whether a distributed pattern of activity in cortical face-selective areas reflects distinct modes of face categorization.
Subjects were trained to carry out one of two face-categorization tasks: gender (male vs. female) or race (Caucasian vs. Asian), in alternating blocks of trials. The difficulty of the face categorizations was manipulated by using stimuli consisting of a morphed mixture of Asian males with Caucasian females and Asian females with Caucasian males. MVPA was performed on a set of regions of interest defined by a functional localizer (faces selective regions: fusiform gyrus, inferior occipital cortex and superior temporal sulcus, and a house selective region: parahippocampal gyrus) and by retinotopic mapping (e.g., V1).
A linear classifier (support vector machine) was trained to discriminate the gender or race categorization task using leave-one-run-out cross validation. The average classification accuracy was significantly greater than chance (77%) in the cortical face network whereas that in the PPA is not different from chance (55%). The average classification accuracy in V1 was also greater than chance (65%).
These results suggest that the pattern of activity in face-selective visual processing areas contains information about the modes of processing associated with different categorization tasks. These patterns could reflect top-down attentional biases on sub-population of neurons in these areas that process task-diagnostic facial features or spatial regions of the face.