Abstract
Filtering faces to remove all but horizontal information largely preserves face-specific processing. Conversely, preserving only vertical face information leads to dramatic recognition impairments. We conducted an fMRI study to explore where in the visual system the horizontal tuning of face perception emerges. Subjects were scanned while viewing blocks of faces filtered to preserve a 20°-orientation range centered either on horizontal, vertical, or oblique (45° and 135°) orientations. Orientation-filtered faces were presented upright, inverted or (phase-)scrambled. We localized Fusiform Face Area (FFA) and V1 in each subject based on independent functional data. FFA responded most strongly to upright-horizontal faces. Upright versus scrambled and upright versus inverted differences in activation were strongest for horizontal faces. In V1, the average activation profile did not show preference for any orientation. We trained linear support vector machines to classify the stimulus category (upright, inverted, scrambled) or decode the orientation content (horizontal, vertical, left-oblique, right-oblique) based on FFA and V1 voxel activation patterns. In the FFA, the classification of stimulus category was well above chance. Furthermore, upright-horizontal faces (≈75%) were classified better than upright-vertical faces (≈60%). The classification of inverted and scrambled faces showed no orientation preference. In V1, classifying stimulus category was better with horizontal than vertical information; in contrast to FFA, this advantage was present for all stimulus categories. When decoding orientation content based on V1 spatial activation patterns, accuracy was high (55%) for upright and inverted faces, but dropped for scrambled faces (≈27%). In FFA, orientation decoding was slightly above chance for upright faces; it was at chance level for inverted and scrambled faces. These results indicate that FFA is tuned to horizontal orientation selectively for upright faces. The orientation tuning profile of FFA was not passively inherited from V1 processing as this region showed horizontal tuning for all tested categories of faces.
Meeting abstract presented at VSS 2012