Abstract
Faces can be recognized across a remarkable degree of transformations, including viewpoint, which greatly shifts the position and orientation of facial features. Face identity processing is thought to involve a distributed network of several brain areas including the occipital face area (OFA) and fusiform face area (FFA). FMRI studies in humans and parallel studies in non-human primates, have suggested a hierarchical organization of face viewpoint coding, from orientation dependence through mirror invariance culminating in viewpoint invariance. Here we present data from 12 subjects with a total of 35 face sensitive intracranial electroencephalography (iEEG) electrodes implanted in face sensitive patches of the human ventral visual stream showing that single regions dynamically change their representation over time, in contrast to traditional hierarchical models. Specifically, multivariate classification and representational analyses showed that the early representation (80–180 ms after viewing a face) reflected a mirror symmetric code, with a relatively strong representation for forward facing and away facing faces (90 degree profiles). In the same regions, the representation then shifted to where intermediate face viewpoints (45 degree profiles) were represented strongly from approximately 200–240 ms. Viewpoint dependent/invariant identity coding for faces was then examined by training a classifier to predict identity using one set of face viewpoints and testing it on identities from other viewpoints. The early representation for identity (80–200 ms) was viewpoint dependent and showed some suggestion of mirror invariant identity coding. A viewpoint invariant identity code emerged later (400–600 ms) in many of the same regions that showed viewpoint dependence earlier. Taken together, these results provide support for a model wherein the representation for face viewpoint, and identity as it relates to viewpoint, changes over time with single regions supporting the computation for multiple aspects of face coding over time, likely through network-level interactions with the extended face processing system.
Acknowledgement: NSF 1734907, NIH R01MH107797