Abstract
Although recent research has mapped a network of cortical regions supporting neural representations of facial identity, the nature of the neural code within these regions is largely unknown. Here, we combine neuroimaging (fMRI) and computational investigations to address this issue, both in human adults with normal recognition abilities and in individuals with congenital prosopagnosia (CP). First, we employ multivariate mapping to localize ventral regions able to support individual face discrimination. Then, we use a computational model of face encoding based on independent component analysis, along with high-dimensional regularized regression, to estimate the neural patterns elicited by different faces within these regions. And last, we reverse this procedure in order to reconstruct facial images from relevant patterns of neural activation. Our results show that BOLD patterns related to face processing can be estimated fairly well, particularly in a region of the anterior fusiform gyrus. Moreover, this region is shown to support above-chance facial image reconstruction. Specifically, our assessment of neural-based image reconstructions shows that they are able to support above-chance identity recognition across variation in emotional expression. These results, in participants with normal recognition, also carry overalbeit to a more limited extentto CP participants. The present findings shed light on the nature of high-level visual representations involved in face processing. At the same time, they also open up the possibility of a broad range of applications based on facial image reconstruction.
Meeting abstract presented at VSS 2014