RT Journal Article
SR Electronic
T1 Ramp coding with population averaging predicts human cortical face-space representations and perception
JF bioRxiv
FD Cold Spring Harbor Laboratory
SP 029603
DO 10.1101/029603
A1 Johan D. Carlin
A1 Nikolaus Kriegeskorte
YR 2015
UL http://biorxiv.org/content/early/2015/10/21/029603.abstract
AB Face space provides a popular metaphor for the representation of individual faces. Although computer-graphics models of face space have a long history, their relationship to the cortical code for individual faces remains unclear. We used such a model to generate animations of faces with realistic 3D shape and texture and analyzed fMRI responses to each individual face. We developed and evaluated multiple neurobiologically plausible computational models of face-space coding, each of which predicts a representational geometry and a regional-mean activation profile. A population code of units with sigmoidal ramp tuning over the face-space dimensions explained both pattern and regional-mean fMRI effects better than alternative models, but only in conjunction with a readout-level population averaging mechanism. This model also accounted for perceptual similarity judgments. Our study demonstrates the importance of modeling readout-level population averaging and provides a computational account of the cortical representation underlying human face processing.