A computational model of dysfunctional facial encoding in congenital prosopagnosia

Congenital prosopagnosia is a selective deficit in face identification that is present from birth. Previously, behavioral deficits in face recognition and differences in the neuroanatomical structure and functional activation of face processing areas have been documented mostly in separate studies. Here, we propose a neural network model of congenital prosopagnosia which relates behavioral and neuropsychological studies of prosopagnosia to theoretical models of information processing. In this study we trained a neural network with two different algorithms to represent face images. First, we introduced a predisposition towards a decreased network connectivity implemented as a temporal independent component analysis (ICA). This predisposition induced a featural representation of faces in terms of isolated face parts. Second, we trained the network for optimal information encoding using spatial ICA, which led to holistic representations of faces. The network model was then tested empirically in an experiment with ten prosopagnosic and twenty age-matched controls. Participants had to discriminate between faces that were changed either according to the prosopagnosic model of featural representation or to the control model of holistic representation. Compared to controls prosopagnosic participants were impaired only in discriminating holistic changes of faces but showed no impairment in detecting featural changes. In summary, the proposed model presents an empirically testable account of congenital prosopagnosia that links the critical features--a lack of holistic processing at the computational level and a sparse structural connectivity at the implementation level. More generally, our results point to structural differences in the network connectivity as the cause of the face processing deficit in congenital prosopagnosia.