Abstract
Face-space models (e.g., Valentine, 1991) have successfully accounted for a range of important phenomena in face recognition. Much of the recent literature has focused on differentiating two possible accounts: Norm-based models, where faces can be encoded with respect to the direction and distance with which they deviate from the central tendency, or norm face (e.g., Rhodes & Jeffery, 2006), and exemplar-based models, where faces are encoded with respect to their distance from all or a subset of previously experienced face exemplars (e.g., Lewis, 2004). High-level visual aftereffects, induced by adaptation from brief exposure to a study face, have provided putative evidence for the norm-based encoding of faces (e.g., Leopold et al., 2001; Susilo, McKone, & Edwards, 2010). Exemplar-based models are purportedly falsified because they are generally believed to make predictions that are qualitatively inconsistent with the observed findings, for example by predicting that aftereffects ought to be centred on the adapting stimulus rather than on the norm. Despite the apparent consensus in the literature regarding norm-based and exemplar-based accounts of these findings, there have been few attempts to explicitly simulate norm-based and exemplar-based models to assess their qualitative and quantitative predictions. Here we explored the theoretical accounts of face adaptation, instantiating two versions of the norm-based model, a traditional norm-based model and a two-pool norm-based model, along with an exemplar-based model. In contrast to the consistent claims in the literature, we found that both the exemplar-based model and the two-pool model, but not the traditional norm-based model, made predictions that were qualitatively and quantitatively consistent with the findings in the literature across a wide range of model parameters.