Logan Trujillo¹ (); ¹Texas State University

Previous computational modeling research investigating the mental representation of faces suggests that the positive relationship consistently observed between the perceived typicality and attractiveness of young adult Caucasian female faces may be best explained by a representational format where faces are encoded in memory with respect to a hypothetically perfect representative of a face population (ideal face). This study extends this prior research by modeling the mental representation of young adult Caucasian, African American, and East Asian female faces. Human judgments of typicality, attractiveness, and pairwise similarity were gathered for 17 faces from each face group. Perceived typicality and attractiveness judgments were then predicted using four computational models expressed within a perceptual embedding space (“face space”) derived from the similarity ratings via multidimensional scaling: the Ideal Dimension Model (IDM), in which typicality and attractiveness reflect the similarity of a face to the ideal face of a face population; the Generalized Context Model (GCM), in which typicality and attractiveness reflect the similarity of a face to other faces encoded in memory; the Central Prototype Model (CPM), in which typicality and attractiveness reflect the similarity of a face to the central tendency of a population of faces; and the Varying Abstraction Model (VAM), in which typicality and attractiveness reflect the similarity of a face to clusters of related faces (“subprototypes”) within a face population. Face typicality for all three face groups was best explained by the IDM as quantified by Bayesian metrics of model generalizability. The attractiveness of the Caucasian and African American faces was also best explained by the IDM, whereas the VAM best explained East Asian facial attractiveness. These findings suggest that facial typicality is mentally represented by humans via an ideal format; however, facial attractiveness may be mentally represented using different formats depending on the face population.

Acknowledgements: This work was partially supported by NIH grant #HD021332.