Nakwon Rim¹, Marc G. Berman¹, Stefan Uddenberg², Wilma A. Bainbridge¹; ¹University of Chicago, ²University of Illinois Urbana-Champaign

Faces are among the most important visual stimuli to remember when navigating the social world. Previous research has shown that some faces are more reliably remembered or forgotten than others, suggesting that faces have objective memorability. Interestingly, the subjective memorability of faces, or how memorable people "think" a face is, is positively correlated with objective memorability but does not capture the whole variance. This signals that there are both commonalities and dissociations between objective and subjective face memorability. To further explore the relationship between these two notions of face memorability, evaluating them in a common space is crucial. Here, we build models of objective and subjective memorability of faces using the latent dimensions of a Generative Adversarial Network (StyleGAN2) pre-trained on face images. We first collected objective memorability scores from 1,004 synthetic faces generated by StyleGAN2 via a continuous recognition task. As with past research using real-face images, the objective memorability of synthetic faces was reliable across participants. Furthermore, the subjective memorability of faces, acquired via slider ratings made by a separate group of participants, was positively correlated with objective memorability but did not account for the full variance. Building on this, we trained two models predicting objective and subjective memorability from the latent vectors corresponding to the synthetic faces. The two models showed comparable predictive power, measured by cross-validated R-squared, but differed in terms of the weights of the latent vectors. This implies that there are differences between objective and subjective face memorability that we can observe in the common latent face space. Importantly, these models can be also used to transform synthetic faces, allowing us to create a diverse array of face stimuli originating from the same starting point that can be used in further studies of both objective and subjective face memorability.

Acknowledgements: The research was supported by the National Eye Institute (R01-EY034432) to W.A.B.