Hidekazu Nagamura¹, Wilma A. Bainbridge², Shunichi Kasahara^1,3; ¹Sony Computer Science Laboratories, Inc., ²University of Chicago, ³Okinawa Institute of Science and Technology Graduate University

Understanding the factors that determine face memory is crucial for elucidating the encoding mechanisms of face perception. Previous research suggests that face memory is driven by both "intrinsic memorability," which is consistent across observers, and "self-relevance," which depends on observer attributes such as ethnicity (e.g., the other race effect; Meissner & Brigham, 2001). However, the relative contributions of these factors remain unclear, leaving optimal intervention strategies undefined regarding whether to prioritize universally memorable features or to personalize stimuli for individuals. To answer this question, we quantitatively compared the contributions of intrinsic memorability and self-relevance. We generated a balanced face dataset (160 images: 4 ethnicities × 2 genders × 20 images) based on features estimated by the DeepFace neural network model to minimize confounds. We then measured memory performance using a continuous recognition task with a demographically balanced participant group (N = 120: 4 ethnicities × 2 genders × 15 participants). Participants viewed a continuous stream of faces (750 ms duration) and identified repeated images appearing after an average interval of approximately 1 minute. Results showed that while gender congruence between participant and image significantly improved memory performance, ethnicity congruence did not. To quantify the contribution of each factor, we employed a Bayesian mixed-effects logistic regression analysis. This approach allowed us to decompose the variance in memory performance into image-specific effects (intrinsic memorability), participant-image interaction effects (self-relevance), and participant-specific effects (individual memory ability). The impact of intrinsic memorability on memory performance was approximately three times greater than that of self-relevance and comparable to individual differences in participants' memory ability. These findings demonstrate that intrinsic image features are the dominant factor in face memory, outweighing the contribution of self-relevant demographic cues. This suggests that leveraging intrinsic visual information offers a more potent approach to memory augmentation than relying solely on demographic matching.

Acknowledgements: This work was supported by JST, PRESTO Grant Number JPMJPR23I4.