Learned foveal representations, not gaze-mediated information access, underlie perceptual advantages for canonical face configurations.

Altering face feature positioning significantly impairs the recognition of identity, gender, and emotional expression. We evaluated the possible cause of these impairments in terms of three distinct mechanisms: (1) peripheral information loss due to foveated vision when diagnostic features fall distant from fixation, (2) suboptimal adaptation of eye-movement strategies to altered configurations, or (3) mismatch between the incoming face features and learned face templates tuned to upright configuration. In this study, we combined computational modeling and psychophysics to isolate the contributions of these factors to the degradation of perceptual judgments of scrambled dynamic emotional faces. Our results indicate that, although the relative positioning of facial features with respect to observer fixations influences performance through both peripheral information loss and suboptimal fixation strategies, these factors together cannot fully account for the observed performance decrements. Using convolutional neural network models, we show that the use of a learned internal face representations based on the long-term use of a consistent eye movement strategy for viewing upright faces, combined with a foveated visual system, can best explain the performance decrements for altered face configurations. These findings enhance our understanding of how peripheral visual constraints, oculomotor strategies, and learned high-level representations tuned to the statistical distribution of visual information contribute to configuration-dependent face perception.