| dc.description.abstract | Object recognition, a fundamental function of the brain’s visual system, has been ex-
tensively studied, yet the precise neural mechanisms underlying it remain only par-
tially understood. This study aims to deepen our understanding of object-selective
responses in the human brain, focusing on faces as highly salient stimuli. Faces pro-
vide a uniquely powerful lens for investigating how complex object representations
emerge. Using high-resolution fMRI data from the Natural Scenes Dataset (NSD), an
analysis pipeline identified face-selective regions and examined their feature encod-
ing.
Key findings showed that the face-selective middle temporal sulcus (+MTS/OFA),
face-suppressive fusiform gyrus (–FFG), and face-selective fusiform gyrus (+FFG)
exhibited strong trial-repeatability and encoded low-level features like face position
and size. High-level features (gender, age) were not consistently encoded, possibly
due to NSD’s passive viewing and limited statistical power. Representational sim-
ilarity analysis (RSA) on DNNs (GenderAge, RetinaFace) revealed hierarchical en-
coding, with low-level features dominating early layers and task-specific high-level
features emerging deeper, reflecting training objectives. DNN representations were
deterministic, contrasting with fMRI data’s inherent noise.
Despite limitations (NSD stimulus constraints, passive task, manual ROI definition),
this research offers insights into early face representation, emphasizing spatial con-
figuration. By comparing human fMRI patterns with computational models, the
study highlights fundamental differences in noise, task specificity, and feature pro-
gression, providing a versatile pipeline for future semantic face encoding investiga-
tions. | |
