Yinuo Yang¹, Yuening Yan², Philippe Schyns², Jiayu Zhan¹; ¹Peking University, ²University of Glasgow

Human faces comprise a high-dimensional information space, characterized by rich and complex variations in both 3D shape and complexion. Identifying individual faces thus requires feature-level analysis sufficiently detailed and efficiently versatile to support accurate and rapid recognition. While previous research has identified brain regions selectively involved in face processing (as distinct from those for place or tool processing), the feature-level representations of face-related brain activity remain poorly understood. To address this gap, we asked participants to evaluate the perceptual similarity between randomly generated face identities, each parametrized with 200 shape features using a generative face model. In each trial, two of these faces were randomly paired, each displayed a fixed duration (500 ms) and separated by a blank interval (800-1200 ms); Behavioral responses were allowed only after the second face disappeared. Concurrently, we measured the single trial source-reconstructed MEG activity on 8196 sources throughout the task. To trace the representational dynamics, we computed the mutual information between feature samplings and brain responses of each source on every 4 ms during the first face presentation. As expected, feature representations started early in the occipital region (~80 ms post-stimulus) and propagated along the ventral pathway towards the fusiform gyrus. Upon reaching the fusiform gyrus, the representational dynamics revealed a global-to-local processing pattern: the first peak, around 120 ms post-stimulus, represented the global contour information defining the overall length and width of the face; A second peak, around 240 ms post-stimulus, represented more localized feature information, including the shape details of eyes, nose and cheekbones. Notably, this global-to-local processing pattern was not observed in the occipital region. Our results provide the first detailed account of the dynamic feature representation of 3D faces along the occipito-ventral pathway in the human brain.