Task-optimized neural networks reveal distinct contributions of specialized and broader visual learning to neural representations of face familiarity

Neural activity patterns within the ventral visual pathway contribute differently to face recognition, with specialized computations in fusiform cortex playing a distinct role compared to broader visual processes, according to a new study. Researchers employed source-resolved magnetoencephalography (MEG) alongside task-optimized neural networks, acting as computational models to examine visual experiences.

Instead of leveraging human observers' expertise, the study systematically altered the learning objective of these models—what the networks were trained to recognize—while maintaining the architecture and loss function constant. By comparing the representational geometries of the learned models with neural responses, researchers aimed to identify the best alignment across various brain regions.

Neural alignment was measured millisecond-resolution, across visual areas V1, lateral occipital cortex (LOC), and fusiform cortex while participants viewed familiar, unfamiliar, and scrambled faces. Across all stimuli, familiar faces exhibited an earlier brain-model alignment in LOC, particularly in the M170 range, an effect most prominent in models trained specifically for face-recognition. In contrast, broader learning objectives led to more variable peak alignment timings in these models.

Fusiform cortex showed stronger alignment around the M200 range, suggesting a fusiform advantage not unique to face-recognition training. Models trained for dual categories (face and object) and object categorization displayed greater correspondence with fusiform representations compared to models solely trained for face recognition.

These findings support the concept of stage-dependent specialization, where training for face-identity recognition influences intermediate-stage timing, while later fusiform representations maintain compatibility with representational structures acquired through broader visual computations.