Facial expression discrimination emerges from partially overlapping neural subspaces of detection and identity

How does the human brain decode facial expressions? Addressing this question requires models that tightly link neural representations to behavior. Here, we first establish rhesus macaques as an animal model of human facial expression discrimination, showing strong behavioral correspondence across six expressions. To probe underlying neural mechanisms, we leverage artificial neural network (ANN) models of the ventral stream as explicit computational hypotheses explaining macaque behavior. Despite action unit-based stimulus generation and high classification accuracies of face-optimized models, broadly-trained ANNs best match macaque image-level behavior. Interestingly, ANNs showing the strongest behavioral alignment are those whose internal representations best resemble macaque inferior temporal (IT) representational geometry. Neural recordings further reveal that early IT responses best predict our measured expression judgments and exhibit a heterogeneous coding architecture in which identity and expression information coexist along partially overlapping dimensions. Together, these findings constrain mechanistic accounts of facial expression discrimination in high-level visual cortex.