“Computational Single-Neuron Mechanisms of Visual Coding and Attention in the Primate Brain”

Shuo Wang, PhD
Associate Professor of Radiology, Neurosurgery and Biomedical Engineering
WashU

How does the human brain encode, differentiate and attend to faces? In this talk, I will address four core questions: (1) What are the neural computational mechanisms for general face and object encoding? (2) How do neural representations of faces change with learning? (3) What are the network mechanisms underlying visual attention? (4) What are the neural population dynamics underlying visual attention? Drawing on human single-neuron recordings and advanced computational analyses, I will present evidence for a novel region-based feature coding mechanism in the medial temporal lobe (MTL). Specifically, sparse-coding neurons in the MTL encode visually related faces and objects, whereas feature neurons exhibit localized tuning that remains stable across changes in identity and familiarity. These findings point to a coding scheme grounded in visual feature space rather than conceptual identity. I will then describe a computational pathway by which face and object representations evolve from dense, feature-based encodings in the ventral temporal cortex (VTC) to sparse, semantic-based codes in the MTL. This transformation supports robust visual discrimination and forms the neural foundation for memory and recognition. Learning further sharpens these representations: Repeated exposure increases representational distance between similar faces in the MTL, reflecting experience-dependent refinement that enhances familiar face discrimination. Next, I will discuss how attention dynamically modulates neural activity across distributed cortical networks. Attentional engagement enhances MTL responses to visual targets and is associated with gamma-band synchronization between the MTL and medial frontal cortex (MFC). Distinct synchronization patterns between the MTL and dorsal anterior cingulate cortex (dACC) support different phases of visual search, working memory and decision-making. MTL-MFC synchronization also shapes the geometry of neural representations, linking attention to representational dynamics. Finally, I will highlight neural population dynamics that support goal-directed visual search. Large-scale recordings from V4, IT, OFC and LPFC revealed distinct populations encoding attention and category information. Population activity maintained cue representations across delays, differentiated categories during search and predicted search efficiency. An orthogonal subspace provided a latent structure for sustaining task-relevant information, while foveal attention enhanced peripheral representations through context-dependent changes in representational geometry. Together, these dynamics flexibly coordinated attention, memory and category representations across search stages. Together, these findings reveal a unified neural framework for how faces and objects are encoded, refined through experience and prioritized through attention, and how distributed population dynamics orchestrate goal-directed visual search — illuminating the computational architecture of social visual cognition in the human and non-human primate brain.