Bottom-up and generative computations uniquely explain neural responses across the social brain

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Social evaluations from visual input engage the superior temporal sulcus (STS) and temporoparietal junction (TPJ).
• A prevailing hypothesis suggests a spatial hierarchy: posterior STS (pSTS) for bottom-up social perception and TPJ for top-down mentalizing.
• This computational-neural mapping lacks empirical testing due to a deficit in computational models of social processing.

Purpose of the Study:

• To formally test the computational-neural mapping of social processing in the brain.
• To compare graph neural network (GNN) and inverse-planning models against neural responses during social interaction perception.
• To investigate the roles of pSTS and TPJ in social perception and mentalizing.

Main Methods:

• Developed a GNN model for social interaction recognition based on relational visual information.
• Developed a generative inverse-planning model for social interaction recognition by inverting agent goals and physical world models.
• Collected fMRI data from participants viewing videos of animated shapes depicting social interactions and compared neural responses to both computational models.

Main Results:

• Both the GNN and inverse-planning models explained neural responses in pSTS and TPJ, even when controlling for the other model's variance.
• Exploratory analyses indicated a temporal shift from early perceptual processing to later higher-order reasoning within these regions.
• This suggests a temporal rather than a strict spatial hierarchy in social brain regions.

Conclusions:

• Social perception and mentalizing regions perform a combination of bottom-up and inferential computations, potentially on distinct timescales.
• Challenges the strict division of labor hypothesis in the social brain.
• Provides the first neural comparison of an inverse-planning model and demonstrates the predictive power of cognitive models for fMRI responses to social scenes.