FROMANNUAL REVIEWS

CogSci 2025

•

August 02, 2025

•

San Francisco, United States

keywords:

cognitive neuroscience

social cognition

fmri

artificial intelligence

neural networks

Understanding social scenes depends on tracking relational visual information, which is prioritized behaviorally and represented in the superior temporal sulcus (STS), a region involved in processing social scenes. Despite its importance, relational information has been underutilized in computational models of social vision. In this study, we evaluate two neural network models—SocialGNN and RNN Edge—that explicitly incorporate relational cues, and compare their performance to state-of-the-art (SOTA) AI vision models. SocialGNN utilizes a graph neural network to organize input information about each video frame into a graph structure with nodes representing faces and key objects, and edges encoding relational information such as gaze direction and physical contact. RNN Edge is an even simpler model that processes only relational information without node features or graph-based structures. These models were tested on behavioral and neural data from 3-second natural videos of two people engaged in everyday activities, as well as on the PHASE dataset, a collection of 2D animations depicting agent-object interactions inspired by Heider and Simmel. Across both datasets, SocialGNN and RNN Edge achieved strong performance in predicting human behavioral ratings of social interactions and were comparable to SOTA AI models in behavioral encoding tasks, despite being trained on significantly less data and with simpler architectures. Notably, the success of RNN Edge suggests that additional visual features and the graph-based framework of SocialGNN do not significantly enhance performance, underscoring the primacy of gaze and physical contact as essential relational cues. These findings emphasize the importance of integrating relational information into computational models to develop better models of social perception and human-aligned AI.

Downloads

PaperTranscript English (automatic)

Next from CogSci 2025

Heterogeneity in Loss Aversion Estimates across Modeling Approaches
technical paper

Heterogeneity in Loss Aversion Estimates across Modeling Approaches

CogSci 2025

Akhil Abburu and 2 other authors

31 July 2025

Similar lecture

Graph-Conditional Flow Matching for Relational Data Generation
technical paper

Graph-Conditional Flow Matching for Relational Data Generation

AAAI 2026

Davide Scassola
Luca Bortolussi and 2 other authors

23 January 2026

Stay up to date with the latest Underline news!

Select topic of interest (you can select more than one)

PRESENTATIONS

  • All Presentations
  • For Librarians
  • Resource Center
  • Free Trial
Underline Science, Inc.
1216 Broadway, 2nd Floor, New York, NY 10001, USA

© 2026 Underline - All rights reserved

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.