Poster Presentation

Contributed Talk Session: Thursday, August 6, 1:45 – 2:45 pm, Skirball Theater
Poster B43 in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Self-Supervised Modelling of Social Primitives in Human Dyads

Thibaut Chataing¹, Thomas Maillart², Nada Kojovic², Giuseppe Chindemi², Sara Seyed Akhavan¹, Camilla Bellone¹, Marie Schaer¹; ¹University of Geneva, ²ETH Zurich

Presenter: Thibaut Chataing

Social interactions rely on dyadic primitives that are central to brain social perception. Yet they have rarely been explicitly modeled to computationally. We introduce HumanLISBET, a self-supervised pose model that encodes clinician-child dyads. Using 65 hours of unlabeled interactions for training, we evaluate the model on 119 children across 19 clinical targets (ADOS-2, Vineland-II). We evaluate HumanLISBET against several benchmarks, including a kinematic null model. Our model proves competitive for autism diagnosis (AUROC=0.77) and significantly outperforms baselines on relational Vineland-II subscales (Personal, Community), though it underperforms on ADOS-2 severity and the non-relational Domestic subscale. This selective pattern suggests that the embedding captures relational social dynamics, while session-level aggregation discards part of the temporal structure needed for social affect.

Topic Area: Development, Individual Differences & Clinical Populations

Extended Abstract: Full Text PDF Spotlight Presentation