Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Model-based Characterization of the Lateral Visual Stream

Yingtian Tang¹, Ming Zhou², Martin Schrimpf¹, Leyla Isik²; ¹EPFL - EPF Lausanne, ²Johns Hopkins University

Presenter: Yingtian Tang

While neural responses to static scenes have been characterized extensively, much less is known about dynamic visual processing. Here, we use a state-of-the-art brain-encoding model based on V-JEPA 2 as a digital twin to infer the preferred stimuli along the lateral visual stream. Screening over 700k naturalistic videos reveals a progression from low-level features to body motion, physical interactions, and language-related content. These findings are further supported by a searchlight-based representational similarity analysis. Since the co-occurrence of low- and high-level features can obscure region-specific representations, we disentangle them via a novel modulation objective that maximizes responses in a target region while suppressing others. These results reveal distinct features such as communicative signals in pSTS beyond motion and language-related content in aSTS. Taken together, we develop a scalable framework for probing dynamic visual processing which can be applied to model-based synthesis for neural modulation.

Topic Area: Computational Models of Vision & Visual Cortex