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

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Latent-fMRI: Latent Neural Representations in Naturalistic fMRI using Contrastive Learning

Kajal Singla¹, Prathamesh Dinesh Joshi², Felix Filius¹, Nico Scherf¹; ¹Max Planck Institute for Human Cognitive and Brain Sciences, ²Vizuara AI Labs

Presenter: Kajal Singla

Functional Magnetic Resonance Imaging (fMRI) studies of naturalistic, within-subject analyses commonly rely on event segmentation methods such as Hidden Markov Models (HMM) and pattern-based approaches like Multivoxel Pattern Analysis (MVPA) and Representational Similarity Analysis (RSA). However, these methods often struggle to capture key aspects of neural dynamics, including geometric structure, trajectories between brain states, non-linear representations, temporal continuity, and noise robustness. To address these limitations, we explore Consistent Embeddings of High-Dimensional Recordings using Auxiliary Variables (CEBRA), a contrastive learning based framework, for within-subject modeling of Blood Oxygen Level Dependent signals. In this study, we reanalyzed an audio-based naturalistic fMRI dataset using CEBRA. Our objectives were to estimate an optimal temporal offset for naturalistic fMRI data and to investigate the resulting latent representations in a 3D space. We then annotated these representations with narrative labels to assess whether the learned embeddings align with meaningful story events. Overall, CEBRA recovered temporally coherent latent structure, but the current analyses do not demonstrate that this structure encodes narrative content beyond intrinsic BOLD autocorrelation.

Topic Area: Methods, Tools, Theory & Neural Coding