English

VIBE: Video-Input Brain Encoder for fMRI Response Modeling

Machine Learning 2025-07-28 v2 Artificial Intelligence Computer Vision and Pattern Recognition

Abstract

We present VIBE, a two-stage Transformer that fuses multi-modal video, audio, and text features to predict fMRI activity. Representations from open-source models (Qwen2.5, BEATs, Whisper, SlowFast, V-JEPA) are merged by a modality-fusion transformer and temporally decoded by a prediction transformer with rotary embeddings. Trained on 65 hours of movie data from the CNeuroMod dataset and ensembled across 20 seeds, VIBE attains mean parcel-wise Pearson correlations of 0.3225 on in-distribution Friends S07 and 0.2125 on six out-of-distribution films. An earlier iteration of the same architecture obtained 0.3198 and 0.2096, respectively, winning Phase-1 and placing second overall in the Algonauts 2025 Challenge.

Cite

@article{arxiv.2507.17958,
  title  = {VIBE: Video-Input Brain Encoder for fMRI Response Modeling},
  author = {Daniel Carlström Schad and Shrey Dixit and Janis Keck and Viktor Studenyak and Aleksandr Shpilevoi and Andrej Bicanski},
  journal= {arXiv preprint arXiv:2507.17958},
  year   = {2025}
}
R2 v1 2026-07-01T04:16:09.254Z