Shapley (MSA) Insights:
Feature attribution maps align with neuroanatomy, but crazily enough, textual features from the transcripts are the most predictive of all.

We have strong lifts over baseline both in-distribution and OOD. VIBE (final): r=0.3225 (ID, Friends S07) & 0.2125 (6×OOD).
Competition submission (earlier iteration): r=0.3198 ID / 0.2096 OOD → 1st in Phase-1, 2nd overall.
vs baseline 0.2033/0.0895 → +0.119/+0.123.

We present VIBE: Video-Input Brain Encoder, which is a 2-stage Transformer. First stage merges text, audio, and visual features per timepoint (plus subject embeddings). And the second stage models temporal dynamics with rotary positional embeddings.

Competition & Data: Algonauts 2025 tests how well we can predict brain activity while people watch naturalistic movies. Multi-modal stimuli (video, audio, text) → whole-brain fMRI, split into parcels. Train on movies/TV; evaluate in-distribution and on out-of-distribution films.

We did it! 🏆 We won Phase 1 and placed 2nd overall in the Algonauts 2025 Challenge. So proud of the crew
@keckjanis.bsky.social,Viktor Studenyak,Daniel Schad,Aleksandr Shpilevoi. Huge thanks to @andrejbicanski.bsky.social and @doellerlab.bsky.social for support. Report: arxiv.org/abs/2507.17958

DCGAN Case Study:
Pixel-wise Shapley Modes reveal the inverted CNN hierarchy: first transposed-conv layer shapes high-level facial parts; final layer merely renders RGB channels.

LLM Case Study:
Calculated expert-level contributions of an MOE-based LLM across arithmetic, language ID, and factual recall. Found an expert which was super-important for all domains. Also found redundant experts, removing which does not decrease performance much.

MLP Case Study:
Neural computations within a three-layer MNIST MLP were analysed. L1/L2 regularisation funnels computations into a few neurons. Also, contrary to popular belief, large weights do not equal high importance of neural units.