Stéphane d’Ascoli

Stéphane d’Ascoli

@sdascoli.bsky.social

Research scientist at Meta AI, working on decoding brain signals 🧠

99 followers97 following11 posts

View full thread
Stéphane d’Ascoli
@sdascoli.bsky.social

This was joint work with Jeremy Rapin, Yohann Benchetrit, Teon Brooks, Katelyn Begany, Josephine Raugel, Hubert Banville and Jean-Rémi King. Thanks to Elisa_Cascardi, Diego Marcos, Dominic Giardini, Meta AI, and the open-source & neuroscience communities on which this work builds 🙏

View full thread
Stéphane d’Ascoli
@sdascoli.bsky.social

🗺️ By ablating modalities we can further map how vision, audition & language integrate across cortex. Largest multimodal gains at the temporo-parietal-occipital junction — a known multisensory hub. RGB overlays reveal the topography of sensory convergence.

View full thread
Stéphane d’Ascoli
@sdascoli.bsky.social

💬 Same for language: speech vs. non-speech, emotional vs. physical pain, sentences vs. word-lists — all replicated in-silico. It recovers subtle phenomena: Broca's lateralization, orthography in the VWFA, syntax in STG — without training on these tasks.

Stéphane d’Ascoli
@sdascoli.bsky.social

🔍 What has the model learned? ICA on its weights reveals 5 components mapping onto known functional networks: primary auditory, language, motion, default mode & visual. The brain's functional organization — rediscovered from naturalistic data alone.

View full thread
Stéphane d’Ascoli
@sdascoli.bsky.social

🎯 Without retraining, the predictions of TRIBE v2 are more correlated with group-averaged brain responses than almost any individual fMRI scan! A short finetuning step vastly improves over linear models trained, from scratch, on each individual.

Stéphane d’Ascoli
@sdascoli.bsky.social

👁️ Can we do useful experiments with TRIBE v2? Yes: classic paradigms replicate in-silico! Flash faces, places, bodies, words to the model — TRIBE v2 zero-shot recovers the FFA, PPA, EBA & VWFA, matching decades of empirical results.

Stéphane d’Ascoli
@sdascoli.bsky.social

❓How does it work? Stemming from our v1, which won the Algonauts 2025 challenge, it combines video, audio & language embeddings to predict brain activity for any brain, then adapts to each individual.

Stéphane d’Ascoli
@sdascoli.bsky.social

📈 TRIBE v2 accurately predicts brain activity across cortical & subcortical regions, significantly better than standard linear models. Encoding accuracy follows a log-linear scaling law with data volume — no plateau in sight.

You are up to date.