←
Joséphine Raugel50 posts
Joséphine Raugel avatar

Joséphine Raugel

@JRaugel

PhD student @ Meta & ENS, on Large models and Brains processing language and vision together. jraugel.github.io/

◷ Joined October 2021▧ 15 Photos and videos
Krunoslav Lehman Pavasovic avatar
⟳ Joséphine Raugel retweeted
Krunoslav Lehman Pavasovic@KrunoLehman·Oct 30

WaiT code is out! 🚀 there's plenty of room to push these models further - would love to see people build on it and maybe finally beat latent-space models 👀 github.com/facebookresearch/…

WaiT for the Signal: Simple Frequency-Aware Flow-Matching. We show how to natively incorporate a fundamental property of images directly into diffusion models; setting a new pixel-space SOTA on ImageNet, while reducing compute. 📄 arxiv.org/abs/2607.28760 Full breakdown below👇

1:00
Joséphine Raugel avatar
Joséphine Raugel@JRaugel·Sep 17

The EEG/EMG Foundation Challenge 2026 starts in 4 days 💫 🔗 neural-interfaces26.github.i… Four tracks 🧠 : 🖼️ EEG-to-IMG ➗ BCI decoding 😴 Predicting sleep onset ✋ EMG-to-Pose Let's push the frontiers of neural decoding further 🔥

Alexandr Wang avatar
⟳ Joséphine Raugel retweeted
Alexandr Wang@alexandr_wang·Jun 29

some exciting new work from our AI teams at Meta on non-invasive brain computer interfaces!

We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2. Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. It advances beyond character-level performance to decoding words and semantics, enabling accuracy for overall communication. We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating. 🧵👇

Joséphine Raugel avatar
Joséphine Raugel@JRaugel·Jun 29

Brain2Qwerty V2 is out 💫

🧠⌨️ Decode language from brain activity without surgery. 🧠⌨️ Brain2Qwerty V1 is officially published in Nature Neuroscience. Today, we're releasing Brain2Qwerty V2. We achieve unprecedented performance for a non-invasive MEG setup. Details below 🧵👇

Jean-Rémi King avatar
⟳ Joséphine Raugel retweeted
Jean-Rémi King@JeanRemiKing·Jun 29

We’re happy to announce 2 releases today: - 🧠Brain2qwerty v1 is published at @NatureNeuro - 🚀 Brain2Qwerty v2 is now publicly released Explore how we decode sentences from non-invasive brain recordings: facebookresearch.github.io/b… Links: 📄v1 Nature Neuro: nature.com/articles/s41593-0… 📑v2 Meta preprint: facebookresearch.github.io/b… 💻Code: github.com/facebookresearch/… 📊Data: huggingface.co/datasets/bcbl… 📝Blog: ai.meta.com/blog/brain2qwert… 🧵Thread: x.com/JarodLevy/status/20715…

0:26

🧠⌨️ Decode language from brain activity without surgery. 🧠⌨️ Brain2Qwerty V1 is officially published in Nature Neuroscience. Today, we're releasing Brain2Qwerty V2. We achieve unprecedented performance for a non-invasive MEG setup. Details below 🧵👇

Yohann Benchetrit avatar
⟳ Joséphine Raugel retweeted
Yohann Benchetrit@BenchetritYoha1·Jun 9

We’re very happy to share our latest paper: “Boosting Brain-to-Image Decoding with TRIBE v2 Data Augmentation“ with up to 68% gain in brain-to-image image decoding! 📝arxiv.org/abs/2606.06345 🧵Details in thread below:

Joséphine Raugel avatar
Joséphine Raugel@JRaugel·May 28

11/What does this mean?💡It may suggest that while deepnets and brains share similar representational content (what they see), they likely use different optimization algorithms (how they learn). Understanding this gap is key to building the next generation of energy-efficient AI.

Joséphine Raugel avatar
Joséphine Raugel@JRaugel·May 28

10/ We also tracked this across training. 📈 While forward/backward alignment grows with training, "residual backprop" alignment is non-monotonic: it peaks early (10k iterations) and then subsides as the model converges.

Joséphine Raugel avatar
Joséphine Raugel@JRaugel·May 28

9/ The divergence between the learning brain and the hierarchy of backpropagated gradients is common across all these models.

Post attachment
Joséphine Raugel avatar
Joséphine Raugel@JRaugel·May 28

8/ This finding is robust across architectures from a benchmark of 8 diverse vision models – from supervised CNNs to self-supervised ViTs and vision-language models. For each of them, forward, backprop and residual backprop signals map onto the learning human brain.

Post attachment
Joséphine Raugel avatar
Joséphine Raugel@JRaugel·May 28

7/ The order of these computations also mismatches between AI and the brain. ⏱️ A biological backpropagation should lead the deeper layers to peak after the forward pass. Instead, we observe that gradients peak simultaneously with the forward activations.

Post attachment
Joséphine Raugel avatar
Joséphine Raugel@JRaugel·May 28

6/Regarding forward activations, deeper layers tend to align with higher-level brain areas, and shallow layers with sensory areas. However, backpropagated gradients do not show the same anatomical mapping: sensory areas are best predicted from the gradients of the deepest layers.

Post attachment