🌙 Lune Bellec

Program - MAIN educational 2025 Website of the educational workshop organized during the Montreal Artificial Intelligence and Neuroscience Conference 2025

Attending the Montreal AI and Neuroscience (MAIN) Conference this week? #MontAIN2025
We have put together some exciting educational workshops on cognitive benchmarking large models, RL and video games and dynamical systems! More info and registration here: main-educational.github.io/program/

Registration closes tonight!

MAIN educational 2025 Website of the educational workshop organized during the Montreal Artificial Intelligence and Neuroscience Conference 2025

Interested in getting hands on with methods at the intersection of AI and neuroscience? Register to the Educational Workshop of the Montreal AI and Neuroscience (MAIN) Conference 2025! main-educational.github.io

I've created a @gatsbyucl.bsky.social starter pack!

Let me know if you’d like to be included, or just jump in to see what we're talking about.

Either way, a retweet would be greatly appreciated! 🚀
go.bsky.app/4g6Ro4U go.bsky.app/AErHDon

From memories to maps: Mechanisms of in context reinforcement learning in transformers Humans and animals show remarkable learning efficiency, adapting to new environments with minimal experience. This capability is not well captured by standard reinforcement learning algorithms that re...

Humans and animals can rapidly learn in new environments. What computations support this? We study the mechanisms of in-context reinforcement learning in transformers, and propose how episodic memory can support rapid learning. Work w/ @kanakarajanphd.bsky.social : arxiv.org/abs/2506.19686

When an audio AI Model Play 🧠 Dress-Up This really is about fine-tuning an audio artificial network to align representations with human brain data, and seeing what happened next

CNeuroMod is now on substack, and our first post highlights a new study showing that a tiny 2M-parameter audio model can be meaningfully fine-tuned on an individual brain with benefits for downstream AI tasks. open.substack.com/pub/cneuromo...

very much looking forward to try this.

How much are MPs entitled to know about research grants? Not as much as they think A parliamentary committee is asking for academics’ private information on a strange anti-DEI crusade

a gross parliamentary overreach — article by @picardonhealth.bsky.social in the globe and mail; #academicsky #neuroskyence #psychscisky www.theglobeandmail.com/gift/91587b9...

The 2nd CogBases Workshop is this 4 & 5 Nov at Institut Pasteur!
We'll discuss the latest in open science methods for analysing brain imaging data. Registration free, but mandatory
neuroanatomy.github.io/cogbases-2025/
@k4tj4.bsky.social @cmaumet.bsky.social @bthirion.bsky.social @demw.bsky.social

OpenNeuro @openneuro.bsky.social just hit a huge milestone: 1500 datasets! Congrats to the team on making this project so successful over the last 7 years.

The Biological Psychiatry family of journals is now officially on Bluesky!

Follow us for the latest research in psychiatric neuroscience, cognitive neuroimaging, and global open science from our three leading journals.

And to match their spirit of openness, we’ve released the code, containers, and data. Anyone can rerun the entire analysis.

Co-lead authors: @clarken.bsky.social and @surchs.bsky.social
Paper: doi.org/10.1093/giga...
Github: github.com/SIMEXP/autis...
Zenodo archive: doi.org/10.5281/zeno...

End/🧵

The signature was discovered in a balanced cohort of ~1,000 individuals and replicated in an independent sample (thanks to the ABIDE I & II wonderful participants and the researchers who shared their data 💜💜💜). 5/🧵

Scatterplot showing individual risk (positive predictive value) versus prevalence in the general population for different autism risk markers. Rare monogenic syndromes (green diamonds) confer very high risk but are extremely rare; common genetic variants (yellow triangles) are widespread but confer very low risk; copy number variants (pink triangles) sit in between. Previous imaging-based models (red dots) achieve modest risk. The new High-Risk Signature (orange circle) replicates across datasets, confers a sevenfold increased risk of autism, and is present in about 1 in 200 people.

A positive result means someone is about seven times more likely to actually have an autism diagnosis. This rivals the best imaging markers, while still being found in about 1 in 200 people in the general population. 4/🧵

Diagram comparing how different autism risk markers identify individuals. Each circle represents the overlap between people labeled by a marker (grey), people with autism (purple), and those labeled who actually have autism (blue). Monogenic syndromes label very few people but with high accuracy; existing imaging models label many people but with low accuracy; the High-Risk Signature (HRS) approach identifies a small subset with a higher proportion of true autism cases.

We turned the problem on its head. Instead of trying to classify everyone, we built a brain signature that only makes predictions when it’s confident. 3/🧵

Real life isn’t balanced. Autism affects about 1% of the population. In that setting, a biomarker with 80% balanced accuracy would catch one true case for every twenty false alarms. 2/🧵

Many brain imaging “biomarkers” for autism have been proposed. Most aim for balanced accuracy (matching sensivity/specificity) on datasets where cases and controls are split 50/50. 1/🧵

"A murder of butterflies" has a nice ring to it.

This year at #CCN25 we showed the importance of OOD evaluation to adjudicate between brain models. Our results demonstrate these trivial but key facts :
- high encoding accuracy ≠ functional convergence
- human brain ≠ NES console ≠ 4-layers CNN
- videogames are cool

w/ @lune-bellec.bsky.social 🙌

Mapping cerebral blood perfusion and its links to multi-scale brain organization across the human lifespan | doi.org/10.1371/jour...

How does blood perfusion map onto canonical features of brain structure and function? @asafarahani.bsky.social investigates @plosbiology.org ⤵️

Poster titled "Neuromod: The Courtois Project on Neuronal Modelling" with logos from Université de Montréal and the Centre de recherche de l'Institut universitaire de gériatrie de Montréal. Large bold text reads: 6 BRAINS – 987H-fMRI – 18 TASKS Followed by the subtitle: Naturalistic & Controlled – Multimodal / Perception + Action Each letter in "18 TASKS" contains thumbnails from various visual tasks. The central table summarizes 32 datasets grouped by primary domain (Vision, Audition, Language, Memory, Action, Other). For each dataset, the table indicates which stimulus modalities were used (Vision, Speech, Audio, Motion), what responses were collected (Physiology, Eye tracking, Explanations, Actions), and how many sessions and subjects were scanned. The overall visual style is playful and bold, with rainbow colors for modality types and rich iconography indicating data types.

In 2019, the CNeuroMod team and 6 participants began a massive data collection journey: twice-weekly MRI scans for most of 5 years. Data collection is now complete! 1/🧵

Automated testing with GitHub Actions Better Code, Better Science: Chapter 4, Part 7

Automated testing with GitHub Actions - the latest in my Better Code, Better Science series russpoldrack.substack.com/p/automated-...

I find AI coding most useful to comment / suggest on what I do. Your disastrous experience with code generation matches mine. But as a side kick it's incredibly positive IMO.

🥁... we are SO happy to officially announce that registration is now OPEN for our OHBM Virtual Satellite Meeting, taking place September 10-12!

This has been a major goal of the SEA-SIG for a while now and we're so excited to show you what we've been working on!

🌱🌎✨🧠

four brain maps showing noise ceiling estimates in response to image presentation

New CNeuroMod-THINGS open-access fMRI dataset: 4 participants · ~4 000 images (720 categories) each shown 3× (12k trials per subject)· individual functional localizers & NSD-inspired QC . Preprint: arxiv.org/abs/2507.09024 Congrats Marie St-Laurent and @martinhebart.bsky.social !!

Precision functional mapping reveals less inter-individual variability in the child vs. adult human brain Human brain organization shares a common underlying structure, though recent studies have shown that features of this organization also differ significantly across individual adults. Understanding the...

1/11 Very excited to say that our preprint, Precision functional mapping reveals less inter-individual variability in the child vs. adult human brain, is up on biorxiv!
www.biorxiv.org/content/10.1...

Google’s Gemini 2.5 paper has 3295 authors

arxiv.org/abs/2507.06261

Advancing neural decoding with deep learning - Nature Computational Science A recent study introduces a neural code conversion method that aligns brain activity across individuals without shared stimuli, using deep neural network-derived features to match stimulus content.

Excited to share our News&Views on Kamitani Lab's NatComputSci paper! Their neural code converter enables transformation of brain activity patterns across individuals, and it doesn't need shared stimuli or connectivity information!
www.nature.com/articles/s43...

Excited to co-organize our NeurIPS 2025 workshop on Foundation Models for the Brain and Body!
We welcome work across ML, neuroscience, and biosignals — from new approaches to large-scale models. Submit your paper or demo! 🧠 🧪 🦾

This is why I think the platonic rep hypothesis doesn’t apply to brain-ANN alignment, since most existing (functional?) models are implicitly or explicitly trained to mimic humans.

The assumption of PRH is that the networks are trained independently which doesn’t hold in brain-ANN comparisons.