iiinova — where your digital self keeps evolving

Reposted by @julienbrandoit.bsky.social

Terence Tao @teorth.bsky.social · 11/09/2026

A group of 25 Fields Medalists, including myself, have made a joint declaration on Math and AI: mathandai.org . We welcome additional signatories. See also this article in the Economist announcing the declaration: www.economist.com/science-and-...

mathandai.org

Declaration — Math and AI

Read the declaration and add your name.

432073934

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 23/05/2026

My first journal paper is out in PLOS Computational Biology! 🎉 We do SBI, but not in parameter space. We perform it on interpretable feedback gains, making inference fast, interpretable, and degeneracy-aware. lnkd.in/eZC4CYXK #ComputationalNeuroscience #SBI @neuroenguliege.bsky.social

The figure illustrates the pipeline. The top panel is distributions in the DICs space. The middle panel is the signal input-output reconstruction. The bottom panel is the box plot distributions of conductances.

183

Reposted by @julienbrandoit.bsky.social

Gilles Louppe @glouppe.bsky.social · 22/05/2026

"Accept (spotlight)" at ICML'26 😎 Our paper brings particle filters back to life: autoregressive diffusion models + posterior sampling yield optimal proposals for Bayesian filtering, scaling up to GenCast-sized systems. arxiv.org/abs/2605.20028 w/ Thomas Savary and @francois-rozet.bsky.social

arxiv.org

Training-Free Bayesian Filtering with Generative Emulators

Bayesian filtering is a well-known problem that aims to estimate plausible states of a dynamical system from observations. Among existing approaches to solve this problem, particle filters are theoret...

03911

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

Hybrid architectures combining transient dynamics and multistability outperform both alone, pointing toward a concrete path for scalable, generalizing RL agents. Work with Asad Bakija, Florent De Geeter, Pierre Sacré and Guillaume Drion @ @neuroenguliege.bsky.social

011

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

The good news? Not a dead end. We are developing parallelizable blocks with persistent, multistable memory: Bistable MRUs: arxiv.org/abs/2601.09495 Improving the training reliability (ICML 2026 Spotlight 🎉): arxiv.org/abs/2605.11855

121

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

The bad news? Most parallelizable RNNs, including linear RNNs and SSMs, are monostable by construction. Their memory fades, and fading memory is fundamentally incompatible with THG. The field is betting on these architectures for scalability. We think this blind spot deserves attention.

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

Because the hidden state converges to different attractors depending on history, information is not gradually washed out: it is encoded in which attractor the network settles into, and it stays there.

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

We formalized THG, derived a necessary and sufficient condition for it, and connected it to a key dynamical property of RNNs: multistability, the coexistence of multiple stable equilibria in the network dynamics.

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

Why is this hard? In partially observable environments, agents must remember past observations to act correctly. The longer the horizon, the harder it is to retain that memory.

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

Imagine training a robot on short tasks, then watching it seamlessly handle tasks that are 1000x longer, with no additional training. That's the promise of Temporal Horizon Generalization (THG): a policy that remains optimal for arbitrarily longer horizons.

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/05/2026

📜 New preprint: "On the Importance of Multistability for Horizon Generalization in Reinforcement Learning" arxiv.org/abs/2605.12206 🧵

132

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 13/05/2026

This work is a collaboration with Arthur Fyon, Damien Ernst, and Guillaume Drion within the @neuroenguliege.bsky.social.

010

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 13/05/2026

⚡ Ultra‑low‑power analog hardware: The same trained weights map 1‑to‑1 onto an analog circuit where the physics of the transistors runs the model directly - no microcontroller, no clock. That's how we achieve sub-µW keyword spotting, making intelligent always‑on sensors and implants a real prospect.

111

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 13/05/2026

🧠 Large-scale ML: We make persistent memory reliably trainable in parallelizable RNNs. Persistent memory is present in many dynamical systems - e.g. the brain - but SSMs (including Mamba) will never be able to implement it. We turn persistent memory into a trainable, competitive building block.

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 13/05/2026

Very happy to share that our paper got a Spotlight at #ICML2026! 💡 This work explores two entirely distinct research directions, and we’d be excited to discuss it with anyone interested or curious! 📜 : arxiv.org/abs/2605.118... #AnalogAI #PersistentMemory #ML #UltraLowPower #Multistability

132

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 08/04/2026

But it also raises a real question: if safeguarding global digital infrastructure is this important, why is it being led by private companies with profit motives ? Where are governments and public institutions in building and governing this kind of capability in the public interest ?

000

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 08/04/2026

Project Glasswing from Anthropic and others shows that the AI era will not just happen on its own; we need to actively build the tools and institutions to make it safe and manageable.

100

Reposted by @julienbrandoit.bsky.social

Neuroengineering Lab ULiège @neuroenguliege.bsky.social · 22/03/2026

Three researchers from our lab presented at #COSYNE2026 in Lisbon last week! Posters on calcium-based plasticity and on identifying neuromodulatory targets from spike recordings. Great discussions and some well-deserved sun. #Neuroscience #ComputationalNeuroscience

Julien is explaining the poster to someone. Arthur is on the right.

Gustave is explaining the poster to someone. Julien is on the left.

The three people are enjoying the sunshine in front of a magnificent building in Lisbon.

131

Reposted by @julienbrandoit.bsky.social

Neuroengineering Lab ULiège @neuroenguliege.bsky.social · 22/03/2026

What if biological muscles could inspire a new control theory? Christian Fernandez & Alessio Franci are working on it, guided by Elena Petri (TU/e) 🧠⚙️ Neuromorphic hybrid systems — watch this space. #neuromorphic #controltheory

A photo of the three researchers standing in front of a blackboard covered with equations

031

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 13/03/2026

Neuromodulation moves entire clouds of conductances from one degenerate set to another. We make that shift visible and interpretable. See our poster on Saturday! #COSYNE2026 @neuroenguliege.bsky.social

161

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 04/03/2026

In case it helps anyone: I just put 6+ million conductance-based neuron simulations (STG & DA models) on Zenodo. Took a few days of cluster time to generate. Spike trains with various sampling strategies and noise levels, ready to use. Code and information : shorturl.at/BAALO

030

Reposted by @julienbrandoit.bsky.social

Neuroengineering Lab ULiège @neuroenguliege.bsky.social · 03/03/2026

🦋 We’re now on Bluesky! The Neuroengineering Lab at University of Liège studies intelligence from neurons to hardware. Follow our journey 👇 🌍 www.neuroengineering.uliege.be 🐦 X: x.com/NeuroEngULiege 🟦 LinkedIn: www.linkedin.com/company/neur... #Neuroengineering #Neuromorphic #BrainInspired #AI

123

Reposted by @julienbrandoit.bsky.social

Neuroengineering Lab ULiège @neuroenguliege.bsky.social · 03/03/2026

Great to welcome a local TV crew into our lab! 🎥🧠 We had the pleasure of hosting the LABO4 team for a behind-the-scenes look at our daily research on the mysteries of the human brain. Catch the segment tonight at 19:30 on QU4TRE! 👉 www.news.uliege.be/cms/c_206393...

news.uliege.be

LABO4 | Le cerveau humain

Découvrez comment nos chercheurs explorent le cerveau humain pour comprendre comment il se construit, ce qui le rend unique et comment il inspire les machines de demain.

043

Reposted by @julienbrandoit.bsky.social

Richard Gao @rdgao.bsky.social · 03/03/2026

Once you get past the first 5 minutes of me bumbling around clearly in need of media training, we settle into a nice groove talking about fitting mechanistic models, ML/AI for neuroscience, underconstrainedness vs. degeneracy, and its potential benefits for biology (plus some good chuckles).

2478

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 14/01/2026

We’ve heard a lot about deep nudes and sexual deepfakes made without consent. Yet for weeks now, social media platforms from Meta have been running ads for deepfake apps, available directly on the App Store. I’ve reported at least 5 of them in the past 2 weeks. How did this become normalized?

000

Reposted by @julienbrandoit.bsky.social

Irénée Régnauld @ireneer.bsky.social · 04/01/2026

"Une partie des dizaines de milliers de livres, documents et revues qu’accueillait l’établissement sera jetée à la poubelle" www.huffingtonpost.fr/internationa...

huffingtonpost.fr

La plus grande bibliothèque de la NASA a fermé ses portes, et Trump n’y est pas pour rien

Une partie des dizaines de milliers de livres, documents et revues qu’accueillait l’établissement sera jetée à la poubelle.

45138

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 03/01/2026

Those abuses are real, but they’re not the same story as AI used for discovery, modeling, or understanding nature. We probably need to be much clearer (and push the media to do the same...) about which "AI" we’re talking about, and not let the worst uses define the whole field.

010

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 03/01/2026

At the same time, there are obvious bad uses that are sensational, harmful, and media-friendly (non-consensual sexual image manipulation, scams, etc.).

110

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 03/01/2026

I think part of the problem is that "AI" has collapsed into a single word covering wildly different things. AI for science has already delivered real, measurable progress and will clearly continue to do so.

120

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 29/12/2025

Excited to be at #COSYNE2026! 🧠 We’ll present a poster in #computational #neuroscience on population inference and #neuromodulation plausible targets investigation, extending our NeurIPS workshop work. The COSYNE 2026 abstract is available here: hdl.handle.net/2268/339005

The review process for COSYNE 2026 is now complete. We are pleased to inform you that your submission #11:

"Neuronal population inference reveals how neuromodulation reshapes conductance space"

has been accepted to COSYNE 2026 for a POSTER presentation. Congratulations!

010

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 07/12/2025

Super interesting talk! Perceptions and actions shouldn’t be viewed as opposite or orthogonal concepts, but rather as two sides of the same coin under the Bayesian brain hypothesis.

020

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 07/12/2025

Thanks for this talk @cpehlevan.bsky.social ! The call for a science of "AI for Science" and for caution when interpreting learned mechanisms in a purely data-driven context was particularly insightful.

000

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 07/12/2025

Why can neurons with completely different ion channels fire identically? 🧠 Neuronal degeneracy isn't a bug - it's a feature. Come see how we reconstruct entire degenerate populations from spike times within an interpretable feedback framework. Poster @dataonbrainmind.bsky.social #NeurIPS2025 #DBM

0134

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 23/09/2025

Curious? You can find our extended abstract on ORBi: hdl.handle.net/2268/336182 And check out our recent preprint on the same topic here: - arXiv: arxiv.org/abs/2509.12783 - ORBi: hdl.handle.net/2268/335923 Damien Ernst, Guillaume Drion and Arthur Fyon

000

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 23/09/2025

Thrilled to announce that our work on reconstructing degenerated neuronal populations has been accepted at the NeurIPS 2025 Workshop on Data on the Brain & Mind 🎉 I will be in San Diego in early December to present a poster, and I’m really looking forward to it!

100

julienbrandoit.bsky.social @julienbrandoit.bsky.social · 17/09/2025

🧠✨ Thrilled to announce my FIRST preprint is now live! ✨🧠 Sharing my first research paper on arXiv, tackling a challenge in computational neuroscience. 📄 arXiv preprint: arxiv.org/abs/2509.12783 📚 ORBi (with supplementary materials): hdl.handle.net/2268/335923

A schematics of the computational approach we used.

030