Franck Iutzeler

I love the #eurips initiative! But to live up to its potential, it should be accepted an official #neurips conference location (similar to Mexico City) and not just an addon! Then we would save C02, rather than adding to it!!! What can we do to achieve this?

Paper ( #ICML 2025 ) arxiv.org/abs/2503.16398

See also our previous work the invariant distribution of SGD ( #ICML 2024 ) arxiv.org/abs/2406.09241

🔧 Technical approach: Building on our previous paper on the invariant measures of SGD, we leverage large deviations theory (Freidlin-Wentzell) to obtain matching upper/lower bounds on the hitting time to a neighborhood of the global minima.

📊 Here's how it looks on a standard non-convex example. We plot the log of the global convergence time as a function of the step-size. Our theory predicts a linear relationship and the slope of this line (red). R² = 0.97 for theory vs 0.99 for data fit ✔️

🔗 Our quantity E captures the interplay between geometry and SGD convergence, enabling transfer from loss landscape literature to convergence guarantees.

💡 Key consequence: Our results explain how the favorable properties of the landscape of NNs drive SGD's convergence

🌄 The key quantity E is a geometric measure of what makes optimization hard, eg:
• When f has no spurious minima → E = 0 → sub-exponential convergence
• Even more: E can be controlled by the depth of spurious minima - shallow minima → small E → fast convergence 🚀

🔑 Main result: we find that the mean time τ to hit a global minimum starting from x is

𝔼[τ] ≈ exp(E(x) / η)

where:
• η = (constant) step size of SGD
• E(x) = energy-like function that captures the geometry of the non-convex landscape and the statistics of the noise

🎯 The challenge: Most SGD theory gives local guarantees or requires special structure (overparameterization, specific models, etc).

We ask: for ANY smooth non-convex function, how much time does SGD need to reach the global optimum (escaping all spurious minima along the way)?

❓ How long does SGD take to reach the global minimum on non-convex functions?
With W. Azizian, J. Malick, P. Mertikopoulos, we tackle this fundamental question in our new ICML 2025 paper: "The Global Convergence Time of Stochastic Gradient Descent in Non-Convex Landscapes"

📣 New preprint 📣

**Differentiable Generalized Sliced Wasserstein Plans**

w/
L. Chapel
@rtavenar.bsky.social

We propose a Generalized Sliced Wasserstein method that provides an approximated transport plan and which admits a differentiable approximation.

arxiv.org/abs/2505.22049 1/5

lets using fucking git plz

Attention #JuliaLang users & scientific computing fans: #JuliaCon Local Paris 2025 is drawing closer 😱
- Talk submissions are still open (for 1️⃣ more week!)
- Tickets are now available (at early 🐦 prices)
Tell your colleagues, your neighbors, your grandma 👵 details at juliacon.org/local/paris2...

Épisode 5 de notre série "Les nouveaux visages de l’Académie des sciences" : La statistique, une science en son temps.

Retrouvez la vidéo sur Youtube : www.youtube.com/watch?v=2AhT...

The PAISS summer school is back with an incredible line of speakers (and more to come). Spread the word !

The logo for JuliaCon Local Paris 2025. Pretty cool tbh. Design by cormullion.

Fellow #JuliaLang enthusiasts, here's your chance to shine: the call for proposals of JuliaCon Local Paris 2025 is finally open. Send us your coolest, nerdiest contributions and we'll prepare a scientific program for the ages!
pretalx.com/juliacon-loc...

Mathématiques : cette spécialité nationale que la France doit populariser Le Centre national de la recherche scientifique ouvre une consultation pour recueillir l'opinion du grand public sur les mathématiques. Il s'agit d'améliorer la perception de cette matière si utile au...

Algorithme, intelligence artificielle… les #mathématiques sont plus que jamais au cœur des technologies. Mais cette science devient un sujet d’inquiétude pour les instances françaises : une illustration supplémentaire du french paradox.
📻 À écouter sur @franceinfo.fr bit.ly/41F9hCC
@cnrs.fr

The 2025 Barkley Marathons is over. There are no finishers. #BM100

🎓 Concours Normalien Étudiant

Le Concours Normalien Étudiant : une opportunité unique d’accéder à une formation d’excellence, de bénéficier d’un environnement intellectuel stimulant et de rejoindre une communauté engagée.

🎥 Toutes les infos sur la page de notre site : bit.ly/3YjeQTz

Fellow #JuliaLang enthusiasts in Europe, I've got 4 words for you.

JuliaCon Local Paris 2025

Yes, it's happening. Like the global @juliacon.bsky.social event, but with more croissants. Stay tuned!

Claire Vernade - European career opportunities European Academic Career Opportunities in 2025

I’ve put together a short list of opportunities for early career academics willing to come to Europe: www.cvernade.com/miscellaneou...

This mostly covers France and Germany for now but I’m willing to extend it. I build on @ellis.eu resources and my own knowledge of these systems.

Citoyennes, citoyens, les #mathématiques ont besoin de vous 📢

Du 10 mars au 30 avril 2025, venez donner votre point de vue sur les maths en participant à notre consultation citoyenne "Aux maths citoyennes, citoyens !" 💬

Pour participer 👉 buff.ly/3ZDVI82

👋 @cnrs.fr

La consultation nationale « Aux maths citoyennes et citoyens » chez @franceinter 🎙️

En ce 1er jour de la semaine des #Mathématiques et jour de lancement de la consultation, Christophe Besse, directeur de @cnrs-insmi.bsky.social , est l'invité de 6h20.

À écouter 👉 buff.ly/MaDQaks

#Communiqué 🗞️ Face aux enjeux des #maths, le @cnrs.fr lance, le 10 mars 2025, une grande #ConsultationNationale pour interroger les citoyennes et les citoyens sur la place des mathématiques dans la société.
👉 www.cnrs.fr/fr/presse/au...

📣 New preprint 📣

Learning Theory for Kernel Bilevel Optimization

w/ @fareselkhoury.bsky.social E. Pauwels @michael-arbel.bsky.social

We provide generalization error bounds for bilevel optimization problems where the inner objective is minimized over a RKHS.

arxiv.org/abs/2502.08457

Great post! Bring back a lot of memories and gives some interesting lessons. You're encourage to read it with an onion on the belt, which was the style at the time!

Open Assistant Professor position in Statistical Learning or Data Science at École Polytechnique at Paris/Saclay. Reasonable teaching hours on ML/Stat or DS courses with very good students. Deadline is March 28th 2025, contact us if interested. gargantua.polytechnique.fr/siatel-web/a...

PhD Position: Probabilistic approaches to tipping points in natural systems with application to cloud transitions Hello everyone! I would like to advertise a PhD position I got funding for, at the University of Exeter: Probabilistic approaches to tipping points in natural systems with application to cloud trans...

Cool PhD position available, where climate modeling meets probabilistic programming and #JuliaLang!
discourse.julialang.org/t/phd-positi...

an electrical transmission tower with some insulators and wires. some insulators are not where they should be. the bottom right of the photo has the date "11.11.2018"

this electrical transmission tower has a little problem. can you spot it? actually, it's not a small problem--it cost us 16.65 *billion* dollars and caused the deaths of 85 people.

Videos of the CNRS optimization conference now online (in French):
- Claire Mathieu : www.youtube.com/watch?v=_ZXZ...
- Gabriel Peyré : www.youtube.com/watch?v=vQOF...
- Jérôme Bolte : www.youtube.com/watch?v=tjkg...
- Axel Parmentier : www.youtube.com/watch?v=DohO...

Enjoy 🙂