Sofiane Tanji
@sofianetanji.com
I study Mathematical Optimization @UCLouvain
Reposted by Sofiane Tanji
New paper out! Together with my postdoc advisors @esoubies.bsky.social and Cédric Févotte :)
We study the optimization landscape of L0-Bregman relaxations, which naturally arise when one optimizes non-quadratic fidelity terms coupled with the L0-norm.
1/n
We study the optimization landscape of L0-Bregman relaxations, which naturally arise when one optimizes non-quadratic fidelity terms coupled with the L0-norm.
1/n
Jonathan Chirinos-Rodriguez, C\'edric F\'evotte, Emmanuel Soubies
Optimization landscape of $\ell_0$-Bregman relaxations
https://arxiv.org/abs/2511.12157
Optimization landscape of $\ell_0$-Bregman relaxations
https://arxiv.org/abs/2511.12157
November 20, 2025 at 8:45 PM
New paper out! Together with my postdoc advisors @esoubies.bsky.social and Cédric Févotte :)
We study the optimization landscape of L0-Bregman relaxations, which naturally arise when one optimizes non-quadratic fidelity terms coupled with the L0-norm.
1/n
We study the optimization landscape of L0-Bregman relaxations, which naturally arise when one optimizes non-quadratic fidelity terms coupled with the L0-norm.
1/n
Reposted by Sofiane Tanji
Lately, I have been obsessed with developing theoretically based optimization algorithms that actually attain the best practical performance.
Alas, the classic model of minimax optimal methods is overly conservative; it overfits to tune its worst-case.
We found a path forward 1/
Alas, the classic model of minimax optimal methods is overly conservative; it overfits to tune its worst-case.
We found a path forward 1/
November 18, 2025 at 2:59 PM
Lately, I have been obsessed with developing theoretically based optimization algorithms that actually attain the best practical performance.
Alas, the classic model of minimax optimal methods is overly conservative; it overfits to tune its worst-case.
We found a path forward 1/
Alas, the classic model of minimax optimal methods is overly conservative; it overfits to tune its worst-case.
We found a path forward 1/
Reposted by Sofiane Tanji
I have several offers for Master internships / PhDs on graph ML funded by ERC MALAGA for 2026. Don't hesitate to contact me to apply!
All infos here: nkeriven.github.io/malaga/
All infos here: nkeriven.github.io/malaga/
MALAGA: Reinventing the Theory of Machine Learning on Large Graphs (ERC StG)
nkeriven.github.io
November 6, 2025 at 1:56 PM
I have several offers for Master internships / PhDs on graph ML funded by ERC MALAGA for 2026. Don't hesitate to contact me to apply!
All infos here: nkeriven.github.io/malaga/
All infos here: nkeriven.github.io/malaga/
Reposted by Sofiane Tanji
A regrettable sign of the times, but it shouldn’t be misreported: they aren’t restricting most CS papers. They are restricting ONLY review/survey papers and position papers, which weren’t common in CS anyway prior to the AI onslaught. No new restrictions on CS papers consisting of original research.
November 3, 2025 at 5:13 PM
A regrettable sign of the times, but it shouldn’t be misreported: they aren’t restricting most CS papers. They are restricting ONLY review/survey papers and position papers, which weren’t common in CS anyway prior to the AI onslaught. No new restrictions on CS papers consisting of original research.
Reposted by Sofiane Tanji
I’m recruiting a postdoctoral researcher on a fully funded 2-year contract. The topic is developing and implementing new parallel graph algorithms, with applications to data-driven transportation and logistics.
Details and application procedure below ⬇️ Reach out if you have any questions!
Details and application procedure below ⬇️ Reach out if you have any questions!
acme-mobidec.github.io
October 27, 2025 at 5:43 PM
I’m recruiting a postdoctoral researcher on a fully funded 2-year contract. The topic is developing and implementing new parallel graph algorithms, with applications to data-driven transportation and logistics.
Details and application procedure below ⬇️ Reach out if you have any questions!
Details and application procedure below ⬇️ Reach out if you have any questions!
Reposted by Sofiane Tanji
An interesting knock-on effect of various recent hype cycles at the intersection of LLMs and Maths is that I've found some motivation to care a bit more about what people really mean by "open problem".
October 18, 2025 at 1:45 PM
An interesting knock-on effect of various recent hype cycles at the intersection of LLMs and Maths is that I've found some motivation to care a bit more about what people really mean by "open problem".
Reposted by Sofiane Tanji
#ICLR2026 PCs consider prompt injection as an ethics problem, in particular an attempt to collude with a potential malicious reviewer (using an LLM). I can't agree more.
blog.iclr.cc/2025/08/26/p...
blog.iclr.cc/2025/08/26/p...
August 27, 2025 at 6:50 AM
#ICLR2026 PCs consider prompt injection as an ethics problem, in particular an attempt to collude with a potential malicious reviewer (using an LLM). I can't agree more.
blog.iclr.cc/2025/08/26/p...
blog.iclr.cc/2025/08/26/p...
Reposted by Sofiane Tanji
My newish go-to "intro to RL from first principles" is Ch. 11/12 of Moritz Hardt & @beenwrekt.bsky.social's ML book. They're not 100% standalone chapters but can work that way. I now like this as an intro more than Sutton & Barto (more concise, conceptually clearer, connections w/ control theory).
Sequential decision making and dynamic programming
mlstory.org
April 29, 2025 at 4:17 PM
My newish go-to "intro to RL from first principles" is Ch. 11/12 of Moritz Hardt & @beenwrekt.bsky.social's ML book. They're not 100% standalone chapters but can work that way. I now like this as an intro more than Sutton & Barto (more concise, conceptually clearer, connections w/ control theory).
Reposted by Sofiane Tanji
Announcing : The 2nd International Summer School on Mathematical Aspects of Data Science
mathsdata2025.github.io
EPFL, Sept 1–5, 2025
Speakers:
Bach @bachfrancis.bsky.social
Bandeira
Mallat
Montanari
Peyré @gabrielpeyre.bsky.social
For PhD students & early-career researchers
Apply before May 15!
mathsdata2025.github.io
EPFL, Sept 1–5, 2025
Speakers:
Bach @bachfrancis.bsky.social
Bandeira
Mallat
Montanari
Peyré @gabrielpeyre.bsky.social
For PhD students & early-career researchers
Apply before May 15!
Mathematical Aspects of Data Science
Graduate Summer School - EPFL - Sept. 1-5, 2025
mathsdata2025.github.io
April 14, 2025 at 5:00 PM
Announcing : The 2nd International Summer School on Mathematical Aspects of Data Science
mathsdata2025.github.io
EPFL, Sept 1–5, 2025
Speakers:
Bach @bachfrancis.bsky.social
Bandeira
Mallat
Montanari
Peyré @gabrielpeyre.bsky.social
For PhD students & early-career researchers
Apply before May 15!
mathsdata2025.github.io
EPFL, Sept 1–5, 2025
Speakers:
Bach @bachfrancis.bsky.social
Bandeira
Mallat
Montanari
Peyré @gabrielpeyre.bsky.social
For PhD students & early-career researchers
Apply before May 15!
Reposted by Sofiane Tanji
The problem with most machine-based random number generators is that they’re not TRULY random, so if you need genuine randomness it is sometimes necessary to link your code to an external random process like a physical noise source or the current rate of US tariffs on a given country.
April 9, 2025 at 7:15 PM
The problem with most machine-based random number generators is that they’re not TRULY random, so if you need genuine randomness it is sometimes necessary to link your code to an external random process like a physical noise source or the current rate of US tariffs on a given country.
Reposted by Sofiane Tanji
1-day workshop on Bilevel optimization and hyperparameter tuning at ENS de Lyon, on March 25th : gdr-iasis.cnrs.fr/reunions/bil...
Keynote talks by @jmairal.bsky.social @tonysf.bsky.social @samuelvaiter.com, Saverio Salzo and Luce Brotcorne, contributed talks are welcome!
Keynote talks by @jmairal.bsky.social @tonysf.bsky.social @samuelvaiter.com, Saverio Salzo and Luce Brotcorne, contributed talks are welcome!
Bilevel Optimization and Hyperparameter Learning - GdR IASIS
Bilevel optimization has become a crucial framework in machine learning for addressing hierarchical problems, where one optimization task depends on the outcome of another. This approach plays a pivot...
gdr-iasis.cnrs.fr
February 28, 2025 at 8:56 AM
1-day workshop on Bilevel optimization and hyperparameter tuning at ENS de Lyon, on March 25th : gdr-iasis.cnrs.fr/reunions/bil...
Keynote talks by @jmairal.bsky.social @tonysf.bsky.social @samuelvaiter.com, Saverio Salzo and Luce Brotcorne, contributed talks are welcome!
Keynote talks by @jmairal.bsky.social @tonysf.bsky.social @samuelvaiter.com, Saverio Salzo and Luce Brotcorne, contributed talks are welcome!
Reposted by Sofiane Tanji
🎓🌞 Join the PhD course on Theoretical Foundations of Machine Learning at 🇮🇹 ELLIS Unit Genoa, June 23-27, 2025, hosted by Machine Learning Genoa Center.
🗓️ Apply by: March 16, 2025
📍 University of Genoa (in-person only)
💡 More info: malga.unige.it/education/sc...
🗓️ Apply by: March 16, 2025
📍 University of Genoa (in-person only)
💡 More info: malga.unige.it/education/sc...
TFMLUniGe ¦ MaLGa
malga.unige.it
February 28, 2025 at 10:25 AM
🎓🌞 Join the PhD course on Theoretical Foundations of Machine Learning at 🇮🇹 ELLIS Unit Genoa, June 23-27, 2025, hosted by Machine Learning Genoa Center.
🗓️ Apply by: March 16, 2025
📍 University of Genoa (in-person only)
💡 More info: malga.unige.it/education/sc...
🗓️ Apply by: March 16, 2025
📍 University of Genoa (in-person only)
💡 More info: malga.unige.it/education/sc...
Reposted by Sofiane Tanji
The University of Pittsburgh confirmed Friday that there would be no new Ph.D. offers of admission while Pitt works to understand how reduced federal aid could impact the institution.
The University of Pittsburgh pauses its Ph.D. admissions process amid research funding uncertainty
A spokesperson for the University told WESA Friday that the school has "temporarily paused additional Ph.D. offers of admission," while Pitt works to understand how proposed federal funding cuts could...
www.wesa.fm
February 21, 2025 at 9:56 PM
The University of Pittsburgh confirmed Friday that there would be no new Ph.D. offers of admission while Pitt works to understand how reduced federal aid could impact the institution.
Reposted by Sofiane Tanji
Happy to share our work on learning theory for kernel bilevel optimization! Grateful to Edouard, @samuelvaiter.com, and @michael-arbel.bsky.social for their mentorship and insightful discussions that shaped this paper!
📣 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
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
February 20, 2025 at 9:30 PM
Happy to share our work on learning theory for kernel bilevel optimization! Grateful to Edouard, @samuelvaiter.com, and @michael-arbel.bsky.social for their mentorship and insightful discussions that shaped this paper!
Reposted by Sofiane Tanji
📣 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
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
February 20, 2025 at 1:55 PM
📣 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
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
Reposted by Sofiane Tanji
Very glad of this new collaborative work "Herglotz-NET: Implicit Neural Representation of Spherical Data with Harmonic Positional Encoding" with Théo Hanon, Nicolas Mil-Homens Cavaco, John Kiely, Laurent Jacques https://arxiv.org/html/2502.13777v1
Herglotz-NET: Implicit Neural Representation of Spherical Data with Harmonic Positional Encoding
arxiv.org
February 20, 2025 at 6:15 AM
Very glad of this new collaborative work "Herglotz-NET: Implicit Neural Representation of Spherical Data with Harmonic Positional Encoding" with Théo Hanon, Nicolas Mil-Homens Cavaco, John Kiely, Laurent Jacques https://arxiv.org/html/2502.13777v1
Reposted by Sofiane Tanji
🧗♂️Why GD converges beyond [step size]<2/[smoothness]? We investigate loss functions and identify their *separation margin* is an important factor. Surprisingly Renyi 2-entropy yields super fast rate T=Ω(ε^{-1/3})!
arxiv.org/abs/2502.04889
arxiv.org/abs/2502.04889
February 10, 2025 at 3:04 AM
🧗♂️Why GD converges beyond [step size]<2/[smoothness]? We investigate loss functions and identify their *separation margin* is an important factor. Surprisingly Renyi 2-entropy yields super fast rate T=Ω(ε^{-1/3})!
arxiv.org/abs/2502.04889
arxiv.org/abs/2502.04889
Reposted by Sofiane Tanji
Very relatable. It would be nice to say that my research reading habits are more tactical than they are, but for me, it is really much closer to "an offer which I can't refuse"; effortlessly enjoyable and an embedded part of my routine. I've described it in the past as being a bit like gardening.
I personally read at least a couple of hours per day. It is not particularly focused and I might "waste" time but I just enjoy it.
February 10, 2025 at 12:07 PM
Very relatable. It would be nice to say that my research reading habits are more tactical than they are, but for me, it is really much closer to "an offer which I can't refuse"; effortlessly enjoyable and an embedded part of my routine. I've described it in the past as being a bit like gardening.
Reposted by Sofiane Tanji
I personally read at least a couple of hours per day. It is not particularly focused and I might "waste" time but I just enjoy it.
January 8, 2025 at 8:05 AM
I personally read at least a couple of hours per day. It is not particularly focused and I might "waste" time but I just enjoy it.
Reposted by Sofiane Tanji
The Riesz Representation Theorem tells you that everything in the dual space, denoted (R^n)* when the primal is R^n, looks like row vectors. So a dual element is like z^T but you must think of it as a *function* not a vector. Evaluation is the inner product.
February 2, 2025 at 7:34 PM
The Riesz Representation Theorem tells you that everything in the dual space, denoted (R^n)* when the primal is R^n, looks like row vectors. So a dual element is like z^T but you must think of it as a *function* not a vector. Evaluation is the inner product.
Reposted by Sofiane Tanji
Early stopping on validation loss? This leads to suboptimal calibration and refinement errors—but you can do better!
With @dholzmueller.bsky.social, Michael I. Jordan, and @bachfrancis.bsky.social, we propose a method that integrates with any model and boosts classification performance across tasks.
With @dholzmueller.bsky.social, Michael I. Jordan, and @bachfrancis.bsky.social, we propose a method that integrates with any model and boosts classification performance across tasks.
February 3, 2025 at 1:03 PM
Early stopping on validation loss? This leads to suboptimal calibration and refinement errors—but you can do better!
With @dholzmueller.bsky.social, Michael I. Jordan, and @bachfrancis.bsky.social, we propose a method that integrates with any model and boosts classification performance across tasks.
With @dholzmueller.bsky.social, Michael I. Jordan, and @bachfrancis.bsky.social, we propose a method that integrates with any model and boosts classification performance across tasks.
Reposted by Sofiane Tanji
Does anyone have pointers to tutorials on double descent and what it tells us about "overfitting?"
February 3, 2025 at 2:58 PM
Does anyone have pointers to tutorials on double descent and what it tells us about "overfitting?"
Reposted by Sofiane Tanji
The death toll from Israel’s war on Gaza has been revised to more than 61,700 people according to government officials, after reclassifying thousands of people who had been missing in the rubble.
February 3, 2025 at 8:50 AM
The death toll from Israel’s war on Gaza has been revised to more than 61,700 people according to government officials, after reclassifying thousands of people who had been missing in the rubble.
Reposted by Sofiane Tanji
This review paper by @guillaume-garrigos.com on SGD-related algorithms is a fantastic resource, offering elegant, self-contained, and concise proofs in a single, accessible reference. arxiv.org/pdf/2301.11235
January 29, 2025 at 4:15 PM
This review paper by @guillaume-garrigos.com on SGD-related algorithms is a fantastic resource, offering elegant, self-contained, and concise proofs in a single, accessible reference. arxiv.org/pdf/2301.11235
Reposted by Sofiane Tanji
🎓 💫 We are opening post-doc positions at the intersection of AI, data science, and medicine:
• Large Language Models for French medical texts
• Evaluating digital medical devices: statistics and causal inference
• Large Language Models for French medical texts
• Evaluating digital medical devices: statistics and causal inference
January 29, 2025 at 8:19 AM
🎓 💫 We are opening post-doc positions at the intersection of AI, data science, and medicine:
• Large Language Models for French medical texts
• Evaluating digital medical devices: statistics and causal inference
• Large Language Models for French medical texts
• Evaluating digital medical devices: statistics and causal inference