Théo Guyard

Postdoctoral researcher at @polymtl working on OR/ML topics.

[Théo Guyard @TheoGuyard]

Definitely a wonderful time to be alive

[Théo Guyard @TheoGuyard]

What a great time to live in

[Théo Guyard @TheoGuyard]

After some rest at the end of my PhD, I'll be back to maths tomorrow with a new postdoctoral position at @polymtl 🇨🇦 Thanks @vidalthi for this opportunity !

[Théo Guyard @TheoGuyard]

Hey people ! I will defend my PhD on Wednesday at 2pm (UTC+1). It's about designing Branch-and-Bound algorithms to solve L0-regularized problems. You can watch it live on YouTube at https://t.co/3s4809YpVf

Feel free to share the link with anyone who might be interested 🤓

[Théo Guyard @TheoGuyard]

Next Monday, I'll be talking about relaxations of the L0-norm at the GdR IASIS workshop organized at @InHenriPoincare (thanks
@EmmanuelSoubies
and Luca Calatroni for the invitation 🙌). Let's meet there !

[Laurent Duval (research amateur) @laurentduval]

16h10-16h30: Theo Guyard: Toward stronger relaxations for L0-regularized problems
16h30-16h50: Diego Delle Donne: A Novel Integer Linear Programming Approach for Global L0 Minimization
16h50-17h10: Nicolas Nadisic: Global optimization for L0-constrained simultaneous sparse coding

[Laurent Duval (research amateur) @laurentduval]

Lundi 16 septembre, Luca Calatroni et Emmanuel Soubies @EmmanuelSoubies optimisent et relaxent parcimonieusement le GdR IASIS avec "L0-based minimization: from continuous relaxations and non-convex algorithms to global optimization" https://t.co/6BA0cocA7F

[Théo Guyard @TheoGuyard]

I'll be presenting our paper on Branch-and-Bound algorithms for L0-regularized problems at @icmlconf Wednesday at 11:30 (poster #1101) 😎 Feel free to join !

The paper is available here: https://t.co/9Y9lVThjDb

[Théo Guyard @TheoGuyard]

📢New paper out at @icmlconf

"A New Branch-and-Bound Pruning Framework for L0-Regularized Problems"

The paper aims to provide tractable solvers for 𝗹𝗮𝗿𝗴𝗲-𝘀𝗰𝗮𝗹𝗲 L0-regularized problems. Check-out how we managed to do this via BnB algorithms!

https://t.co/9Y9lVThjDb

[Théo Guyard @TheoGuyard]

@bozavlado Note that to use Gurobi/Cplex/Mosek/... optimizers, you need a valid licence and install the python dependencies required for its usage.

[Théo Guyard @TheoGuyard]

🚀 Need a fast and flexible solver for L0-norm regularized problems ? The first release of el0ps is out ! 🚀

https://t.co/n1ayC6llYO

The package is just out of the oven so any feedback is welcome 🥳

[Théo Guyard @TheoGuyard]

@bozavlado b) No GPU support yet for the `BnbSolver`. However, with the `MipSolver`, you can leverage off-the-shelf optimizers such as Gurobi that do allow GPU usage.

`BnbSolver`: https://t.co/Y0mjLGus8S
`MipSolver`: https://t.co/OrgUdj7Z29

[Théo Guyard @TheoGuyard]

@bozavlado Hey !

a) It really depends on the problem structure. Some instances (including real-world ones) where the variable has ~10^5 entries can be solved in seconds. However, some instances in much smaller dimensions can also turn out to be intractable.

[Théo Guyard @TheoGuyard]

Just wrote a 5 pages long rebuttal and found out that only 2500 characters are allowed 🙃

[Théo Guyard @TheoGuyard]

Today is the beginning of a loooong journey 😵‍💫

[Théo Guyard @TheoGuyard]

How dare you, Python ?

[Théo Guyard @TheoGuyard]

@KatjaZia @Inria_Rennes @virt_us 👏👏👏

[Moreau Thomas @tomamoral]

We just finished three days of benchmarking sprint with benchopt!! With 40 participants, we had lots of fun, energy, and nice achievements!

Here are a few takeaways from this sprint and our discussions!

Blog post: https://t.co/790adUA80P

Doc: https://t.co/Gkf91NQ4V0

[Théo Guyard @TheoGuyard]

🔈 New paper accepted at EUSIPCO : https://t.co/0L65TcDtFJ

« Safe Peeling for L0-Regularized Least-Squares » with G. Monnoyer, C. Herzet and C. Elvira.

L0-problems have too poor relaxations so that BnB can solve them efficiently? Check out our idea for strengthening them!

[Théo Guyard @TheoGuyard]

Meet me at #SIAMOP23 in Seattle on Friday at 3.15 PM. I’ll present a part of my PhD work about solution methods for L0-penalized problems 🤓