La semaine de l’IMB

• Monday 02/02/2026

13h30 – 14h00 : Welcome and introduction; Guillaume Ferré

14h00 – 17h30 : Tutorial - Introduction to programming on Anduino – what is an Arduino?; Marwane Rezzouki

• Tuesday 03/02/2026

9h00 – 10h30 : Lecture - How to choose a communication technique?; Marwane Rezzouki

11h00 – 12h00 : Lecture - From object to cloud: establishing communication, data framing and display; Marwane Rezzouki

13h30 – 17h00 : Tutorial - From object to cloud: establishing communication, data framing and display; Marwane Rezzouki

17h00 – 18h00 : Lecture - Thibault Pirson

• Wednesday 04/02/2026

9h00 – 10h30 : Lecture - Internet of Things and big data; Nicolas Hanusse

11h00 – 12h00 : Lecture - Learning in the Internet of Things framework; Jérémie Bigot

13h30 – 17h00 : Tutorial - Outliers processing; Pierrick Legrand

• Thursday 05/02/2026

11h00 – 12h00 : Lecture - Energy harvesting; Philippe Laurent

13h30 – 17h00 : Tutorial - From rapid prototyping to maker culture; Julien Allali

• Friday 06/02/2026

11h00 – 12h00 : Lecture - Deep learning for big data in the Internet of Things; Akka Zemmari

13h30 – 17h00 : Tutorial - From raw data to learning and deep learning; Romain Giot

We introduce the Kit Update Problem, a new online optimization problem motivated by a real-world application at Doctors Without Borders. Emergency response kits -- pre-packed sets of medical items -- must be maintained in a continuously ready-to-ship state. Because many items are perishable, maintaining readiness incurs maintenance costs (manual effort) and potentially some additional expected disposal costs (due to expired or prematurely replaced items). The objective is to minimize cumulative cost over time while ensuring continuous kit operability. We show that the problem is NP-hard (indeed APX-hard) and develop a 2-approximation algorithm. We further prove that this algorithm naturally extends to a fast 2-competitive online algorithm. The resulting policy has been adopted and integrated into the organization’s operations, where it has led to reported cost reductions of 30–40% in practice. We also compare our approach with a stochastic look-ahead policy on synthetic instances calibrated to real data provided by the organization. The online algorithm achieves nearly the same performance, while requiring no forecasts, minimal computation, and exhibiting robustness to misspecification of cost parameters, making it well suited to such settings.

Attention-based models, such as Transformer, excel across

various tasks but lack a comprehensive theoretical understanding. To

address this gap, we introduce the single-location regression task,

where only one token in a sequence determines the output, and its

position is a latent random variable, retrievable via a linear

projection of the input. To solve this task, we propose a dedicated

predictor, which turns out to be a simplified version of a non-linear

self-attention layer. We study its theoretical properties, both in terms

of statistics (gap to Bayes optimality) and optimization (convergence of

gradient descent). In particular, despite the non-convex nature of the

problem, the predictor effectively learns the underlying structure. This

highlights the capacity of attention mechanisms to handle sparse token

information. Based on Marion et al., Attention Layers Provably Solve

Single-Location Regression, ICLR 2025, and Duranthon et al., Statistical

Advantage of Softmax Attention: Insights from Single-Location

Regression, submitted.

L'ordre du jour sera le suivant :

1) Présentation des IAGs par Laurent Simon (Professeur à Bordeaux INP, Google) : fonctionnement général, possibilités et risques

2) Illustration sur 3 problèmes pratiques correspondant à chacune des activités administration, enseignement et recherche

3) Rappel des règles générales d'utilisation et synthèse des consignes actuelles de nos tutelles (Bordeaux INP, CNRS, Inria, Université de Bordeaux)

4) Questions diverses

This work concerns the numerical approximations of the weak solutions of scalar hyperbolic conservation laws. After showing how to bypass the discrete entropy inequality barrier theorems for the linear advection, the derivation of a second-order entropy-satisfying scheme is presented for non-linear equations. The fully discrete stability result is established for regular strictly convex entropy and under a parabolic CFL-like condition. Some numerical experiments are done to assess the accuracy and the stability of the proposed scheme.

In this talk, we revisit some of the fundamental results on unique continuation for elliptic operators with variable coefficients. Inspired by the recent examples constructed by Mandache and Krymskii-Logunov-Pagano, we sharpen some quantitative unique continuation properties for solutions $u$ of divergence-form elliptic equations with variable coefficients $A$. In the local setting, we prove that if $A$ is log-Lipschitz, then strong unique continuation still holds for solutions, which sharpens and clarifies the Lipschitz threshold pioneered by Han-Lin. I will also discuss joint work with B. Davey, where we prove optimal Landis-type results for solutions of Schrodinger-type equations with variable coefficients in cylinders.

Il résulte du théorème d'Aubin-Yau (1978) qu'une variété projective complexe lisse X de dimension n canoniquement polarisée est revêtue par la boule unité de C^n (donc est un quotient étale de cette dernière) si et seulement si une certaine égalité entre les deux premières classes de Chern est réalisée (pour n=2, c'est la fameuse égalité c_1^{^2}=3c_2). Je présenterai un analogue où l'on s'autorise à regarder des quotients non étales et donc des variétés X qui sont singulières. Il s'agit d'un travail en commun avec Benoît Claudon et Patrick Graf.

La cohomologie de certains systèmes locaux étales p-adiques universels sur le demi-plan de Drinfeld encode la correspondance de Langlands p-adique dans le cas dit spécial.

Dans cet exposé, j’expliquerai comment l’étude de certains anneaux de déformation, appelés anneaux de Kisin spéciaux, permet d’affiner ce résultat en une description complète de la cohomologie de ces systèmes locaux en termes de la correspondance de Langlands p-adique en famille. Le résultat se présente sous une forme factorisée, analogue à la description d’Emerton de la cohomologie des courbes modulaires.

Je présenterai ces résultats, ainsi que les idées et les étapes clés de leurs démonstrations.