Agenda

Département
Langue
Date
Thématique
2026

March

  • 13:00
    14:00

    Between 2018 and 2024, the American company Digital Realty built four data centers at the Port of Marseille-Fos, an industrial hub located on the northern coast of Marseille. The port seized the opportunity offered by the city's growing role in global data exchanges. Over the past decade, Marseille has risen from 44th to 5th place among the world's most connected cities. Since 2021, the city of Marseille and activist groups have been contesting the establishment of data centers. In this context, this presentation will examine how the convergence (or divergence) of interests between regional state agencies and private infrastructure players influences the establishment of data centers in Marseille.

    Biography: Loup Cellard holds a PhD in Interdisciplinary Studies from the Centre for Interdisciplinary Methodologies (University of Warwick, UK). He is currently a postdoctoral researcher at LISN and a member of the observatory on the environmental impact of AI at the ENS's AI and Society Institute. Previously, Loup Cellard was a researcher at the Datactivist cooperative.

     

    Salle 178
  • 12:30
    13:00
    Café science ouverte

    Anna Loeff, Research Data Project Manager, Cross-functional Research Support Service (DIRDOC), will host a webinar.
    Presentation of the multidisciplinary national data warehouse Recherche Data Gouv, the institutional space of the University of Bordeaux, and details of the procedure for submitting and promoting your data.

    Registration.

    This short format (30 min) aims to raise awareness of the challenges and practices of open science among the entire Bordeaux university community, particularly researchers and doctoral students.
     

  • 14:00
    10:49
    Séminaire Benoit Carré

    From the spectacular emergence of generative AI for music to voice cloning, AI agents, and the possibility of coding applications without being a coder, I will share my journey as a musician through examples, use cases, and artistic projects involving several types of AI. I will also share my point of view, backed by scientific studies, on the evolution of pop music and the contradictory feelings that generative AI provokes.

    Biography: Benoit Carré (SKYGGE) has been at the forefront of music creation with AI since 2015. He collaborates with research teams (Sony CSL, Spotify) to develop AI prototypes for music. His first pop album designed with AI, Hello World (2018), was followed by American Folk Songs, Melancholia, and Clo_w_nes. He is currently working on a live concept called “Les Chansons Impossibles” (The Impossible Songs). Expert on the Ministry of Culture's AI Committee (2024). Songwriter (Lilicub, Maurane, Françoise Hardy, Imany). Composes for the film “Les Rêveurs” (2025).

     

    Salle Hémicyclia
  • 14:00
    15:00

    In automated learning problems, the task is to find a model that maps given inputs to their corresponding outputs as accurately as possible. Over the past 30 years, machine learning and deep learning have achieved tremendous success in solving this type of problem. However, while the resulting models can be used to make predictions, they offer limited interpretability, i.e. they provide little insight into *how* they solve the problem. For instance, if we train a recurrent neural network to predict whether a sequence of events will lead to a crash, the model cannot describe the sequences of events that lead to crashes and provide solutions to remediate the issue. Program synthesis is a framework for solving learning problems with models that are programs in a domain-specific language, which allows creating interpretable models in the domain of the problem.

    In this talk, I will present an enumerative approach for learning Linear Temporal Logic (LTL) formulas from data. I will begin with an introduction to program synthesis, using examples from program de-obfuscation and anomaly explanation. In the second part, I will present the main techniques used in our algorithm: observational equivalence and domination, two pruning techniques used to reduce the search space, and a connection between LTL Learning and the Boolean Set Cover problem. Finally, I will discuss the engineering choices we made when implementing this algorithm in Bolt, an open-source tool available at https://github.com/SynthesisLab/Bolt.

    This talk is based on joint work with Nathanaël Fijalkow, Théo Matricon, Baptiste Mouillon and Pierre Vandenhove.

    English
    LaBRI

April