Agenda depuis 2025
Conference : Seminaire IACD Invité - Gianluca Bontempi (Université libre de Bruxelles)
19/05/2025 à 11h30
Titre: From feature selection to causal discovery in large dimensional settings Lieu: Salle de conseil (St Jerome) Résumé: We are drowning in data and starving for knowledge” is an adage of data scientists that nowadays should be rephrased into ”we are drowning in associations and starving for causality”. The democratisation of machine learning software and big data platforms increases the risk of ascribing causal meaning to simple and sometimes brittle associations. This risk is particularly evident in settings (like bioinformatics, social sciences, and economics) characterised by high dimensions, multivariate interactions, and dynamic behaviour where direct manipulation is unethical and impractical. The conventional ways to recover a causal structure from observational data are score-based and constraint-based algorithms. Their limitations, mainly in high dimension, opened the way to alternative learning algorithms which pose the problem of causal discovery as the classification of probability distributions. The rationale of those algorithms is that the existence of a causal relationship induces a constraint on the observational multivariate distribution. In other words, causality leaves footprints in the data distribution that can hopefully be used to reduce the uncertainty about the causal structure. The first part of the presentation will introduce some basics of causal discovery and will present a causal extension of a feature selection method commonly used in bioinformatics. The second part of the talk will focus on the D2C approach which featurizes observed data using information theory asymmetric measures to extract meaningful hints about the causal structure. The D2C algorithm performs three steps to predict the existence of a directed causal link between two variables in a multivariate setting: (i) it estimates the Markov Blankets of the two variables of interest and ranks its components in terms of their causal nature, (ii) it computes several asymmetric descriptors and (iii) it learns a classifier (e.g. a Random Forest) returning the probability of a causal link given the value of the descriptor. The final part of the presentation is more prospective and will introduce some recent work to implement counterfactual prediction in a data-driven setting.Seminaire : Séminaire I&M - Sabrine djedjiga OUCHERIF
28/03/2025 à 14h00
Titre : APPORTS D’UN SYSTÈME DE VISION PLÉNOPTIQUE POUR L’ANALYSE DES COMPORTEMENTS.Lieu : Luminy, TPR2 salle de réunion 04.03.
Résumé :
La reconnaissance des expressions faciales constitue depuis plusieurs décennies une méthode de choix pour l’analyse fine des émotions humaines, et représente un enjeu majeur dans l’étude du comportement. Dans une perspective de recherche interdisciplinaire, ce domaine s’enrichit aujourd’hui grâce à l’intégration de travaux académiques pointus et de technologies avancées. Ce projet s’inscrit dans une collaboration entre le Centre d’Étude et de Recherche en Gestion d’Aix Marseille (CERGAM), le Laboratoire d’Informatique et Systèmes (LIS) et l’Institut de Mathématiques de Marseille (I2M). L’utilisation de caméras plénoptiques, capables de capturer simultanément l’intensité et la direction des rayons lumineux, permet d’obtenir des images multi-perspectives en une seule acquisition [1, 2]. Ces images facilitent la génération de cartes de profondeur détaillées et optimisent l’analyse des subtilités des expressions faciales, offrant ainsi un éclairage nouveau sur les réponses émotionnelles des consommateurs.
Ce projet vise à approfondir l’analyse comportementale des individus en environnement de vente en se focalisant sur l’étude et la reconnaissance des expressions faciales. À cet effet, nous mobilisons des dispositifs de capture avancés (caméras plénoptiques) fournissant une diversité d’images, incluant des prises de vue en sous-ouverture, en total focus et des cartes de profondeur. Ces différents supports permettent de constituer un jeu de données riche et varié, indispensable à une analyse fine des interactions en situation réelle.
Pour traiter ces données, nous proposons d’implémenter des architectures hybrides combinant réseaux de neurones convolutifs (CNN) et récurrents (RNN), afin d’extraire efficacement les informations spatiales et temporelles. Par ailleurs, l’adoption d’une approche de fusion multimodale constitue un levier majeur pour intégrer et exploiter l’ensemble des modalités capturées, dans le but d’obtenir une analyse comportementale précise et approfondie
[1] T. -W. Shen et al., "Facial expression recognition using depth map estimation of Light Field Camera," 2016 IEEE International Conference on Signal Processing, Communications and Computing(ICSPCC), 2016, pp.1-4, doi: 10.1109/ICSPCC.2016.7753695.
[2] A. Sepas-Moghaddam, A. Etemad, F. Pereira and P. L. Correia, "Facial Emotion Recognition Using Light Field Images with Deep Attention-Based Bidirectional LSTM," ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 3367-3371, doi: 10.1109/ICASSP40776.2020.9053919
Seminaire : Séminaire CANA : Alana Spak
18/03/2025 à 14h00
Title : Thermalization via quantum walksSalle : 04.05 TPR2
Abstract :
A central question in quantum thermodynamics and the foundations of quantum mechanics is how and when thermalization emerges in quantum systems. An intriguing framework for studying this phenomenon is the quantum walk, a model that allows the analysis of quantum correlations. Entanglement, for example, can induce decoherence, ultimately leading to thermalization. In this seminar, we introduce a quantum walk in energy space to investigate thermodynamic equilibrium in quantum systems. Our approach is based on monodesmic processes, derived from Markov chains, in the asymptotic limit. By analyzing the classical limit, we show that the equilibrium temperature can be determined solely by the transition probabilities between the energy levels. This approach offers valuable insights into the thermalization phenomena.
Seminaire : Séminaire JOURNEE DE LA BALEINE (TOULON)
13/03/2025 à 17h00
Bienvenue à la soirée DYNI / Centre d’IA en Acoustique Naturelle en streaming sur Youtube : https://youtube.com/live/y12oI_3wha4?feature=share ou en présentiel Jeudi 13 mars 2025, 17h - 20h • Amphi M, SeaTech • Campus de La Garde Cette soirée découverte s’ouvre sur deux exemples de ’Cocktail Party’ inédits, captés et analysés par le LIS CIAN. Le premier est l’unique symphonie de vocalises simultanées de quatre superprédateurs arctiques (orques, cachalots, baleines à bosse et rorquals). Le second, est le concert de biosonars entrelacés d’une meute de six cachalots que nous suivons en 3D durant 3h dans leurs chasses abyssales par -1000 mètres au large de Monaco par l’acoustique passive de notre mission de drones surface ’Sphyrna Odyssée’. Ainsi s’ouvre une face de nos recherches : quelles sont les informations, les interactions et les stratégies de leurs communications ? Nous en illustrons quelques pistes et méthodes. Nous présentons donc quelques instruments de la start-up française ALPAI, dont ses hydrophones innovants. Puis, nous explorons les dynamiques culturelles des Baleines à bosse sur les Antilles, et des chorus de cétacés en Arctique. Nous donnons un aperçu de la boucle perception-action des biosonars et de ses ambiguïtés. Nous nous concentrons alors sur celle du cachalot : comment séparer les locuteurs d’un groupe en socialisation, puis étudier leurs communications individu par individu (codas à Maurice). Nous analysons aussi leur présence non loin de Toulon-Monaco, et comment l’IA embarquée pourrait aider à les préserver des risques anthropiques. Puis, avec notre équipe australe nous découvrons les sonars convergents en ’NBHF’ des dauphins méconnus des canaux patagoniens chiliens. Enfin, nous offrons un ’Cocktail Party’ pour en discuter... Intervenant.e.s : Docteur.e.s ou doctorant.e.s. Justine Girardet, Lara Berkenbaum, Pascale Giraudet, Julie Patris, Stéphane Chavin, Nicolas Deloustal, Sébastian Marzetti, Valentin Gies, Alain Barcelo, Hervé Glotin et collaboratrices/eurs de CIAN / UMR LIS-LAB. Partenaires de ces recherches : LIS-LAB, IM2NP, Feder CARIMAM & OFB AGOA & partenaires CARIMAM, Longitude181, OrcaLab, PolarPod-Persévérance - Ocean Polaire, ValhallaB, Biodiversa EUROPAM, Feder SeaStMar, OFB PNPC, Pelagos, Chaire IA ADSIL, ANR, AID, Région PACA ETHAC, Sphyrna Odyssée FBVMarine, L181 WhaleWay, EXPLORE, SMIoT, Université de Toulon, IRP CNRS ’Detection & AI Classification of the Biosonars in the Southern Cone’Seminaire : Séminaire CANA : Nasra Daher Ahmed
04/03/2025 à 14h00
Title Introduction to quantum information theorySalle : 04.05 TPR2
Abstract :
Quantum information (QI) theory studies the ways the laws of quantum mechanics allow the storage, processing and transmission of information. Its most well-known applications are the fields of quantum computing, communication and key distribution. However, it also has applications in other fields of physics such as condensed matter and gravitation physics.
This introductory seminar aims to give the basic tools needed to understand current research in quantum information theory and its applications. First, we review some basic concepts of classical information theory and the mathematical formalism of quantum mechanics. We will then introduce some important topics in QI such as entanglement, nonlocality and no-go theorems about the manipulation of quantum information.
Seminaire : Séminaire ACRO : Abderrahim Arabi (03/03, 10h00, REU 4.05)
03/03/2025 à 10h00
Abderrahim Arabi (LIS, Aix-Marseille Université) Title: On Some Properties of Specific Polyominoes and Polycubes 03/03/2025 10h00, salle REU 04.05 (LIS Luminy) Polyominoes and polycubes are classic objects in combinatorics, yet their exact enumeration remains an open problem. One effective approach to tackle this issue is to study these structures under specific constraints. In this talk, we explore the relationships between certain families of convex and parallelogram polyominoes and polycubes. By making connections, we establish new enumeration results for polyominoes and polycubes.Seminaire : Séminaire I&M - Moncef Boujou
28/02/2025 à 14h00
Titre : Analyse du comportement des consommateurs dans les surfaces de venteLieu : Luminy, TPR2 salle de réunion 04.03.
Résumé :
L’analyse du comportement des consommateurs dans les surfaces de vente est essentielle pour mieux comprendre leurs interactions avec l’espace commercial. Toutefois, il est tout aussi important de procéder de manière non invasive afin de respecter la vie privée des clients et de ne pas biaiser leurs comportements par des dispositifs intrusifs. Dans cette optique, nous proposons une méthode de suivi multi-caméras multi-personnes reposant sur un cadre hiérarchique intégrant une phase de ré-identification et de projection, afin d’extraire automatiquement les trajectoires des clients. Toutefois, les occlusions, les changements d’apparence et les réapparitions dans différentes zones du champ de vision constituent un défi majeur pour assurer la continuité des identités. Pour pallier ces difficultés, nous présentons une nouvelle approche de ré-identification qui fusionne des descripteurs visuels et des informations extraites de la démarche, renforçant ainsi la robustesse de l’association des identités. Ce séminaire détaillera notre méthode de suivi multi caméras multi-personnes, ainsi que la nouvelle approche de ré-identification, en mettant en lumière les avancées obtenues et les perspectives qu’elles ouvrent pour l’analyse du comportement des consommateurs.
Seminaire : Séminaire CANA : Ugo Nzongani
25/02/2025 à 14h00
Title : Introduction to Quantum ComputingSalle : 04.05 TPR2
Abstract :
Quantum computing leverages the principles of quantum mechanics to develop algorithms that, in some cases, significantly outperform their classical counterparts. In this seminar, we will introduce the fundamental building blocks of quantum computation, including qubits, quantum gates, and quantum circuits. We will then briefly discuss key quantum algorithms such as Shor’s algorithm for factoring, Grover’s search algorithm, and variational methods for optimization (QAOA). Additionally, we will provide a detailed explanation of how the Deutsch-Jozsa algorithm offers an exponential speedup over classical approaches.
Seminaire : Séminaire ACRO : Nicolas Trotignon
10/02/2025 à 10h00
Nicolas Trotignon (ENS Lyon) Title: A survey of layered wheels 10/02/2025 10h00, salle REU 04.05 (LIS Luminy) Layered wheels are constructions of graphs first discovered by Ni Luh Dewi Sintiari and the speaker. They disprove several conjectures and answer several questions. For instance, jointly with Maria Chudnovsky, the speaker could use them to disprove a conjecture de Dallard, Milanič and Štorgel about the tree-independence number and a conjecture of Hajebi about the induced subgraphs contained in graphs of high treewidth. The goal of the talk is to present all the situations where layered wheels turned out to be useful and to present in more details the variant that disproved the two conjectures mentioned above.Seminaire : Séminaire I&M - DE SOUSA SILVA, Raul Alfredo
31/01/2025 à 14h00
Titre : Modélisation du comportement complexe autistique chez la souris en utilisant des approches d'apprentissage profonde et vision par ordinateurLieu : Luminy, TPR2 salle de réunion 04.03.
Résumé :
Le Trouble du Spectre Autistique (TSA) est une trouble neurodéveloppemental qui se caractérise par un déficit de communication et interaction sociale ainsi que par des comportements répétitifs et des intérêts restreints. Les souris sont des animaux sociaux et sensibles comme nous, c’est justement à cause de ce comportement social que l’on étudie les souris pour mieux comprendre l’autisme chez l’humain. Dans le cadre de ce projet, nous visons retrouver des séquences de comportement qui distinguent une souris contrôle d'une souris modèle à l'aide des algorithmes d'apprentissage profond et vision par ordinateur. Pendant la dernière année, nous avons fait plusieurs progrès dans l'amélioration de la détection de certaines actions, de la gestion des données bruitées et de la création des protocoles analytiques permettant de sortir des informations pertinentes à propos des souris. Je vous présenterai en détail ces avancées, les justificatifs méthodologiques et les prochains pas vers l'achèvement de ce projet.
Seminaire : Séminaire ACRO : Simon Vilmin (27/01, 10h00, REU 4.05)
27/01/2025 à 10h00
Simon Vilmin (LIS, Aix-Marseille Université) Title: Enumerating the minimal keys of an implicational base 27/01/2025 10h00, salle REU 04.05 (LIS Luminy) An implicational base (IB) over a groundset X is a collection of statements A --> B, called implications, where A and B are subsets of X. An IB induces a convexity space where a set C is convex if it satisfies each implication A --> B of the IB: if C includes A, it includes B. The convex hull of a subset S of X is the inclusion-wise minimal convex set including S. A key is a subset K of X whose convex hull is X, i.e. a generating set of X. In this talk, we will discuss the following problem: enumerate all the minimal keys of a convexity space given by an IB. This problem has notably been studied within the context of database theory, from which the term "key" originates and where implications are better known as "functional dependencies". It also relates to the enumeration of prime implications of a pure Horn CNF, and can be seen as a generalization of the problem of enumerating minimal hull sets---an other name for minimal keys---of (binary) graph convexities. I will present the result of Lucchesi and Osborn 78 that gives a polynomial-delay algorithm listing all minimal keys associated to an IB. Their algorithm implicitly uses solution-graph traversal to solve the problem. I will also present an alternative proof of their result using elements from the framework of proximity search.Seminaire : Séminaire CANA : Nicolas Schabanel
21/01/2025 à 14h00
Title : How do we make DNA origami and what are the challenges we tackle Nicolas SchabanelJoint work with: Nicolas Levy (LIP), Julie Finkel (CBS), Allan Mills (CBS), Pierre Marcus (LIP), Octave Hazard (LIP), Daria Pchelina (LIP), Joris Picot (LIP) and Gaëtan Bellot (CBS)
Salle : 04.05 TPR2
Abstract :
Curved shapes are ubiquitous in both natural and engineered structures contributing to their intricate functionalities and mechanical resilience. Replicating these shapes at the nanoscale using DNA nanotechnology poses significant challenges due to the inherent constraints of DNA geometry. Here, we introduce a geometric model for curved DNA helices and an algorithm for automatic routing of DNA helices along non-planar trajectories to fold predefined 3D DNA origami nanostructures. We provide an automated design process that enables the self-assembly of curved DNA helix bundles, hollow shapes, nested spheres, and biomimetic structures such as vault-like cages, using a novel spiral-based paradigm. This process, integrated in ENSnano software, revisites DNA origami principles to go beyond current structural confinements and opens opportunities for creating general 3D spatial configurations with advanced programmability for enhanced functional integration. L’exposé sera spécialement conçu pour des personnes qui ne connaissent rien du tout à l’ADN ni au origami ADN.
Seminaire : Séminaire cybersécurité : Privacy-Preserving Data Processing at Scale: How Much Can You Trust Your Cloud Provider?
21/01/2025 à 10h30
Séminaire cybersécurité : le mardi 21/01 à 10h30 en salle 4.05 TPR2 LuminySpeaker: Pascal Felber, Université de Neuchâtel
Title : ’‘ Privacy-Preserving Data Processing at Scale: How Much Can You Trust Your Cloud Provider?’‘
Abstract : The processing of large amounts of data requires significant computing power and scalable architectures. This trend makes the use of Cloud computing and off-premises data centres particularly attractive but exposes companies to the risk of data theft. This is a key challenge toward outsourcing data processing to external Cloud providers, as data represents for many companies their most valuable asset. In this talk, we will discuss recent and emerging mechanisms to support privacy-preserving data processing, i.e., confidential computing, on untrusted architectures.
Seminaire : Séminaire ACRO : Guilherme Gomes (20/01, 14h00, REU 4.05)
20/01/2025 à 14h00
Guilherme Gomes (Federal University of Minas Gerais, Brasil) Title: Matching Multicut: Parameterized Complexity and Enumeration 20/01/2025 14h00, salle REU 04.05 (LIS Luminy) The Matching Cut problem asks if a graph admits a partition of its vertex set in two such that no vertex has more than one neighbor across the cut. This problem has drawn considerable attention of the algorithms and complexity community in the last decade, becoming a canonical example for parameterized enumeration algorithms and kernelization. In this talk, we discuss one of its generalizations: Matching Multicut - can we partition the vertex set of a graph in at least ℓ parts such that no vertex has more than one neighbor outside its part? In particular, we will present some results on cubic graphs and more generally on parameterized complexity, the latter focused on enumeration kernelization. Joint work with Emanuel Juliano, Gabriel Martins, and Vinicius Santos.Seminaire : Séminaire ACRO : Florian Galliot (13/01/2025, 10h00, REU 4.05)
13/01/2025 à 10h00
Florian Galliot (I2M, Aix-Marseille Université) Title: Graph reconstruction from queries on triples 13/01/2025 10h00, salle REU 04.05 (LIS Luminy) We consider the problem of reconstructing a graph G from a query Q which, for every k-subset of vertices S, provides some information Q(G)(S) about the induced subgraph G[S]. The vertices are labelled from 1 to n, so reconstruction up to isomorphism is not sufficient, we need to know which label is where. We have studied the case k=3. Given a query Q on triples, we are interested in two things: a structural characterization of all graphs G that are uniquely reconstructible from the function Q(G) (i.e. such that Q(H)=Q(G) if and only if H=G), and a polynomial-delay enumeration algorithm of all graphs that are consistent with some input query answers. In 2023, Qi and Bastide et al. respectively have managed this for the connectivity query (meaning that, for every triple S, Q(G)(S) indicates whether G[S] is connected). We have obtained the same results for all 13 other non-trivial queries on triples. This presentation will go into details of a select few of these queries. Joint work with Hoang La, Raphaëlle Maistre, Matthieu Petiteau and Dimitri Watel.Seminaire : Séminaire CANA : Antoine Soulas
07/01/2025 à 14h00
Title : Quantifying quantum coherence and the deviation from the total probability formulaSalle : 04.05 TPR2
Abstract :
Quantum coherence is the main resource exploited by quantum computers. Unsurprisingly, over the past few years, there has been a strong interest in the task of finding appropriate measures of coherence. In this talk, we propose a novel approach to quantify quantum coherence which, contrary to the previous ones, does not rely on resource theory but rather on ontological considerations. In this framework, coherence is understood as the ability for a quantum system's statistics to deviate from the total probability formula (TPF). After a recap on the basics of quantum theory, we motivate the importance of the TPF in quantum foundations. We then propose a new set of axioms that a measure of coherence should satisfy, and show that it defines a class of measures different from the main previous proposal. Finally, we prove a general result about the dependence of the l2-coherence norm on the basis of interest, namely that it is well approximated by the square root of the purity in most bases. Such a behaviour is actually expected for any measure of coherence, because of the mathematical phenomenon known as « concentration of measure ».