ED Mathématiques et Informatique
Concept Drift in machine Learning : Detection, Update and Correction
by Maxime FUCCELLARO (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 14h00 - Amphitéatre 351, cours de la Libération F-33405 Talence cedex
in front of the jury composed of
- Akka ZEMMARI - Professeur des universités - Université de Bordeaux - Directeur de these
- Laurent SIMON - Professeur des universités - Labri - CoDirecteur de these
- Fleur MOUGIN - Professeure des universités - Labri - Examinateur
- Astrid JOURDAN - Professeure des universités - CY Tech école d'ingénieurs Pau - Examinateur
- Alexandre TERMIER - Professeur des universités - IRISA - Rapporteur
- Arnaud LALLOUET - Professeur des universités - Huawei Technologies - Rapporteur
The widespread adoption of Artificial Intelligence has led the number of machine learning (ML) models created and used in production to exponentially grow. The ability of ML models is constantly improving, thus allowing models to continuously reach new domains of application. The core definition of Concept Drift (CD) is of a change in data distribution. Several categories of drift exist based on amplitude, speed or recurrence. CD can appear naturally over the course of a long time period. An example is the way we travel: models used to estimate public transport delays will not perform well in ten years. CD can also be abrupt and promptly render predictive models obsolete, several examples can be found with the covid pandemic. As the probability that a model can be affected by CD increases with the model's age, concept drift is becoming a common phenomenon that needs to be considered before a model is built. CD can be split into two main categories, real drift and virtual drift. Real drift is said to happen when a predictive model's ability to correctly classify samples is negatively impacted by the distribution change. A drift is said to be virtual when the change of distribution has little to no impact on the model's predictive performances. When a drift is virtual, it must not be detected. CD has received a lot of attention in the past few years. However, the focus of most of the research has been on detecting real drift when true labels or ground truth is known. Such algorithms are based on the premise that drift happens when the model's predictions deteriorate. Various windowing techniques combined with statistical tests are applied to precisely find drift events. These methods can achieve almost perfect detection but do rely on the true labels to be available immediately after the inference phase. This is rarely the case in real world scenarii where class labels can take a long time to arrive or are never known. Some drift detecting algorithms work without requiring the true class labels, but often mistake virtual drift for real drift causing unnecessary and costly retraining. The focus of our work consists of creating drift detectors algorithms that do not require true labels to function while not mistakenly detect virtual drift as real drift. Thus, the contributions of the thesis are the following: (i) Partially Supervised Drift Detector (PSDD), an algorithm which partition the feature space using decision trees to find distribution changes using a statistical test; (ii) Real Drift Detector (RDD) improves PSDD by modifying the partitioning mechanism, applying several filters and by using different tests and by setting a user parameter that enables finer control over misdetections; (iii) Gaussian Split Detector (GSD), an algorithm that detects real drifts based on the assumption that some variables follow a Gaussian distribution using the Expectation Maximization algorithm and a Gaussian split criterion which enables to precisely find the amplitude of a drift event. All of our contributions manage to positively detect real drift events while not detecting virtual drift. Our contributions have been thoroughly tested against the state of the art on several drift datasets. Our testing methodology allowed us to evaluate real drift detection and on false positives created by virtual drift.
ED Sciences Chimiques
Beyond conventional synchrotron-based X-ray spectroscopy: fundamental insights and application to positive electrodes for metal-ion batteries
by Jazer Jose TOGONON (ICMCB - Institut de Chimie de la Matière Condensée de Bordeaux)
The defense will take place at 14h00 - ICMCB, Amphitheater 87 Avenue du Dr Albert Schweitzer, 33600 Pessac
in front of the jury composed of
- Laurence CROGUENNEC - Directrice de recherche - Université de Bordeaux, Institute of Condensed Matter Chemistry of Bordeaux (ICMCB) - CNRS - Directeur de these
- Sandrine LYONNARD - Directeur de recherche - IRIG - CEA - Grenoble - Rapporteur
- Iztok ARCON - Professor - University of Nova Gorica - Rapporteur
- Jean-Noël CHOTARD - Maître de conférences - Université de Picardie Jules Verne, Laboratoire de Réactivité et Chimie des Solides (LRCS) - CoDirecteur de these
- Antonella IADECOLA - Ingénieure de recherche - Sorbonne Université, PHysicochimie des Electrolytes, Nanosystèmes InterfaciauX (PHENIX) - Examinateur
- Lorenzo STIEVANO - Professeur - Université de Montpellier, Institut Charles Gerhardt Montpellier (ICGM) - Examinateur
- Cyril AYMONIER - Directeur de recherche - Université de Bordeaux, Institute of Condensed Matter Chemistry of Bordeaux (ICMCB) - CNRS - Examinateur
- Alessandro LONGO - Directeur de recherche - Istituto per lo Studio dei Materiali Nanostrutturati (ISMN) - CNR, UOS Palermo - Examinateur
The design of positive electrode materials for metal-ion batteries has been driven towards the goal of having enhanced safety and achieving high capacity and energy density. These targets mainly rely on the capacity of the electrode material to exchange Li+ ions (or Na+ and K+ ions) quickly and reversibly while operational at larger potential windows. The control of the composition and stoichiometry, as well as the crystal structure, of these electrode materials is a crucial point in design to manifest improved electrochemical properties and higher structural stability for metal-ion batteries. Hence, a holistic understanding of the structure and property relationship is essential. Various positive electrode materials, including polyanionic compounds and layered transition metal oxides for metal-ion batteries, are investigated using novel characterization techniques utilizing large-scale facilities, particularly synchrotron radiation sources. This study presents a comprehensive approach to explain the structure-property relationships of the positive electrode materials by combining structural analysis through X-ray diffraction measurements with hard X-ray spectroscopy studies. Mainly, the project explores the electronic structure of various battery chemistries using beyond conventional X-ray spectroscopic techniques like, high-energy resolution fluorescence-detected X-ray absorption near-edge structure (HERFD-XANES) spectroscopy, non-resonant X-ray emission spectroscopy (XES), and X-ray Raman scattering (XRS) spectroscopy. Specifically, the geometric and electronic distinctions between Tavorite-type LiVPO4F and KTP-type KVPO4F are analyzed using HERFD-XANES and XES. Additionally, the ligand and electronic environment in the mixed polyanionic compound KVPO4F1-xOx (x = 0, 0.25, 0.5, 0.75, 1) are investigated using non-resonant valence-to-core XES. Meanwhile, a combination of hard X-ray spectroscopy techniques is employed to probe the strong covalent nature of Ni-O bonds in layered LiNiO2. Finally, both ex situ and operando X-ray spectroscopy measurements are used to examine the cationic and anionic redox present in Li/Mn-rich layered oxides. Overall, the application of multiple synchrotron-based X-ray spectroscopy techniques is intended broadly to improve fundamental material understanding and future development strategies.
ED Sciences de la Vie et de la Santé
Development of a novel biohybrid polymersome γδT-cell conjugated technology : Combination of cell therapy and drug delivery
by Selma CORNILLOT-CLEMENT (Immunologie Conceptuelle, Expérimentale et Translationnelle)
The defense will take place at 14h00 - Amphithéatre LOUIS ISPED, 146 rue Léo Saignat, 33076 Bordeaux cedex
in front of the jury composed of
- Mary POUPOT - Chargée de recherche - Université de Toulouse - Rapporteur
- Harm-Anton KLOK - Professeur des universités - Ecole Polytechnique Fédérale de Lausanne (EPFL) - Rapporteur
- Emmanuel SCOTET - Directeur de recherche - Université de Nantes - Examinateur
- Julie DECHANET-MERVILLE - Docteure - Université de Bordeaux - Directeur de these
Development of a Bio-Hybrid Combining γδ T Lymphocytes and Polymersomes: Merging Cell Therapy and Drug Delivery Background: The combination of cell therapy and nanomedicine has advanced in recent years to improve the management of various pathologies. γδ T lymphocytes (γδ T cells), a cell population at the interface between innate and adaptive immunity, are involved in immune response to infectious agents and to cancers. They are divided into two main subpopulations: Vγ9Vδ2 T cells and non-Vγ9Vδ2 T cells, mainly composed of Vδ1 T cells. Vδ1 T cells have multiple activation mechanisms independent of antigen presentation by HLA and various effector functions, making them particularly attractive for overcoming the limitations associated with therapies that use more conventional populations such as αβ T cells. Additionally, these cells have tissue-targeting properties that make them promising vectors for developing nanoparticle-based therapies. Among the nanoparticles developed in recent years, polymeric nanoparticles, or polymersomes, stand out due to their many attractive features: a membrane with tunable size and composition, significant encapsulation capacity, and relatively simple formulation. Moreover, advances in metabolic engineering and click chemistry using cyclooctynes such as Dibenzocyclooctyne (DBCO) have facilitated their use in the development of new targeted therapies. Objective: This thesis aimed to develop a click chemistry protocol to create a biohybrid composed of DBCO-functionalized polymersomes and Vδ1 T cells (DOT) amplified and modified through metabolic engineering. Results: Initially, metabolic engineering and click chemistry protocols were developed using fluorescent DBCO. These protocols were sequentially optimized on a cell line, peripheral blood mononuclear cells (PBMCs), and then on the cells of interest, the DOT cells. Toxicity and functionality tests were conducted to assess the feasibility of this biohybrid. Subsequently, the same protocol was tested and optimized on DOT cells and empty polymersomes functionalized with DBCO. Perspectives: This thesis provides a first proof of concept for combining polymersomes with DOT cells. However, further validation tests are necessary to fully evaluate the feasibility of this biohybrid. These tests should particularly verify the activity of DOT cells against the target cells, the toxicity associated with click chemistry, and the potential benefits that this new therapeutic approach could offer in treating pathologies such as cancer and CMV infection.
ED Sociétés, Politique, Santé Publique
The non-emergence of Algeria's economy (2000-2023). Blockages of capitalist accumulation from the perspective of the relationships between economic and political elites.
by Faouzi ELMESTARI (Centre Emile Durkheim)
The defense will take place at 14h00 - Amphi Ellul Sciences Po Bordeaux 11 allée Ausone Domaine universitaire 33607 Pessac
in front of the jury composed of
- Antoine ROGER - Professeur des universités - Université de Bordeaux - Directeur de these
- Mylène GAULARD - Maîtresse de conférences - Université Grenoble Alpes - Rapporteur
- Myriam AïT-AOUDIA - Professeure des universités - Université de Picardie Jules Verne - CoDirecteur de these
- Andy SMITH - Directeur de recherche - Sciences Po Bordeaux - Examinateur
- Laurence LOUëR - Professeur associé - Sciences Po Paris - Examinateur
- Luis MARTINEZ - Directeur de recherche - Sciences Po-CERI - Rapporteur
In this research, we propose to clarify the socio-political factors that have hindered Algeria's economic emergence in recent times, despite the developmentalist ambitions of the political authorities. This thesis highlights the tensions and contradictions in which political power has been struggling since the end of the civil war (late 1990s), by, on the one hand, deploying public policies aimed at transitioning towards a productive economy and, on the other, maintaining institutional structures and economic orientations that jeopardize the rise of the industrial class. The Algerian regime seems to be driven by the deep-rooted and contradictory tendency to seek economic development without creating the conditions for the private sector to flourish; in a nutshell, it wants to build a dynamic capitalism without the formation of a powerful capitalist class. While political economy theories are of great value in understanding some of the forces behind economic emergence or non-emergence in the South, this research promotes the hypothesis that the essential factors determining the success or failure of late industrialisation are to be found in the socio-political sphere, and more specifically in the configuration of relations between the ruling class and the owning classes. Our research topic - the relationship between the State and entrepreneurs - is addressed using a qualitative methodology: around sixty interviews with private entrepreneurs (heads of large groups and SME owners), public company leaders and administration directors; collection of a set of data from the analysis of press articles, legal publications and statistics produced by major international institutions. The combination of our theoretical analyses and our empirical investigation shows that the socio-political conditions which have ensured the industrial successes of the South, namely collaborative, if not integrated, relationships between the higher spheres of government and the industrial sector, are not met in in the Algerian case. To date, Algeria has shown no evidence of such a developmental alliance crystallising within the ruling coalition. On the contrary, economic and political interests that are difficult to reconcile between the factions in power and the business leaders, which can sometimes turn into antagonism, are shaping a long-term structural tendency towards the separation of these two types of elites. This process of inter-elite fragmentation hampers capitalist accumulation, traps the country in a circle of economic stagnation, and most importantly, reduces heavily the future possibilities of an Asian-style development, meaning an industrialization driven from the top by the State. However, there is a bottom-up emergence dynamic at work in the productive sectors of the economy, triggering low-key transformations of economic and social structures. But, as political and administrative institutions undermine the rise of social forces for change, Algeria undoubtedly finds itself in historical circumstances whose political outcome remains largely uncertain: either a reproduction of existing social, political and institutional structures and the continuation of the economic stagnation, or popular organisation and mobilisation that forces the ruling coalition to adopt a genuine developmentalist project.