ED Mathématiques et Informatique
Visualization for the explanation of deep neural networks
by Luc-Etienne POMME-CASSIEROU (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 14h00 - 050-AMPHI (198) Domaine universitaire, 351, cours de la Libération, 33405 Talence, Bâtiment A30
in front of the jury composed of
- David AUBER - Professeur des universités - Université de Bordeaux - Directeur de these
- Anne VIALARD - Maîtresse de conférences - Université de Bordeaux - Examinateur
- François QUEYROI - Chargé de recherche - Université de Nantes - Examinateur
- Aurélie BUGEAU - Professeure des universités - Université de Bordeaux - Examinateur
- Arnaud SALLABERRY - Professeur des universités - Université de Montpellier - Rapporteur
- Harold MOUCHERE - Professeur des universités - Université de Nantes - Rapporteur
The understanding of the world that surrounds us requires the observation and analysis of the phenomena that govern it. The observation of these phenomena requires the systematic collection of large quantities of data to build models representative of the reality. Eventually, the models are used to extract knowledge and properties about the object of study. The more complex the data, the more difficult it becomes for a human to find processing rules or concepts to extract. This is where AI (e.g. machine / deep learning) algorithms come in to help. The superior performance of these algorithms being no longer in question, they are widely used in matters where decision making is required. In addition to the worrying ecological footprint of these tools, the ethical aspect is also at stake. Indeed, these algorithms are often considered as “black boxes”, making arbitrary decisions without justification. The question of responsability, in case of bad decision arises. It becomes necessary to build systems enabling to understand the behavior (either good or bad) of such algorithms in order to improve the trust we can put in their decisions. This problem has given rise to a field of research called XAI (eXplainable AI), which is currently in full expansion. This thesis brings together a number of contributions enabling us to analyze the behavior of trained neural networks at several scales and from several angles, based on visualizations intended for both experts and non-expert users. More specifically, we propose two complementary ways of analyzing how a network works. The first focuses on the architecture of a neural network and its components. The aim of this method is to show how class discrimination evolves throughout the network, for a classification problem. The second approach studies the operation of a network from the point of view of the data it processes. On the scale of individual data, then on the scale of groups of data, a visualization method is proposed to illustrate the characteristics of the data that lead a network to make one prediction rather than another. These methods are also accompanied by a visualization to compare the performance of models on a finer scale: that of classes.
ED Sciences Chimiques
Synthesis and study of biomimetic enzymatic responsive reactors from water-in-water emulsions
by Léa WALDMANN (Institut des Sciences Moléculaires)
The defense will take place at h00 - Amphithéatre 1 16 Avenue Pey Berland, ENSMAC Bâtiment A, 33600, PESSAC
in front of the jury composed of
- Valérie RAVAINE - Professeure des universités - Bordeaux INP - Directeur de these
- Stéphane ARBAULT - Directeur de recherche - CNRS - CoDirecteur de these
- Florence AGNELY - Professeure des universités - Université Paris Saclay - Rapporteur
- Sophie GRIVEAU - Professeure des universités - Chimie ParisTech PSL - Rapporteur
- Lazhar BENYAHIA - Professeur des universités - Le Mans Université - Examinateur
- Chrystel FAURE - Professeure des universités - Bordeaux INP - Examinateur
For several years now, systems based on aqueous liquid phase separation (ATPS) have been identified as a possible compartmentalization route for producing biomimetic synthetic cells. This phase separation enables the formation of domains with different compositions and thus different polarities, which can modify the distribution of species in the aqueous media and be the site of biochemical reactions, especially enzymatic ones. Water-in-water emulsions are a type of ATPS, formed by two water-soluble but non-compatible polymers that segregate within each phase. Stabilizing water-in-water emulsions is a real challenge, due to the very low interfacial tensions between the two phases and the thickness of the interface, which extends over several nanometers. In this work, we present a new type of Pickering-type stabilizer, bis-hydrophilic and temperature-sensitive microgels, able to stabilize the prototypical emulsion composed of dextran (Dex) and poly(ethylene oxide) (PEO). New microgels incorporating in the same structure poly(N-isopropylacrylamide) (pNIPAM) chains, with affinity for the PEO phase, and Dex chains, were synthesized with different Dex/NIPAM ratios. Microgels with a higher Dex content showed excellent stabilizing properties for emulsions by adsorbing at the droplet surface, demonstrating the fundamental role of bis-hydrophilicity. In a second step, charges were incorporated into these particles, via the addition of acrylic acid groups. The microgels then became pH-sensitive, and their hydrodynamic diameter varied as a function of the acidity of the medium. These different aqueous compartments were then used to study their effects on enzymatic reactions. Glucose Oxidase (GOx) was chosen as the model enzyme, because of its robustness and ability to function under physiological conditions. It catalyzes the oxidation of glucose by oxygen to form gluconolactone and hydrogen peroxide (H2O2). The enzyme activity was analyzed in different phases by detecting the produced H2O2, firstly using the Amplex Red reagent, which generates a fluorescent product after oxidation (catalyzed by a peroxidase), and then in greater detail using microelectrochemistry, directly detecting the local H2O2 concentration by oxidation. The enzymatic reaction and its kinetics were studied within polymer phases and within emulsion droplets, in order to understand the impact of ATPS systems on the physico-chemistry and on the dynamics of an enzymatic reaction.
ED Sciences Physiques et de l'Ingénieur
Biohybrid neuro-cardiac platform for an electroceutical approach
by Pierre-Marie FAURE (Laboratoire de l'Intégration du Matériau au Système)
The defense will take place at 9h30 - Amphi Dom Laboratoire IMS, 351 Cours de la Libération, 33405 Talence Cedex, France
in front of the jury composed of
- Timothée LEVI - Professeure des universités - Université de Bordeaux - Directeur de these
- Guilhem LARRIEU - Directeur de recherche - LAAS-CNRS - Examinateur
- Noëlle LEWIS - Professeure des universités - Université de Bordeaux - Examinateur
- Michela CHIAPPALONE - Professor - University of Genoa - Rapporteur
- Takashi KOHNO - Professor - Institute of Industrial Sciences, The University of Tokyo - Rapporteur
- Dominique DALLET - Professeur - Bordeaux INP - Examinateur
Today, cardiovascular disease is the world's leading cause of death, and its prevalence will continue to rise in the years to come due to changing lifestyles and demographics. Among these diseases, heart failure affects the heartbeat generation mechanism. This involves the nervous system, which interacts with the cardiac conduction system to generate the heartbeat. To improve the study of these diseases and lay the foundations for a therapeutic solution, a platform capable of reproducing the mechanisms involved in a healthy heart is being developed during this thesis. This solution is based on the most plausible cellular models possible, to target a broad spectrum of biological phenomena, while incorporating real-time computing capability. This methodology makes it possible to insert the platform into a bio-hybrid environment, where the platform's inputs and outputs can be of either biological or artificial origin. In this way, the platform can be incorporated into a multitude of environments, making it easy to deploy. Its implementation also provides a replicable and malleable environment for studying the functioning of the neurocardiac axis and for integration in therapeutic solutions over the long-term. To demonstrate this set of capabilities, tests were carried out using the electrical activity of neural networks as input to mimic the nervous system, while cardiomyocytes were used as output to act as the heart. These experimental setups were designed to approximate the functioning of the neurocardiac axis as it exists in the body.