ED Mathématiques et Informatique
Analysis and development of new optimisers for machine learning
by Bilel BENSAID (IMB - Institut de Mathématiques de Bordeaux)
The defense will take place at 9h30 - Salle de Conférence Institut de Mathématiques de Bordeaux (A33) 351 Cours de la Libération 33405 Talence cedex
in front of the jury composed of
- Julien SALOMON - Directeur de recherche - INRIA et Sorbonne Université - Examinateur
- Pierre-Antoine ABSIL - Professeur des universités - Université Catholique de Louvain - Rapporteur
- Aude RONDEPIERRE - Professeure des universités - INSA de Toulouse - Rapporteur
- Nicolas BOUSQUET - Chargé de recherche - Sorbonne Université et EDF - Examinateur
- Rodolphe LE RICHE - Chargé de recherche - CNRS - Examinateur
- Raphael LOUBERE - Directeur de recherche - CNRS - Examinateur
- Laurent SIMON - Professeur des universités - Université de Bordeaux, Bordeaux-INP - Examinateur
Over the last few years, developping an explainable and frugal artificial intelligence (AI) became a fundamental challenge, especially when AI is used in safety-critical systems and demands ever more energy. This issue is even more serious regarding the huge number of hyperparameters to tune to make the models work. Among these parameters, the optimizer as well as its associated tunings appear as the most important leverages to improve these models HSIC. This thesis focuses on the analysis of learning process/optimizer for neural networks, by identifying mathematical properties closely related to these two challenges. First, undesirable behaviors preventing the design of explainable and frugal network are identified. Then, these behaviors are explained using two tools: Lyapunov stability and geometrical integrators. Through numerical experiments, the learning process stabilization improves the overall performances and allows the design of shallow networks. Theoretically, the suggested point of view enables to derive convergence garantees for classical Deep Learning optimizers. The same approach is valuable for mini-batch optimization where unwelcome phenomenons proliferate: the concept of balanced splitting scheme becomes essential to enhance the learning process understanding and improve its robustness. This study paves the way to the design of new adaptive optimizers, by exploiting the deep relation between robust optimization and invariant preserving scheme for dynamical systems.
ED Sciences Chimiques
Strategies for the elaboration of sustainable self-blown polyurethane foams based on the decarboxylation of cyclic-carbonates
by Florent MONIE (Laboratoire de Chimie des Polymères Organiques)
The defense will take place at 9h30 - Amphi 1 CRPP, 115 Avenue du Dr Albert Schweitzer, 33600 Pessac
in front of the jury composed of
- Henri CRAMAIL - Professeur - Université de Bordeaux - Directeur de these
- Haritz SARDON - Professeur - POLYMAT, University of the Basque Country UPV/EHU - Rapporteur
- Raquel VERDEJO - Research Scientist - Institute of Polymer Science and Technology (ICTP-CSIC) - Rapporteur
- Christophe DETREMBLEUR - Directeur de recherche - Université de Liège - CoDirecteur de these
- François TOURNILHAC - Directeur de recherche - Laboratoire Chimie Moléculaire, Macromoléculaire et Matériaux, Ecole Supérieure de Physique et Chimie Industrielles (ESPCI) - Examinateur
- Karin ODELIUS - Professeur - KTH, Royal Institute of Technology, Department of Fiber and Polymer Technology - Examinateur
This thesis consists in the development of self-foaming isocyanate-free polyurethane formulations, in particular poly(hydroxyurethane)-PHU, for which the blowing agent generation (CO2) originates from the decarboxylation of 5-membered cyclic carbonate (5CC) precursors. First, a system relying on a latent-thiol precursor (i.e. thiolactone) has been designed. The in-situ generated thiol allowed the PHU matrix expansion in a controlled and odorless way. In a second time, various waste biofillers were used as promotors of the foaming behavior thanks to the water they contain. The 5CC hydrolysis was then responsible for providing the foaming power. This led us to also investigate the impact of water on the curing kinetic of a non-foaming thermosetting formulation. Finaly, the developed foam formulations were scaled-up in order to undergo extensive characterization in an industrial context. This allowed to bring insights on the thermal insulation and fire resistance properties of PHU foams.
ED Sciences de la Vie et de la Santé
Contribution of the rat gustatory cortex to stimulus-guided action
by Yacine TENSAOUTI (Institut de neurosciences cognitives et intégratives d'Aquitaine)
The defense will take place at 14h00 - Salle de conférence NEUROCENTRE MAGENDIE - U1215 146, rue Léo saignat 33077 Bordeaux cedex France
in front of the jury composed of
- Shauna PARKES - Chargé de recherche - INCIA – UMR 5287 Université de Bordeaux / CNRS - Directeur de these
- Claire RAMPON - Directeur de recherche - Université Toulouse III Paul Sabatier, CNRS - Examinateur
- Fabien NANEIX - Maître de conférences - University of Aberdeen - Rapporteur
- Sebastien CARNICELLA - Directeur de recherche - Université Grenobles Alpes, U1216 INSERM - Rapporteur
Every day, individuals are faced with numerous decisions that shape their behavior. Typically, when faced with a decision between two courses of action, we and other animals will select the action that delivers the more valuable consequence or outcome; this is known as value-guided action selection. However, it is crucial to also recognize the impact of environmental stimuli on decision-making. The existence of such stimulus-guided action is well-established in both animals and humans. For instance, stimuli associated with food can not only direct us toward nourishment but also trigger cravings, even in the absence of hunger. Furthermore, such stimuli can contribute to maladaptive behaviors, such as drug-related stimuli triggering cravings and relapses in individuals battling addiction, or stimuli giving rise to hallucinatory experiences in individuals with schizophrenia. It is thus important that we understand how internal motivations and external stimuli interact is essential to drive behaviour. The cortical contribution to value-guided actions has been extensively studied, with the gustatory cortex (GC) of the insula identified as crucial for retrieving outcome value to guide choice. However, the cortical contribution to stimulus-guided action remains understudied and, notably, GC's role in this behaviour has never been explored. To uncover the role of the rat GC in stimulus-guided action, we used the Pavlovian-to-Instrumental Transfer (PIT) paradigm . This paradigm comprises three stages: a Pavlovian conditioning phase during which a stimulus-outcome association is learned, followed by an instrumental training phase during which animals learn an action-outcome association in the absence of any stimuli and, finally, a transfer test during which the effects of Pavlovian stimuli on action selection are assessed. The PIT paradigm comes in two forms: a general form (gPIT) where reward-predictive stimuli enhance action performance directed towards obtaining any desired food outcome, relying on the general motivational properties acquired by Pavlovian stimuli; and a specific form (sPIT) where Pavlovian stimuli specifically bias actions towards obtaining the same outcome as predicted by the stimulus, relying on the specific sensory properties of the food outcomes. Using a chemogenetic approach, we first demonstrated that GC inhibition during the general transfer test abolished the ability of Pavlovian reward-predictive stimuli to energize instrumental responding. We then similarly assessed the contribution of GC during the specific transfer test, and observed that Pavlovian stimuli were unable to specifically bias instrumental responding towards the action sharing the same outcome as the presented predictive stimulus. These results demonstrated for the first time the critical role of the GC in stimulus-guided choice, encompassing both the general motivational properties acquired by Pavlovian stimuli and their ability to specifically bias action selection towards specific outcomes. We also obtained preliminary data indicating that input from the GC to the mediodorsal thalamus is critical for the expression of specific PIT but not for general PIT. The findings demonstrate that the GC is essential for Pavlovian reward-predictive stimuli to exert both general and specific effects on action selection, with the latter relying on an intact cortico-thalamic pathway. This work extends our understanding of the GC's role in decision making, highlighting its involvement in stimulus-guided actions in addition to value-guided actions. These results align with human fMRI studies showing morphological changes in the insular cortex of nicotine-dependent individuals and insular activation in response to smoking-related cues during stimulus-induced craving paradigms. Collectively, this suggests that the GC could be a viable target for therapeutic interventions in patients with addiction or other related disorders.