abstract: Recently, a well-designed and well-trained neural network can yield state-of-the-art results across many domains, including data mining, computer vision, and medical image analysis. But progress has been limited for tasks where labels are difficult or impossible to obtain. This reliance on exhaustive labeling is a critical limitation in the rapid deployment of neural networks. Besides, the current research scales poorly to a large number of unseen concepts and is passively spoon-fed with data and supervision. To overcome the above data scarcity and generalization issues, in my dissertation, I first propose two unsupervised conventional machine learning algorithms, hyperbolic stochastic coding, and multi-resemble multi-target low-rank coding, to solve the incomplete data and missing label problem. I further introduce a deep multi-domain adaptation network to leverage the power of deep learning by transferring the rich knowledge from a large-amount labeled source dataset. I also invent a novel time-sequence dynamically hierarchical network that adaptively simplifies the network to cope with the scarce data. To learn a large number of unseen concepts, lifelong machine learning enjoys many advantages, including abstracting knowledge from prior learning and using the experience to help future learning, regardless of how much data is currently available. Incorporating this capability and making it versatile, I propose deep multi-task weight consolidation to accumulate knowledge continuously and significantly reduce data requirements in a variety of domains. Inspired by the recent breakthroughs in automatically learning suitable neural network architectures (AutoML), I develop a nonexpansive AutoML framework to train an online model without the abundance of labeled data. This work automatically expands the network to increase model capability when necessary, then compresses the model to maintain the model efficiency. In my current ongoing work, I propose an alternative method of supervised learning that does not require direct labels. This could utilize various supervision from an image/object as a target value for supervising the target tasks without labels, and it turns out to be surprisingly effective. The proposed method only requires few-shot labeled data to train, and can self-supervised learn the information it needs and generalize to datasets not seen during training. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2020

La BPCO (bronchopathie pulmonaire chronique obstructive) est un problème majeur de santé publique et représentera la 3ème cause de mortalité dans le monde en 2030. L’âge, le tabagisme, ainsi que la pollution atmosphérique via l’exposition aux particules de diesel mais également la pollution domestique – majoritairement représentée par la combustion domestique de biomasse – sont des facteurs de risque bien identifiés d’apparition d’une BPCO. Il n’existe à ce jour aucun traitement curatif pouvant interférer avec l’histoire naturelle de la maladie.Les cellules souches pluripotentes, et notamment les cellules souches humaines pluripotentes induites (hiPSCs), sont définies par deux propriétés fondamentales : l’auto-renouvellement et la capacité à se différencier en tous les types cellulaires de notre corps. Elles offrent une opportunité sans précédent de modéliser le développement humain normal et pathologique de l’appareil respiratoire.Ce projet de recherche a pour objectif de modéliser in vitro les trajectoires de la BPCO, en lien avec une origine développementale (racines pédiatriques) et/ou une susceptibilité au tabac. Afin d’élucider les mécanismes qui sous-tendent la pathogénie de la BPCO et de la susceptibilité au tabac, nous avons constitué deux groupes caricaturaux : i) 4 patients atteints d’une forme sévère de la BPCO, constituant le groupe « hautement susceptibles », ii) 4 patients fumeurs indemnes de BPCO ou tout autre comorbidité liée au tabac « hautement résistants » au tabac.Nous avons utilisés deux modèles de culture cellulaires in vitro : les hiPSCs et la culture de cellules épithéliales primaires bronchiques humaines (HBECs) cultivées en ALI (interface air liquide).Dans un premier temps, nous avons généré des lignées hiPSCs par reprogrammation cellulaire à partir du sang périphérique d’un sujet sain (contrôle), et de trois patients BPCO sévères hautement caractérisés. Dans un second temps, la différenciation dirigée des hiPSCs a permis de récapituler le développement pulmonaire précoce (génération de progéniteurs bronchiques NKX2.1) par la mise au point d’un protocole de différenciation dirigée robuste et reproductible sur plusieurs lignées hiPSCs. La maturation de ces progéniteurs bronchiques en culture 2D ou 3D a permis d’obtenir des structures épithéliales exprimant les marqueurs de cellules basales (KRT5), de cellules Club (CCSP), et ciliées (FOXJ1). Dans un second temps, ces épithélia seront exposés au tabac (CSE- cigarette smoke extract) afin d’induire un phénotype « BPCO-like ». Enfin, la culture des HBECs cultivées en ALI des patients BPCO sévères a été réalisée en condition exposée (CSE) et non exposée. La résistance transépithéliale, la motilité ciliaire, le profil sécrétoire et la diversité ARN ont été collecté.Ce travail a permis de mettre en place les outils nécessaires pour reproduire les trajectoires in vitro de la BPCO et élucider les origines de la pathologie. Les outils de séquençage à haut débit (transcriptomique dans notre étude), permettront de découvrir de nouveaux candidats, représentants de potentielles cibles en vue d’un criblage pharmacologique. / COPD (Chronic Obstructive Pulmonary Disease) is a major public health problem and will be the 3rd leading cause of death in the world in 2030. Age, smoking, and air pollution through the exposure to particulate matter but also domestic pollution - mostly represented by domestic biomass combustion - are well-identified risk factors for the development of COPD. To date, there is no cure that can interfere with the natural history of the disease.Pluripotent stem cells, including induced pluripotent human stem cells (hiPSCs), are defined by two fundamental properties: self-renewal and the ability to differentiate into all cell types in our body. They offer an unprecedented opportunity to model the normal and pathological human development of the respiratory system.This research project aimed to model in vitro the trajectories of COPD, related to a developmental origin (pediatric roots) and / or susceptibility to tobacco. In order to elucidate the underlying mechanisms of COPD and tobacco susceptibility, we established two extreme groups: i) 4 patients with a severe form of COPD, the "highly susceptible" group, ii) 4 patients who are free of COPD or other tobacco-related comorbidity despite heavy smoking, called as "highly resistant" to tobacco.We have used two different but complementary in vitro cell culture models: hiPSCs and human bronchial primary epithelial cell cultures (HBECs) grown in ALI condition (Air Liquid Interface).First of all, we generate hiPSCs cell lines by reprogramming cells from peripheral blood of a healthy subject (control), and three highly characterized severe COPD patients. In a second step, the directed differentiation of hiPSCs allowed to recapitulate the early pulmonary development (NKX2.1 generation of bronchial progenitors) by the development of a robust and reproducible directed differentiation protocol of several hiPSCs lines. The maturation of these bronchial progenitors in 2D or 3D culture allows the generation of epithelial structures expressing markers of KRT5 + basal cells , CSSP + Club cells and FOXJ1 + ciliated cells. In a second step, these epithelia will be exposed to tobacco (CSE-cigarette smoke extract) in order to induce a "COPD-like" phenotype. Finally, ALI culture of HBECs of severe COPD patients was performed in unexposed and exposed condition (CSE). Transepithelial resistance, ciliary motility, secretory profile, and RNA diversity were collected.This work allowed to put in place the necessary tools to reproduce the in vitro trajectories of COPD and to clarify the origins of this pathology. The high throughput sequencing tools (transcriptomic in our study), will allow the discovery of new candidates, that represent potential targets for future pharmacological screening.