Au cours de la dernière décennie, des progrès significatifs ont été réalisés dans le domaine de l’IA, principalement grâce aux progrès de l’apprentissage automatique, de l’apprentissage profond et de l’utilisation de modèles à grande échelle. Cependant, à mesure que ces modèles évoluent, ils présentent de nouveaux défis en termes de gestion de grands ensembles de données et d’efficacité informatique. Cette thèse propose des approches pour réduire les coûts de calcul de la formation et de l’inférence dans les systèmes d’intelligence artificielle (IA). Plus précisément, ce travail étudie les techniques d’apprentissage continu et de calcul adaptatif, démontrant des stratégies possibles pour préserver les niveaux de performance de ces systèmes tout en réduisant considérablement les coûts de formation et d’inférence. Les résultats du premier article montrent que les modèles de base peuvent être continuellement pré-entraînés grâce à une méthode d’échauffement et de relecture, ce qui réduit considérable- ment les coûts de calcul de l’entraînement tout en préservant les performances par rapport à un entraînement à partir de zéro. Par la suite, la thèse étudie comment les stratégies de calcul adaptatif, lorsqu’elles sont combinées avec la mémoire, peuvent être utilisées pour créer des agents d’IA plus efficaces au moment de l’inférence pour des tâches de raisonnement complexes, telles que le jeu stratégique de Sokoban. Nos résultats montrent que les modèles peuvent offrir des per- formances similaires ou améliorées tout en utilisant beaucoup moins de ressources de calcul. Les résultats de cette étude ont de vastes implications pour l’amélioration de l’efficacité in- formatique des systèmes d’IA, soutenant à terme le développement de technologies d’IA plus abordables, accessibles et efficaces. / Over the past decade, significant progress has been made by the field of AI, primarily due to advances in machine learning, deep learning, and the usage of large scale models. However, as these models scale, they present new challenges with respect to handling large datasets and being computationally efficient. This thesis proposes approaches to reducing computational costs of training and inference in artificial intelligence (AI) systems. Specifically, this work investigates how Continual Learning and Adaptive Computation techniques can be used to reducing training and inference costs while preserving the perfor- mance levels of these systems . The findings of the first article show that foundation models can be continually pre-trained through a method of warm-up and replay, which significantly decreases training computational costs while preserving performance compared to training from scratch. Subsequently, the thesis investigates how adaptive computation strategies, when com- bined with memory, can be utilized to create more computationally efficient AI agents at inference time for complex reasoning tasks, such as the strategic game of Sokoban. Our results exhibit that models can deliver similar or improved performances while using signifi- cantly fewer computational resources. Findings from this study have broad implications for improving the computational efficiency of AI systems, ultimately supporting the development of more affordable, accessible, and efficient AI technologies.

There has been significant increase in the occurrence of earthquakes and other disasters around the world recently. With record-breaking storms and natural disasters destroying different parts of the world, many people have to deal with the repercussions of the aftermaths; and homelessness is one of the facts of such incidence. This master thesis is aimed at creating living solution for the victims of earthquakes and other kind of disasters; natural or human-made (war), providing accommodation for homeless people after an earthquake or a housing alternative for refugees. The design of the collapsible home is basically a new concept of emergency hut for accommodating people in emergency after the event of disaster with easy assembling and quick installation feature. Considering the situation of homeless people after catastrophic disasters who need to be accommodated, the design and creation of a quick, accommodating and low-cost emergency hut is advantageous.  This project also provides a living service for both refugees and homeless people. It is low cost, modular, and easy to use. More than that it is feasible and it makes it possible to accommodate people for a longer period of time, rather than a very short temporary housing solution such as tents. / Det har skett en signifikant ökning av jordbävningar och andra katastrofer i världen. Med rekordstora stormar och naturkatastrofer som förstör i delar av världen måste många människor nu hantera efterverkningar och återuppbyggande, och hemlöshet är en av de faktorer som uppstår och som påverkar människor mest. Det här examensarbetet syftar till att skapa en boendelösning för offer för jordbävning och andra naturkatastrofer och krig. Det hopfällbara hemmets design är ett koncept för krisboende med enkelt montage och installering. Om man betänker den situation som människor hamnar i som blir hemlösa vid en katastrof, kan en design av en snabb och billig boendelösning som krisboende ge enormt mycket. Projektet kan också vara en boendeservice till flyktingar och människor som är hemlösa av andra anledningar än katastrofer. Det tillverkas för en låg kostnad och är möjlig att använda som modul och enkel att sätta upp. Men viktigare än det är att den ger möjlighet att härbergera människor under en längre tid till skillnad från mer temporära lösningar som exempelvis tält. Det hopfällbara hemmet är gjort av hållbart material, den använder grön energi och lyser upp staden. Den är utrustad med vertikala trädgårdar för att användarna ska kunna odla egna grönsaker och den återvinner vatten från regn. Den innehåller basala hopfällbara möbler och rumsavdelare som ger känsla av hemtrevnad och bekvämlighet.

台灣位處環太平洋地震帶，為全球地震風險潛勢較高的地區之一；因台灣高科技產業蓬勃發展，地震保險需求殷切，再加上開放保險費率自由化的政策及金控效應，趨使保險業間競爭白熱化，惡性價格競爭及保險經紀人的推波助瀾，保險公司的清償能力面臨重大考驗。 台灣在保險監理方面如同日本、美國採行風險基礎資本額（RBC）制度，惟國際間位處高度地震風險潛勢之國家大都另建立一套地震保險監理機制，以確保保險公司的巨災準備金足以支付回歸期地震所造成的損失，其中以美國加州及加拿大政府均採用地震保險PML申報制度作為地震保險監理之依據最值得台灣學習。 地震保險PML評估可採用CRESTA Zone平均損失幅度表計算或採用認許的地震風險評估電腦軟體推估獲得，實施的關鍵為主管機關應建立具有公信力的CRESTA Zone平均損失幅度表。本文特就二種評估方式的利弊做深入的比較分析，並藉由地震風險評估軟體的架構說明影響地震保險PML的因素與權重，作為保險公司落實地震風險管理之依據。 本研究參考Solvency II的三大支柱提出建立地震保險監理機制之結論與建議如下： 1.鼓勵建立保險公司的地震風險管理機制。 2.公佈CRESTA Zone平均損失幅度表，作為保險公司地震保險PML申報依據，以落實產物保險業之地震保險監理機制。 3.依據保險公司申報資料，提供保險主管機關實施差異化管理之依據。 4.主動揭露經營績效、強化保險市場紀律，建立公平合理的經營環境。 期待藉由建立適當的地震保險監理機制，減輕或消弭產物保險市場面臨自由化的惡性價格競爭與保險經營面的不合理現象，進而達到健全保險經營環境、促進保險業長期穩定發展，並確保社會大眾之保險權益的目標。 關鍵詞：地震保險監理機制、地震保險PML、巨災準備金、風險基礎資本額、地震風險評估軟體 / Located at the Pacific Rim earthquake zone, Taiwan has been recognized as one of the severe seismic hazard areas in the world. With the bloom of high tech industry in the past two decades, the demand of earthquake insurance has been considerably increasing. However, along with the liberalization of insurance market, the new business model of financial holdings and the expanding influence from international brokers, insurance companies’ solvency capacity has been significantly challenged. Taiwanese Government, same as Japan and U.S., adopts Risk-Based Capital (RBC) method in insurance supervision, while most countries with high earthquake potential have set up independent earthquake insurance supervision systems to ensure insurers’ earthquake reserves capable to compensate the huge earthquake losses. Among all the measures, the PML reporting system adopted by Canada and the State of California to regulate and trace insurance companies’ financial statuses could be an adequate paradigm for Taiwan. The PML estimation could be obtained either using computer models or following default mean damage ratio table. This research compares the strength and weakness between these two methods, and presents the importance of parameters and key points in earthquake insurance management. Based on the three pillars of Solvency II, the conclusions and recommendations of this paper are: (1)Encourage insurance companies to build up the earthquake risk management mechanism; (2)Establish the official default mean damage ratio table for PML reporting system; (3)Adopt differential supervision practice to different level insurance companies; (4)Promote the self-disclosure of key business information and enhance market discipline. Establishing a sound earthquake insurance supervision system would not only ease the immoderate low-price competition but the whole insurance environment could also be stabilized and improved. It will ultimately achieve the objective to insure society liability and benefit the public as well. Keywords: Earthquake Insurance, Earthquake Model, Catastrophic Risk Management, Insurance Supervision, Risk-Based Capital, CRESTA Zone, PML

Cette thèse propose une modélisation de la dynamique psychique à l’œuvre dans les contextes d’anorexie. A partir d’une approche processuelle, elle explore la continuité psychodynamique entre les anorexies du bébé-dans-sa-famille et les anorexies mentales de l’adulte. La première partie est consacrée à une revue de la littérature psychanalytique considérant les travaux portant sur l’anorexie mentale de l’adolescente et de l’adulte, les travaux consacrés aux anorexies du bébé et les travaux traitant des anorexies aux différents âges de la vie. La seconde partie présente les fondements épistémologiques de cette recherche. Elle ouvre une réflexion sur la recherche en psychologie clinique et sur ses liens avec la pratique et l’enseignement. Elle précise ma filiation théorique et praticienne. La troisième partie est consacrée à la méthodologie de la recherche clinique, et principalement aux spécificités de l’observation clinique psychanalytique. Elle présente la méthode E. Bick d’observation du bébé dans sa famille, ses apports à la recherche et à la pratique cliniques ainsi que l’adaptation réalisée auprès des patients souffrant d’anorexies. La quatrième partie propose une réflexion sur les enjeux psychiques de la « relation de nourrissage ». Cette relation est envisagée comme une relation commensale (W.R. Bion, 1962), paradigme de la rencontre intersubjective, de la croissance psychique de la subjectivité du bébé et de l’essor de la parentalité, mais aussi terrain privilégié de la transmission psychique inconsciente. La cinquième partie est consacrée à la mise en dialogue entre le bébé-dans-sa-famille et les aspects bébés du self de l’adulte dans les contextes d’anorexie. Les anorexies sont envisagées en terme de refus-impossibilité alimentaire témoignant d’une tentative d’organisation de modalités de survie psychique. Deux angoisses communes apparaissent au centre de la dynamique psychique : une angoisse catastrophique et une angoisse de persécution. L’angoisse catastrophique prend la forme d’une « chute sans fin désintégrante », signe d’une dépression primaire. Elle est surchargée par une angoisse de vampirisation-dévoration qui est notamment envisagée comme une forme d’objectalisation de l’angoisse de « chute sans fin désintégrante ». Les modalités de défenses adhésives ont pour fonction de lutter contre la chute sans fin désintégrante. Elles sont accompagnées d’une inhibition de la vie pulsionnelle —plus particulièrement de l’avidité secondaire— ainsi que de son clivage, son déni et sa projection. Le contact avec la vie émotionnelle du bébé réactive la dépression primaire contre laquelle les aspects bébés du self parental se sont organisés dans une économie de survie psychique. Le refus-impossibilité alimentaire du bébé apparaît en miroir d’un refus-impossibilité d’accueil et de mise en sens de la vie émotionnelle qui caractérise la parentalité anorexique. Le lien entre le bébé et la parentalité, mais aussi entre les aspects bébés du self et la parentalité interne, est caractérisé par une réflexivité opaque et persécutoire intériorisée par le bébé —et les aspects bébés du self— sous la forme d’un cercle pernicieux et involutif. Ce travail se conclut sur la proposition d’un gradient du lien de nourrissage structuré par deux pôles. Le plus intégré, celui de la relation de nourrissage, est caractérisé par une relation d’intimité entre le bébé et la parentalité et l’introjection d’une relation de réflexivité commensale, mutuelle et asymétrique, soutenant l’intégration pulsionnelle, le développement des bases d’un surmoi protecteur et la croissance psychique. Le second est défini comme un lien d’anourrissage, caractérisé par une relation d’ex-timité entre le bébé et la parentalité, l’intériorisation d’un réflexivité opaque et persécutoire sous la forme d’un cercle involutif entraînant la désintrication pulsionnelle et le développement d’un « surmoi sévère et destructeur du moi » (W.R. Bion, 1959). / This dissertation offers a modelisation of the psychological dynamics which are at work in the contexts of anorexia. Process is the cornerstone to the exploration of the ongoing psychodynamic that is at work with the anorexia of the baby-in-its-family and the anorexia nervosa of the adult. The first part reviews the existing psychoanalytical literature and considers the works dealing with the anorexia nervosa of the teenager and of the adult, then the works dedicated to the anorexias of the baby; last but not least, the works offering an approach that considers the anorexias at the different stages of life. The second part deals with the epistemological basis of this work. It explores the research in clinical psychology and the way it is linked to practice and teaching. It looks into the theoretical and practical filiation that operates in our approach to the mental life and to the care relationship. The third part is dedicated to the methodology of the clinical research – mainly the specificities of psychoanalytical observation. It presents E. Bick’s method of infant observation and how this method was adapted for the clinical work with anorexic patients. The fourth part ponders on what is psychologically at stake in the ‘feeding relation’. This relation is seen as commensal (W.R. Bion, 1962), the paradigm of the intersubjective encounter, of the psychological development of the baby’s subjectivity and the emerging parentality, but also the most favored ground of the unconscious psychological transmission. The fifth part creates a dialogue between the baby-in-its-family and the babies aspects of the adult self in the contexts of anorexia. Anorexias are considered in terms of food refusal-impossibility demonstrating an attempt to organise psychological survival. Two common anxieties are at the center of the psychological dynamic : a catastrophic anxiety and a persecution anxiety. The catastrophic anxiety appears as an « endless and disintegrative fall », a sign of a primary depression, which is overloaded with a vampirising and devouring anxiety. The persecution anxiety is also thought as a form of objectalisation of the anxiety of « endless and disintegrative fall ». The adhesive defenses, which are organised to fight against the « endless and disintegrative fall », are accompagnied by an inhibition of the instinctual drives —especially of secondary greed— as well as their splitting, denial and projection. The contact with the baby’s emotional life reactivates the primary depression against which the babies aspects of the parental self are organised in an economy of psychological survival. The food refusal-impossibility of the baby appears as a mirror to the refusal-impossibility to receive and give meaning to the emotionnal life that characterises anorexic parentality. The link between the baby and its parentality, but also between the babies aspects of the self and the internal parentality, presents an opaque and persecutory reflexivity that is interiorized by the baby and the babies aspects of the self, with the form of a pernicious and involutive circle. Finally, this dissertation concludes with the proposition of a range of the feeding relation that is structured by two poles. The first pole is the more integrated one : that of the feeding relation. It is marked by a relation of intimacy between the baby and its parentality and by the introjection of a commensal reflexivity which is both mutual and asymmetrical, and which sustains the integration of instinctual drives, the development of the bases of a protective superego and of the mental growth. The second pole, the un-linked one, is that of the relation of un-feeding characterized by a relation of ex-timity between the baby and its parentality, and the interiorization of a persecutory and opaque reflexivity in the form of a involutive circle leading to the desintegration of the drive and the development of an ego destructive superego (W.R.Bion, 1959).

Dans un problème de prédiction à multiples pas discrets, la prédiction à chaque instant peut dépendre de l’entrée à n’importe quel moment dans un passé lointain. Modéliser une telle dépendance à long terme est un des problèmes fondamentaux en apprentissage automatique. En théorie, les Réseaux de Neurones Récurrents (RNN) peuvent modéliser toute dépendance à long terme. En pratique, puisque la magnitude des gradients peut croître ou décroître exponentiellement avec la durée de la séquence, les RNNs ne peuvent modéliser que les dépendances à court terme. Cette thèse explore ce problème dans les réseaux de neurones récurrents et propose de nouvelles solutions pour celui-ci. Le chapitre 3 explore l’idée d’utiliser une mémoire externe pour stocker les états cachés d’un réseau à Mémoire Long et Court Terme (LSTM). En rendant l’opération d’écriture et de lecture de la mémoire externe discrète, l’architecture proposée réduit le taux de décroissance des gradients dans un LSTM. Ces opérations discrètes permettent également au réseau de créer des connexions dynamiques sur de longs intervalles de temps. Le chapitre 4 tente de caractériser cette décroissance des gradients dans un réseau de neurones récurrent et propose une nouvelle architecture récurrente qui, grâce à sa conception, réduit ce problème. L’Unité Récurrente Non-saturante (NRUs) proposée n’a pas de fonction d’activation saturante et utilise la mise à jour additive de cellules au lieu de la mise à jour multiplicative. Le chapitre 5 discute des défis de l’utilisation de réseaux de neurones récurrents dans un contexte d’apprentissage continuel, où de nouvelles tâches apparaissent au fur et à mesure. Les dépendances dans l’apprentissage continuel ne sont pas seulement contenues dans une tâche, mais sont aussi présentes entre les tâches. Ce chapitre discute de deux problèmes fondamentaux dans l’apprentissage continuel: (i) l’oubli catastrophique d’anciennes tâches et (ii) la capacité de saturation du réseau. De plus, une solution est proposée pour régler ces deux problèmes lors de l’entraînement d’un réseau de neurones récurrent. / In a multi-step prediction problem, the prediction at each time step can depend on the input at any of the previous time steps far in the past. Modelling such long-term dependencies is one of the fundamental problems in machine learning. In theory, Recurrent Neural Networks (RNNs) can model any long-term dependency. In practice, they can only model short-term dependencies due to the problem of vanishing and exploding gradients. This thesis explores the problem of vanishing gradient in recurrent neural networks and proposes novel solutions for the same. Chapter 3 explores the idea of using external memory to store the hidden states of a Long Short Term Memory (LSTM) network. By making the read and write operations of the external memory discrete, the proposed architecture reduces the rate of gradients vanishing in an LSTM. These discrete operations also enable the network to create dynamic skip connections across time. Chapter 4 attempts to characterize all the sources of vanishing gradients in a recurrent neural network and proposes a new recurrent architecture which has significantly better gradient flow than state-of-the-art recurrent architectures. The proposed Non-saturating Recurrent Units (NRUs) have no saturating activation functions and use additive cell updates instead of multiplicative cell updates. Chapter 5 discusses the challenges of using recurrent neural networks in the context of lifelong learning. In the lifelong learning setting, the network is expected to learn a series of tasks over its lifetime. The dependencies in lifelong learning are not just within a task, but also across the tasks. This chapter discusses the two fundamental problems in lifelong learning: (i) catastrophic forgetting of old tasks, and (ii) network capacity saturation. Further, it proposes a solution to solve both these problems while training a recurrent neural network.

Health economists, national decision-makers and global health specialists have been interested in calculating the cost of a disease for many years. Only more recently they started to generate more comprehensive frameworks and tools to estimate the full range of healthcare related costs of illness in a user’s perspective in resource-poor settings. There is now an ongoing trend to guide health policy, and identify the most effective ways to achieve universal health coverage. The user fee exemptions health financing schemes, which grounded the tuberculosis control strategy, have been designed to improve access to essential care for ill individuals with a low capacity to pay. After decades of functioning and substantial progress in tuberculosis detection rate and treatment success, this thesis analyses the extent of the coverage (financial and social protection) of two disease control programs in West Africa. Learning from the concept of the medical poverty trap (Whitehead, Dahlgren, et Evans 2001) and available framework related to the economic consequences of illness (McIntyre et al. 2006), a conceptual framework and a data collection tool have been developed to incorporate the direct, indirect and intangible costs and consequences of illness incurred by chronic patients. In several ways, we have sought to provide baseline for comprehensive analysis and standardized methodology to allow comparison across settings, and to contribute to the development of evidence-based knowledge.<p><p>To begin, filling a knowledge gap (Russell 2004), we have performed microeconomic research on the households’ costs-and-consequences-of-tuberculosis in Burkina Faso and Benin. The two case studies have been conducted both in rural and urban resource-poor settings between 2007 and 2009. This thesis provides new empirical findings on the remaining financial, social and ‘healthcare delivery related organizational’ barriers to access diagnosis and treatment services that are delivered free-of-charge to the population. The direct costs associated with illness incurred by the tuberculosis pulmonary smear-positive patients have constituted a severe economic burden for these households living in permanent budget constraints. Most of these people have spent catastrophic health expenditure to cure tuberculosis and, at the same time, have faced income loss caused by the care-seeking. To cope with the substantial direct and indirect costs of tuberculosis, the patients have shipped their families in impoverishing strategies to mobilize funds for health such as depleting savings, being indebted and even selling livestock and property. Damaging asset portfolios of the disease-affected households on the long run, the coping strategies result in a public health threat. In resource-poor settings, the lack of financial protection for health may impose inability to meet basic needs such as the rights to education, housing, food, social capital and access to primary healthcare. Special feature of our work lies in the breakdown of the information gathered. We have been able to demonstrate significant differences in the volume and nature of the amounts spent across the successive stages of the care-seeking pathway. Notably, pre-diagnosis spending has been proved critical both in the rural and urban contexts. Moreover, disaggregated cost data across income quintiles have highlighted inequities in relation to the direct costs and to the risk of incurring catastrophic health expenditure because of tuberculosis. As part of the case studies, the tuberculosis control strategies have failed to protect the most vulnerable care users from delayed diagnosis and treatment, from important spending even during treatment – including significant medical costs, and from hidden costs that might have been exacerbated by poor health systems. To such devastating situations, the tuberculosis patients have had to endure other difficulties; we mean intangible costs such as pain and suffering including stigmatization and social exclusion as a result of being ill or attending tuberculosis care facilities. The analysis of all the social and economic consequences for tuberculosis-affected households over the entire care-seeking pathway has been identified as an essential element of future cost-of-illness evaluations, as well as the need to conduct benefit incidence assessment to measure equity.<p><p>This work has allowed identifying a series of policy weaknesses related to the three dimensions of the universal health coverage for tuberculosis (healthcare services, population and financial protection coverage). The findings have highlighted a gap between the standard costs foreseen by the national programs and the costs in real life. This has suggested that the current strategies lack of patient-centered care, context-oriented approaches and systemic vision resulting in a quality issue in healthcare delivery system (e.g. hidden healthcare related costs). Besides, various adverse effects on households have been raised as potential consequences of illness; such as illness poverty trap, social stigma, possible exclusion from services and participation, and overburdened individuals. These effects have disclosed the lack of social protection at the country level and call for the inclusion of tuberculosis patients in national social schemes. A last policy gap refers to the lack of financial protection and remaining inequities with regards to catastrophic health expenditure still occurring under use fee exemptions strategies. Thereby, one year before 2015 – the deadline set for the Millennium Development Goals – it is a matter of priority for Benin and Burkina Faso and many other countries to tackle adverse effects of the remaining social, economic and health policy and system related barriers to tuberculosis control. These factors have led us to emphasize the need for countries to develop sustainable knowledge. <p><p>National decision-makers urgently need to document the failures and bottlenecks. Drawing on the findings, we have considered different ways to strengthen local capacity and generate bottom-up decision-making. To get there, we have shaped a decision framework intended to produce local evidence on the root causes of the lack of policy responsiveness, synthesize available evidence, develop data-driven policies, and translate them into actions.<p><p>Beyond this, we have demonstrated that controlling tuberculosis was much more complex than providing free services. The socio-economic context in which people affected by this disease live cannot be dissociated from health policy. The implications of microeconomic research on the households’ costs and responses to tuberculosis may have a larger scope than informing implementation and adaptation of national disease-specific strategies. They can be of great interest to support the definition of guiding principles for further research on social protection schemes, and to produce evidence-based targets and indicators for the reduction and the monitoring of economic burden of illness. In this thesis, we have build on prevailing debates in the field and formulated different assumptions and proposals to inform the WHO Global Strategy and Targets for Tuberculosis Prevention, Care and Control After 2015. For us, to reflect poor populations’ needs and experiences, global stakeholders should endorse bottom-up and systemic policy-making approaches towards sustainable people-centered health systems.<p><p>The findings of the thesis and the various global and national challenges that have emerged from case studies are crucial as the problems we have seen for tuberculosis in West Africa are not limited to this illness, and far outweigh the geographical context of developing countries.<p><p><p>Keywords: Catastrophic health expenditure, Coping strategies, Cost-of-illness studies, Direct, indirect and intangible costs, Evidence-based Public health, Financial and Social protection for health, Health Economics, Health Policy and Systems, Informed Decision-making, Knowledge translation, People-centered policy-making, Systemic approach, Universal Health Coverage<p> / Doctorat en Sciences de la santé publique / info:eu-repo/semantics/nonPublished