Les hydrogels de polysaccharide sont des biomatériaux utilisés comme matrices à libération contrôlée de médicaments et comme structures modèles pour l’étude de nombreux systèmes biologiques dont les biofilms bactériens et les mucus. Dans tous les cas, le transport de médicaments ou de nutriments à l’intérieur d’une matrice d’hydrogel joue un rôle de premier plan. Ainsi, l’étude des propriétés de transport dans les hydrogels s’avère un enjeu très important au niveau de plusieurs applications. Dans cet ouvrage, le curdlan, un polysaccharide neutre d’origine bactérienne et formé d’unités répétitives β-D-(1→3) glucose, est utilisé comme hydrogel modèle. Le curdlan a la propriété de former des thermogels de différentes conformations selon la température à laquelle une suspension aqueuse est incubée. La caractérisation in situ de la formation des hydrogels de curdlan thermoréversibles et thermo-irréversibles a tout d’abord été réalisée par spectroscopie infrarouge à transformée de Fourier (FT-IR) en mode réflexion totale atténuée à température variable. Les résultats ont permis d’optimiser les conditions de gélation, menant ainsi à la formation reproductible des hydrogels. Les caractérisations structurales des hydrogels hydratés, réalisées par imagerie FT-IR, par microscopie électronique à balayage en mode environnemental (eSEM) et par microscopie à force atomique (AFM), ont permis de visualiser les différentes morphologies susceptibles d’influencer la diffusion d’analytes dans les gels. Nos résultats montrent que les deux types d’hydrogels de curdlan ont des architectures distinctes à l’échelle microscopique. La combinaison de la spectroscopie de résonance magnétique nucléaire (RMN) à gradients pulsés et de l’imagerie RMN a permis d’étudier l’autodiffusion et la diffusion mutuelle sur un même système dans des conditions expérimentales similaires. Nous avons observé que la diffusion des molécules dans les gels est ralentie par rapport à celle mesurée en solution aqueuse. Les mesures d’autodiffusion, effectuées sur une série d’analytes de diverses tailles dans les deux types d’hydrogels de curdlan, montrent que le coefficient d’autodiffusion relatif décroit en fonction de la taille de l’analyte. De plus, nos résultats suggèrent que l’équivalence entre les coefficients d’autodiffusion et de diffusion mutuelle dans les hydrogels de curdlan thermo-irréversibles est principalement due au fait que l’environnement sondé par les analytes durant une expérience d’autodiffusion est représentatif de celui exploré durant une expérience de diffusion mutuelle. Dans de telles conditions, nos résultats montrent que la RMN à gradients pulsés peut s’avérer une approche très avantageuse afin de caractériser des systèmes à libération contrôlée de médicaments. D’autres expériences de diffusion mutuelle, menées sur une macromolécule de dextran, montrent un coefficient de diffusion mutuelle inférieur au coefficient d’autodiffusion sur un même gel de curdlan. L’écart mesuré entre les deux modes de transport est attribué au volume différent de l’environnement sondé durant les deux mesures. Les coefficients d’autodiffusion et de diffusion mutuelle similaires, mesurés dans les deux types de gels de curdlan pour les différents analytes étudiés, suggèrent une influence limitée de l’architecture microscopique de ces gels sur leurs propriétés de transport. Il est conclu que les interactions affectant la diffusion des analytes étudiés dans les hydrogels de curdlan se situent à l’échelle moléculaire. / Polysaccharide hydrogels are biomaterials used as controlled drug delivery matrices and serve as model scaffolds for the study of many biological systems like bacterial biofilms and mucus. In every case, the transport of drugs or nutriments across a hydrogel matrix is of prime importance. Therefore, the study of transport properties in hydrogels is an important issue for many fields of application. In this work, curdlan, a neutral bacterial polysaccharide made of β-D-(1→3) glucose repeating units, is used as a model hydrogel. Aqueous suspensions of curdlan can form thermogels of different conformations depending on the incubation temperature. In situ characterization of the preparation of thermo-reversible (low-set) and thermo-irreversible (high-set) curdlan hydrogels was first carried out using variable temperature attenuated total reflection (ATR) Fourier transform infrared spectroscopy (FT-IR). The results allowed optimization of the gelling conditions leading to reproducible gel samples. Structural characterization of fully hydrated hydrogels, carried out by FT-IR imaging, environmental scanning electron microscopy (eSEM) and atomic force microscopy (AFM), allowed visualization of the different gel morphologies susceptible of influencing the diffusion of analytes in hydrogels. Our results show that both types of curdlan hydrogels have distinct microscopic architectures. The combination of pulsed field gradient (PFG) nuclear magnetic resonance (NMR) spectroscopy and NMR profiling allowed the study of self-diffusion and mutual diffusion on the same hydrogel system in similar experimental conditions. We showed that the diffusion of analytes in the gels is slower than in the aqueous solution. The diffusion experiments, carried out on a series of analytes of various sizes in both types of curdlan gels, show a decrease of the relative self-diffusion coefficient as a function of the analyte size. In addition, our results suggest that the equivalence between the self-diffusion and mutual-diffusion coefficients measured in the high-set curdlan gels is mainly due to the fact that the environment probed by the analytes during a self-diffusion experiment is representative of the one probed during a mutual-diffusion experiment. In such conditions, our results show that PFG NMR may present a valuable approach for the characterization of controlled drug release systems. Additional experiments show that the mutual-diffusion coefficient of dextran macromolecules is smaller than its self-diffusion coefficient in the same curdlan hydrogel. The difference between both transport rates is attributed to the different environment volumes probed by the analytes during the measurements. The similarities observed between the self-diffusion and mutual-diffusion coefficients, measured in both types of curdlan gels for all investigated analytes, suggest a limited influence of the microscopic gel architecture on its transport properties. It is therefore concluded that the interactions affecting the diffusion of the investigated analytes in the curdlan hydrogels lie at the molecular scale.

There is concern that climate change may cause mismatches between timing of flowering and activity of pollinators (phenology). However, concluding that mismatches will occur, and have serious consequences for pollination services, requires assumptions that have not yet been tested. I begin by discussing a set of these assumptions, bringing past research into the context of mismatch. Briefly, the assumptions are that 1) dates of first-flowering or emergence (DFFE) correctly describe phenology (and therefore mismatch); 2) differences in DFFE represent the magnitude of mismatch; 3) advancement of DFFE will be the primary phenological change; 4) shifts will be random and independent for each species; 5) populations of plants and pollinators are “bottom-up” regulated by their mutualistic interactions; 6) all interactions are of similar strength and importance; 7) dispersal, and the spatial context of phenological mismatches can be ignored; and ecological processes including 8) phenotypic plasticity and adaptive evolution of phenology, 9) competition and facilitation, and 10) emergence of novel interactions, will not affect mismatches. I then describe novel experiments, which could help to account for some of these assumptions, clarifying the existence and impacts of mismatches. Next, I present an original field experiment on factors affecting seed set in an alpine meadow in the Coast Range of British Columbia, Canada. I found evidence contradicting the assumption that seed set is primarily limited by pollination. My data highlight the roles of phenology, temperature (degree-days above 15°C, and frost hours), and interactions with pollinators (mutualists) and seed-predators (floral antagonists) in driving patterns of seed set. Seed set of early and late-flowering species responded differently to a 400m elevation gradient, which might be explained by phenology of bumble bees. My data suggest that the consequences of mismatch may be smallest for plants that are fly-pollinated and self-fertile. Non-selfing, bee-pollinated species might be more prone to reproductive limitation through mismatch (affected by snowmelt and cumulative degree-days). Plants that are limited by seed-predators might be negatively affected by warming temperatures with fewer frost hours, and extreme events such as late-season frosts and hail storms can prevent plants from setting seed entirely. Overall, my work emphasizes the importance of complementing theory, data-driven simulations, and meta-analyses with experiments carried out in the field. / Graduate

The structure and composition of mixed-conifer forest (MCF) in central Oregon has been altered by fire exclusion and logging. The resulting increased density, spatial contagion, and loss of fire resistant trees decrease the resiliency of this ecosystem to fire, drought, and insects. The historical and current composition and structure of MCF are characterized by steep environmental gradients and a complex mixed-severity fire regime. This inherent variation makes it difficult to determine the magnitude of anthropogenic effects and set objectives for restoration and management. As a result, there is a lack of consensus regarding how MCF should be managed and restored across the landscape. My primary research objectives were to: (1) Characterize the current structure and composition of MCF and how these vary with environmental setting; and (2) Characterize establishment and tree growth patterns in MCF in different environmental settings. To address these objectives, I collected field data on structure and composition and increment cores across a range of environmental conditions in MCF of the eastern Cascades and Ochoco Mountains. I used cluster analysis to identify four stand types based on structure and composition in the eastern Cascades study area and four analogous types in the Ochoco Mountains study area. Variation in understory composition and the presence of large diameter shade tolerant species distinguish each type. Stand types occupied distinct environmental settings along a climatic gradient of increasing precipitation and elevation. At relatively dry PIPO sites understories were dominated by ponderosa pine. At wetter PIPO/PSME and PIPO ABGC sites understories were dominated by shade tolerant species, but ponderosa pine was dominant in the overstory. At the coolest and wettest PIPO/PSME/ABGC sites understories were dominated by grand fir and shade tolerant species were common in the overstory. In the eastern Cascades current density of all live trees and snags was 432, 461, 570, 372 trees per hectare (TPH) for the four stand types identified. Stand types in the drier Ochoco Mountains were currently less dense at 279, 304, 212, and 307 TPH. Current MCF densities in both areas are 2-3 times higher than densities estimated for the late 19th and early 20th centuries from other studies in those two areas. Reconstruction of cuts in each stand type indicates that the density of large diameter ponderosa pine has been reduced by approximately 50% in all stand types in both study regions. Age histograms demonstrate that current density and composition of MCF stand types is a product of abrupt increases in tree establishment following fire exclusion in the late 19th century. The number of trees established increased after 1900 in all stand types, but the timing and composition of changes in establishment varied with climate. At dry PIPO sites increases in establishment were delayed until the 1920s and 1930s and were composed of ponderosa pine. At PIPO/PSME and PIPO/ABGC sites with intermediate precipitation, establishment was dominated by ponderosa pine prior to 1900, but after 1900 establishment was dominated by a large pulse of Douglas-fir and grand fir. At the wettest PIPO/PSME/ABGC there was less evidence of changes in structure and composition over time. My results indicate that compared to dry pine and dry-mixed conifer sites, relatively productive moist mixed-conifer sites were characterized by large changes in structure and composition. Such sites could be considered more ecologically altered by lack of fire than drier forest types that had high fire frequencies but slower rates of stand development and less plant community change. Radial growth patterns of cored ponderosa pines differed between the eastern Cascades and Ochoco Mountains. In the eastern Cascades mean growth rates and variance decreased during favorable climatic periods after 1900. This is likely related to increased competition, and provides evidence that current stand density lacks a temporal analog in the 18th and 19th centuries. Sensitivity of growth to climate and harvest suggest competition for water in the denser forest of the eastern Cascades, and indicates thinning will increase the diameter growth rate of large old pines. In the Ochoco Mountains, ponderosa pine tree growth was less responsive to climate prior to fire exclusion in the late 1800s, and growth did not respond to fire events. This suggests competition among trees was historically low in this region. After fire exclusion growth became more responsive to wet and dry climatic cycles, which may indicate that increased density and competition made trees more responsive to climate variability. Patterns of slow and fast growth appeared to differ between study regions and likely differ at the sub-regional scale. Further analysis of the relationship between growth and climate in different environmental settings is needed to distinguish where stand development has been modified by disruption of fire regimes. / Graduation date: 2013

Morphological diversity of 61 Anabaena populations of 13 morphospecies was described under the field conditions of Czech fishponds and reservoirs. Polyphasic approach was then applied in classification of 45 clonal strains isolated from those populations. Detailed morphological analyses were performed and partial 16S rRNA gene sequences were obtained for 33 of the strains, and secondary metabolite production was evaluated in 20 strains. Plasticity of morphological characteristics under varied conditions of light, temperature, nitrogen and phosphorus was studied in selected strains, as well as their temperature and light growth requirements. The results were then discussed with respect to the delimitation of single Anabaena morphospecies. A new genus Sphaerospermum was defined for the morphospecies Anabaena kisseleviana, A. reniformis and Aphanizomenon aphanizomenoides, whose phenotypic and genotypic features differed considerably from all other Anabaena morphospecies. Unique information was provided on the occurrence and distribution of A. reniformis and Aph. aphanizomenoides in the Czech Republic.

The thesis describes diversity of birds along a complete altitudinal gradient and in forest fragments in lowlands of Papua New Guinea. It focuses separately on the diversity of different feeding guilds, and discusses their links to habitat and food resources. More specifically, it focuses on forest insectivorous birds, their predation pressure on arthropods, feeding specializations and preferences, and some of the ways how insectivores search for food.

La tâche la plus délicate d'un système de reconnaissance faciale est la phase d'extraction de caractéristiques significatives et discriminantes. Dans le cadre de cette thèse, nous nous sommes focalisés sur cette tâche avec comme objectif l'élaboration d'une représentation de visage robuste aux variations majeures suivantes: variations d'éclairage, de pose, de temps, images de qualité différentes (vidéosurveillance). Par ailleurs, nous avons travaillé également dans une optique de traitement temps réel. Tout d'abord, en tenant compte des caractéristiques d'orientation des traits principaux du visages (yeux, bouche), une nouvelle variante nommée ELBP de célèbre descripteur LBP a été proposée. Elle s'appuie sur les informations de micro-texture contenues dans une ellipse horizontale. Ensuite, le descripteur EPOEM est construit afin de tenir compte des informations d'orientation des contours. Puis un descripteur nommée PLPQMC qui intégre des informations obtenues par filtrage monogénique dans le descripteur LPQ est proposé. Enfin le descripteur LPOG intégrant des informations de gradient est présenté. Chacun des descripteurs proposés est testé sur les 3 bases d'images AR, FERET et SCface. Il en résulte que les descripteurs PLPQMC et LPOG sont les plus performants et conduisent à des taux de reconnaissance comparables voire supérieur à ceux des meilleurs méthodes de l'état de l'art. / Centered around feature extraction, the core task of any Face recognition system, our objective is devising a robust facial representation against major challenges, such as variations of illumination, pose and time-lapse and low resolution probe images, to name a few. Besides, fast processing speed is another crucial criterion. Towards these ends, several methods have been proposed through out this thesis. Firstly, based on the orientation characteristics of the facial information and important features, like the eyes and mouth, a novel variant of LBP, referred as ELBP, is designed for encoding micro patterns with the usage of an horizontal ellipse sample. Secondly, ELBP is exploited to extract local features from oriented edge magnitudes images. By this, the Elliptical Patterns of Oriented Edge Magnitudes (EPOEM) description is built. Thirdly, we propose a novel feature extraction method so called Patch based Local Phase Quantization of Monogenic components (PLPQMC). Lastly, a robust facial representation namely Local Patterns of Gradients (LPOG) is developed to capture meaningful features directly from gradient images. Chiefs among these methods are PLPQMC and LPOG as they are per se illumination invariant and blur tolerant. Impressively, our methods, while offering comparable or almost higher results than that of existing systems, have low computational cost and are thus feasible to deploy in real life applications.

O desenvolvimento e aplicação de métodos analíticos na avaliação das concentrações de metais nas frações dissolvida e lábil na bacia do rio Piracicaba são de relevante importância para o estudo do comportamento químico-ambiental desses metais e para seu monitoramento, devido ao impacto das atividades antropogênicas nessa bacia. A técnica de difusão em filmes finos por gradiente de concentração (DGT, do inglês diffusive gradients in thin-films), é uma promissora ferramenta analítica de amostragem, pois permite pré-concentração, análise temporal, remoção de interferentes, avaliação lábil e amostragem in situ. A quantificação dos metais pelas técnicas de fluorescência de raios X dispersiva em energia por reflexão total (TXRF, total reflection X-ray fluorescence analysis) para amostras líquidas e da fluorescência de raios X dispersiva em energia (EDXRF, energy dispersive X-ray fluorescence) para amostras sólidas, mostram-se interessantes devido ao caráter multielementar, simultâneo e não destrutivo da análise. Neste trabalho avaliou-se a DGT para a determinação de Cr, Mn, Co, Ni, Cu, Zn e Pb em águas naturais, utilizando-se como agente ligante a membrana de celulose P 81 e como meio difusivo, o papel cromatográfico 3 MM. Calcularam-se os fatores de eluição e coeficientes de difusão para estes metais. Estudou-se o efeito da substância húmica, em diferentes concentrações, no desempenho da DGT para esses metais, e os resultados foram comparados e concordantes com as frações livre e a ligada à matéria orgânica, calculadas pelo programa Visual MINTEQ. Compararam-se o agente ligante e o meio difusivo à base de celulose, utilizados neste trabalho, com os à base de gel de poliacrilamida, empregando-se o complexante Chelex-100. Avaliou-se a homogeneidade da retenção dos analitos e do grupo ligante fosfato sobre a superfície do agente ligante. Nesses ensaios utilizaram-se soluções sintéticas contendo os analitos em pH 5,5 e força iônica 0,05 mol L-1 (ajustados com NaNO3). As determinações foram realizadas por TXRF e EDXRF. Avaliaram-se os acoplamentos DGT-SRTXRF e da DGT-EDXRF para a determinação da fração lábil de Cr, Mn, Co, Ni, Cu, Zn e Pb em amostras de água da bacia do rio Piracicaba, através de imersão em laboratório e in situ. Esses resultados foram comparados com as concentrações dissolvidas desses metais determinados pela técnica de fluorescência de raios X por reflexão total com excitação por luz síncrotron (SRTXRF, synchrotron radiation total reflection X-ray analysis). Os valores de concentrações de Mn, Cu e Zn nas frações dissolvida e lábil foram comparados com os valores máximos estabelecidos pela legislação brasileira Nas amostragens in situ, nos cinco pontos de coleta, as concentrações de Mn dissolvido e lábil foram maiores que a concentração total máxima permitida para esse elemento. Os elmentos Cr, Co, Ni e Pb, nas amostras coletadas, apresentaram-se abaixo do limite de detecção das técnicas SRTXRF (dissolvido), DGT-SRTXRF (lábil) e DGT-EDXRF (lábil). Os limites de detecção para o Cr, Mn, Co, Ni, Cu, Zn e Pb foram de 2,2 a 10 \'mü\'g L-1 na SRTXRF, de 0,21 a 3,6 \'mü\'g L-1 e de 7,2 a 24 \'mü\'g L-1 na DGT-SRTXRF e DGT-EDXRF, respectivamente / Analytical methods evaluation for metal dissolved and labile fractions determination in Piracicaba river basin is notable for metal-behavior studies in this environment as well as their monitoring due to anthropogenic activities. The diffusive gradients in thin films (DGT) is a promising sampling technique because of its pre-concentration, time-integrated, matrix interference removal, labile evaluation and in situ analytical features. The analytical techniques total reflection X-ray fluorescence (TXRF) for liquid samples and energy dispersive X-ray fluorescence (EDXRF) for solid samples analysis are advantageous due to simultaneous, multielemental and non-destructive characteristics. In this work, it was evaluated the DGT technique for Cr, Mn, Co, Ni, Cu, Zn and Pb determination in freshwater, using P 81 cellulose membrane as bind agent and 3 MM chromatographic paper as diffusive layer. Elution factors and diffusion coefficients were calculated for these metals. The humic acids effect on DGT performance was assessed and compared to free and humic acid bound metal fractions estimated by Visual MINTEQ software, in which agreement results were found. The paper-based DGT and gel-based one with Chelex-100 as bind agent were compared. The metal retention and phosphate group homogenization on P 81 membrane bind agent surface was studied. For these tests it was utilized a synthetic solution these metals containing with pH and ionic strength adjust to 5.5 and 0.05 mol L-1 (NaNO3), respectively. It was used the TXRF and EDXRF techniques for these metal quantifications. The DGT-SRTXRF and DGT-EDXRF were assessed in order to determine Cr, Mn, Co, Ni, Cu, Zn and Pb labile fractions in batch and in situ in Piracicaba river basin freshwater samples. These results were compared to their dissolved fractions by synchrotron radiation total reflection X-ray analysis, SRTXRF. The dissolved and labile Mn, Cu and Zn concentrations levels were compared to maximum allowed set by Brazilian legislation. The dissolved and labile Mn concentrations were higher than the maximum allowed in all five samples in situ sampling. Cr, Co, Ni, and Pb concentrations were lower than their limits of detection for dissolved fractions (SRTXRF) and labile ones (DGT-SRTXRF and DGT-EDXRF). The detection limits for Cr, Mn, Co, Ni, Cu, Zn and Pb were from 2.2 to 10 \'mü\'g L-1 for SRTXRF, from 0.21 to 3.6 \'mü\'g L-1 and from 7.2 to 24 \'mü\'g L-1 for DGT-SRTXRF and DGT-EDXRF, respectively

Lie-detection has been an evergreen and evolving subject. Polygraph techniques have been the most popular and successful technique till date. The main drawback of the polygraph is that good results cannot be attained without maintaining a physical contact, of the subject under test. In general, this physical contact would induce extra consciousness in the subject. Also, any sort of arousal in the subject triggers false positives while performing the traditional polygraph based tests. With all these drawbacks in the polygraph, also, due to rapid developments in the fields of computer vision and artificial intelligence, with newer and faster algorithms, have compelled mankind to search and adapt to contemporary methods in lie-detection. Observing the facial expressions of emotions in a person without any physical contact and implementing these techniques using artificial intelligence is one such method. The concept of magnifying a micro expression and trying to decipher them is rather premature at this stage but would evolve in future. Magnification using EVM technique has been proposed recently and it is rather new to extract these micro expressions from magnified EVM based on HOG features. Till date, HOG features have been used in conjunction with SVM, and generally for person/pedestrian detection. A newer, simpler and contemporary method of applying EVM with HOG features and Back-propagation Neural Network jointly has been introduced and proposed to extract and decipher the micro-expressions on the face. Micro-expressions go unnoticed due to its involuntary nature, but EVM is used to magnify them and makes them noticeable. Emotions behind the micro-expressions are extracted and recognized using the HOG features \& Back-Propagation Neural Network. One of the important aspects that has to be dealt with human beings is a biased mind. Since, an investigator is also a human and, he too, has to deal with his own assumptions and emotions, a Neural Network is used to give the investigator an unbiased start in identifying the true emotions behind every micro-expression. On the whole, this proposed system is not a lie-detector, but helps in detecting the emotions of the subject under test. By further investigation, a lie can be detected. / This thesis uses a magnification technique to magnify the subtle, faint and spontaneous facial muscle movements or more precisely, micro-expressions. This magnification would help a system in classifying them and estimating the emotion behind them. This technique additionally magnifies the color changes, which could be used to extract the pulse without a physical contact with the subject. The results are presented in a GUI. / Gautam: +46(0)739528573, +91-9701534064 Tushal: +46(0)723219833, +91-9000242241 Venu: +46(0)734780266, +91-9298653191 Sai: +91-9989410111

L'objectif de cette thèse est de fournir des modèles et des outils de simulation pour décrire les échanges de masse entre les circuits de ventilation (galeries) et les milieux poreux des ouvrages souterrains d'enfouissement des déchets nucléaires. La modélisation prend en compte le couplage à l'interface poreux-galerie entre les écoulements liquide gaz compositionnels dans le milieu poreux constituant le stockage et les écoulements gazeux compositionnels dans le milieu galerie libre. / The objective of this thesis is to develop models and algorithms to simulate efficiently the mass exchanges occurring at the interface between the nuclear waste deep geological repositories and the ventilation excavated galleries. To model such physical processes, one needs to account in the porous medium for the flow of the liquid and gas phases including the vaporization of the water component in the gas phase and the dissolution of the gaseous components in the liquid phase. In the free flow region, a single phase gas free flow is considered assuming that the liquid phase is instantaneously vaporized at the interface. This gas free flow has to be compositional to account for the change of the relative humidity in the free flow region which has a strong feedback on the liquid flow rate at the interface.

Recurrent neural networks (RNN) are known for their notorious exploding and vanishing gradient problem (EVGP). This problem becomes more evident in tasks where the information needed to correctly solve them exist over long time scales, because it prevents important gradient components from being back-propagated adequately over a large number of steps. The papers written in this work formalizes gradient propagation in parametric and semi-parametric RNNs to gain a better understanding towards the source of this problem. The first paper introduces a simple stochastic algorithm (h-detach) that is specific to LSTM optimization and targeted towards addressing the EVGP problem. Using this we show significant improvements over vanilla LSTM in terms of convergence speed, robustness to seed and learning rate, and generalization on various benchmark datasets. The next paper focuses on semi-parametric RNNs and self-attentive networks. Self-attention provides a way by which a system can dynamically access past states (stored in memory) which helps in mitigating vanishing of gradients. Although useful, it is difficult to scale as the size of the computational graph grows quadratically with the number of time steps involved. In the paper we describe a relevancy screening mechanism, inspired by the cognitive process of memory consolidation, that allows for a scalable use of sparse self-attention with recurrence while ensuring good gradient propagation. / Les réseaux de neurones récurrents (RNN) sont connus pour leur problème de gradient d'explosion et de disparition notoire (EVGP). Ce problème devient plus évident dans les tâches où les informations nécessaires pour les résoudre correctement existent sur de longues échelles de temps, car il empêche les composants de gradient importants de se propager correctement sur un grand nombre d'étapes. Les articles écrits dans ce travail formalise la propagation du gradient dans les RNN paramétriques et semi-paramétriques pour mieux comprendre la source de ce problème. Le premier article présente un algorithme stochastique simple (h-detach) spécifique à l'optimisation LSTM et visant à résoudre le problème EVGP. En utilisant cela, nous montrons des améliorations significatives par rapport au LSTM vanille en termes de vitesse de convergence, de robustesse au taux d'amorçage et d'apprentissage, et de généralisation sur divers ensembles de données de référence. Le prochain article se concentre sur les RNN semi-paramétriques et les réseaux auto-attentifs. L'auto-attention fournit un moyen par lequel un système peut accéder dynamiquement aux états passés (stockés en mémoire), ce qui aide à atténuer la disparition des gradients. Bien qu'utile, il est difficile à mettre à l'échelle car la taille du graphe de calcul augmente de manière quadratique avec le nombre de pas de temps impliqués. Dans l'article, nous décrivons un mécanisme de criblage de pertinence, inspiré par le processus cognitif de consolidation de la mémoire, qui permet une utilisation évolutive de l'auto-attention clairsemée avec récurrence tout en assurant une bonne propagation du gradient.