Search

Recent work has established clear links between the generalization performance of trained neural networks and the geometry of their loss landscape near the local minima to which they converge. This suggests that qualitative and quantitative examination of the loss landscape geometry could yield insights about neural network generalization performance during training. To this end, researchers have proposed visualizing the loss landscape through the use of simple dimensionality reduction techniques. However, such visualization methods have been limited by their linear nature and only capture features in one or two dimensions, thus restricting sampling of the loss landscape to lines or planes. Here, we expand and improve upon these in three ways. First, we present a novel "jump and retrain" procedure for sampling relevant portions of the loss landscape. We show that the resulting sampled data holds more meaningful information about the network's ability to generalize. Next, we show that non-linear dimensionality reduction of the jump and retrain trajectories via PHATE, a trajectory and manifold-preserving method, allows us to visualize differences between networks that are generalizing well vs poorly. Finally, we combine PHATE trajectories with a computational homology characterization to quantify trajectory differences., Accepted at the 20th Symposium on Intelligent Data Analysis (IDA) 2022

L'objectif de cet article est de présenter une démarche de construction d'un exercice de prospective territoriale basé sur un outil de modélisation spatiale. Cette démarche est mise en oeuvre dans le cadre du projet ANR DESCARTES dont l'objectif est de construire un outil de simulation cartographique pour analyser différents scénarios d'affectation de l'usage des sols à l'Ile de La Réunion, en termes de services environnementaux. La plateforme de simulation cartographique est composée de deux applications complémentaires (i) le langage de modélisation Ocelet et son environnement de simulation de paysages dynamiques, et (ii) la plateforme Margouill@. Le développement puis la présentation, en atelier, d'un premier démonstrateur sur la consommation des terres agricoles par l'urbanisation a permis de tester l'outil cartographique comme support d'un processus d'innovation collective entre les parties prenantes, et d'ouvrir de nouveaux champs de recherche sur l'analyse de la démarche par les apprentissages, la spatialisation et la simulation prospective des services écosystémiques, et la prise en compte du changement d'échelle.

After having set out the challenges connected with land ownership and real estate in the insular, micro context of Reunion Island, we plan to create a robot land-use model based on satellite images integrated with a cellular robot. Once the general framework of the modelling has been established, we will make the case for calibrating the cellular robot, and will discuss four scenarios in an approach to long-term territorial planning. La problématique de l'usage du sol à La Réunion.

After having set out the challenges connected with land ownership and real estate in the insular, micro context of Reunion Island, we plan to create a robot land-use model based on satellite images integrated with a cellular robot. Once the general framework of the modelling has been established, we will make the case for calibrating the cellular robot, and will discuss four scenarios in an approach to long-term territorial planning. La problématique de l’usage du sol à La Réunion.

The role of humoral rejection in acute and chronic rejection of human renal allografts other than in hyperacute rejection has not been well established, and its importance may be underestimated. Recently, a specific histological pattern of antibody-mediated rejection of renal allografts has been recognized. The antigens targeted by this mode of rejection are not well defined but are likely located on the endothelium of small vessels (arterioles and glomerular and peritubular capillaries). In both cellular and humoral rejection, the microvasculature of transplanted organs appears to be a main target of injury. This study describes the ultrastructural changes of peritubular capillaries, over a period of up to 8 months, in 14 biopsy specimens obtained from 5 renal allograft recipients diagnosed with "pure" antibody-mediated rejection. In peritubular capillaries, there is progression of injury from necrosis of endothelial cells with lifting and denudation of basement membrane to complete disappearance of capillaries. Acutely, acute tubular necrosis is a constant finding. At 2 to 3 months post-transplantation, the remaining capillaries are dilated, misshapen, and distorted, and are surrounded by a reduplicated and thickened basement membrane. These changes are associated with increased interstitial fibrosis and tubular atrophy, comparable to a sort of renal "asphyxial" death. The author concludes that in "pure" antibody-mediated rejection, the endothelium of peritubular capillaries is a main target of injury. The potential role of antibody-mediated rejection in acute and chronic rejection of renal allografts needs to be explored further.

32nd European Photovoltaic Solar Energy Conference and Exhibition; 2318-2321, Although the high amount of solar irradiance in the tropics is an advantage for a profitable PV production, the local meteorological conditions induce a very high variability which is problematic for a safe and gainful injection into the power grid. This issue is even more critical in non-interconnected territories where network stability is an absolute necessity and the injection of PV power has to be limited. The basis for precise cloud evolution and subsequent irradiance forecasts are high quality atmospheric analyses for NWP. Geostationary meteorological satellites provide valuable observations of cloud properties with high spatio-temporal resolutions and allow a pertinent data assimilation. The shortcoming is that optical and thermal channels of satellite sensors do not provide cloud properties from inside clouds. Different existing data assimilation approaches aim at deriving atmospheric analyses with most realistic cloud features, utilising geostationary satellite observations. The potential of assimilating satellite-derived cloud information in regional NWP with focus on irradiance forecasts in tropical regions has not been evaluated so far. Hence, the present work aims at evaluating the potential of geostationary satellite data assimilation in limited-area models applied to the French tropical oversea territories Reunion Island and French Guiana.

International audience; L'épidémie de chikungunya ayant sévi dans l'Océan Indien en 2005-2007 a eu un impact majeur sur les populations. Au-delà de la densité vectorielle, les facteurs environnementaux et sociaux jouent un rôle important dans la propagation et l'ampleur d'une épidémie d'arbovirose. Ils conditionnent le risque d'exposition d'une population à un agent pathogène à transmission vectorielle L'objectif de cette étude est la caractérisation socio-spatiale des premiers foyers de chikungunya à La Réunion pour poser des hypothèses permettant d'expliquer le démarrage de l'épidémie à partir de ces quartiers. L'étude vise à identifier les facteurs descriptifs significatifs et à définir la carte des quartiers à risque de La Réunion concernant ces facteurs. La période d'étude est celle du premier pic épidémique de février à juillet 2005 où les enquêtes des signalements ont permis le géoréférencement des 2997 premiers cas de chikungunya. La zone d'étude recouvre les zones urbaines situées à moins de 250m d'altitude pour réduire l'impact du facteur altitudinal sur la transmission vectorielle. Les données (DRASS, INSEE, IGN, Cadastre) ont permis de définir 40 variables décrivant le milieu urbain (habitat et densité) et les populations (structure et activités) résidant dans des unités géographiques de référence. Pour valider ces facteurs, des enquêtes de terrain ont été menées dans les quartiers urbains des premiers foyers et des témoins (mêmes facteurs descriptifs significatifs mais sans cas de chikungunya sur la période) pour affiner l'étude environnementale et comportementale des populations y résidant.