Search

Article and Supplementary Material; International audience; The advancement and deployment of artificial intelligent agents brought numerous benefits in knowledge and data gathering and processing. However, one of the key challenges in deploying such agents in an open environment like the Web is their interoperability as they currently mostly run in silos. In this paper we report on a simulation and evaluation based on evolutionary agent-based modelling to empirically test how sustainable different strategies are for knowledge sharing in open multi-agent systems (MAS). Our results show the importance of translation-based approaches and the need for incentives to support these.

The simultaneous clustering of observations and features of data sets (known as co-clustering) has recently emerged as a central machine learning application to summarize massive data sets. However, most existing models focus on continuous and dense data in stationary scenarios, where cluster assignments do not evolve over time. This work introduces a novel latent block model for the dynamic co-clustering of data matrices with high sparsity. To properly model this type of data, we assume that the observations follow a time and block dependent mixture of zero-inflated distributions, thus combining stochastic processes with the time-varying sparsity modeling. To detect abrupt changes in the dynamics of both cluster memberships and data sparsity, the mixing and sparsity proportions are further modeled through systems of ordinary differential equations. The inference of the proposed model relies on an original variational EM procedure whose maximization step trains fully connected neural networks in order to solve the dynamical systems. The usual model selection question is also considered here through an original approach exploiting the ability of the model to work with empty clusters, thanks to the flexibility of deep neural networks. Numerical experiments on simulated and real world data sets demonstrate the effectiveness of the proposed methodology in the context of count data.

International audience; Inspired by the strengths of quadric error metrics initially designed for mesh decimation, we propose a concise mesh reconstruction approach for 3D point clouds. Our approach proceeds by clustering the input points enriched with quadric error metrics, where the generator of each cluster is the optimal 3D point for the sum of its quadric error metrics. This approach favors the placement of generators on sharp features, and tends to equidistribute the error among clusters. We reconstruct the output surface mesh from the adjacency between clusters and a constrained binary solver. We combine our clustering process with an adaptive refinement driven by the error. Compared to prior art, our method avoids dense reconstruction prior to simplification and produces immediately an optimized mesh.

International audience; The Shapes Constraint Language (SHACL) is a W3C recommendation which allows to represent constraints in RDF-shape graphs-, and validate RDF data graphs against these constraints. A SHACL validator produces a validation report whose result is false for a shape graph as soon as there is at least one node in the RDF data graph that does not conform to the shape. This Boolean result of the validation of an RDF data graph against an RDF shape graph is not suitable for discovering new high-potential shapes from the RDF data. In this paper, we propose a probabilistic framework to accept shapes with a realistic proportion of nodes in an RDF data graph that does not conform to it. Based on this framework, we propose an extension of the SHACL validation report to express a set of metrics including the generality and likelihood of shapes and we define a method to test a shape as a hypothesis test. Finally, we present the results of experiments conducted to validate a test RDF data graph against a set of shapes.

International audience; Videos and images from camera traps are more and more used by ecologists to estimate the population of species on a territory. It is a laborious work since experts have to analyse massive data sets manually. This takes also a lot of time to filter these videos when many of them do not contain animals or are with human presence. Fortunately, deep learning algorithms for object detection can help ecologists to identify multiple relevant species on their data and to estimate their population. In this study, we propose to go even further by using object detection model to detect, classify and count species on camera traps videos. To this end, we developed a 3-step process: (i) At the first stage, after splitting videos into images, we annotate images by associating bounding boxes to each label thanks to MegaDetector algorithm; (ii) then, we extend MegaDetector based on Faster R-CNN architecture with backbone Inception-ResNet-v2 in order to not only detect the 13 relevant classes but also to classify them; (iii) finally, we design a method to count individuals based on the maximum number of bounding boxes detected. This final stage of counting is evaluated in two different contexts: first including only detection results (i.e. comparing our predictions against the right number of individuals, no matter their true class), then an evolved version including both detection and classification results (i.e. comparing our predictions against the right number in the right class). The results obtained during the evaluation of our model on the test data set are: (i) 73,92\% mAP for classification, (ii) 96,88\% mAP for detection with a ratio Intersection-Over-Union (IoU) of 0.5 (overlapping ratio between groundtruth bounding box and the detected one), and (iii) 89,24\% mAP for detection at IoU=0.75. Highly represented classes, like humans, have highest values of mAP around 81\% whereas less represented classes in the train data set, such as dogs, have lowest values of mAP around 66\%. Regarding the proposed counting method, we predicted a count either exact or $\pm$ 1 unit for 87\% with detection results and for 48\% with detection and classification results of our test data set. Our model is also able to detect empty videos. To the best of our knowledge, this is the first study in France about the use of object detection model on a French national park to locate, identify and estimate the population of species from camera trap videos.

The treatment of abdominal aortic aneurysm relies on surgical repair and the indication mainly depends on its size evaluated by the maximal diameter (Dmax). The aim of this study was to evaluate a new automatic method based on artificial intelligence to measure the Dmax on computed tomography angiography.A fully automatic segmentation of the vascular system was performed using a hybrid method combining expert system with supervised deep learning. The aorta centreline was extracted from the segmented aorta and the aortic diameters were automatically calculated. Results were compared to manual segmentation performed by two human operators.The median absolute error between the two human operators was 1.2 mm (IQR 0.5-1.9). The automatic method using the deep learning algorithm demonstrated correlation with the human segmentation, with a median absolute error of 0.8 (0.5-4.2) mm and a coefficient correlation of 0.91 (P0.001).Although validation in larger cohorts is required, this method brings perspectives to develop new tools to standardize and automate the measurement of abdominal aortic aneurysm Dmax in order to help clinicians in the decision-making process.

International audience; Direct kinematics (DK) is one of the most challenging problem for cable-driven parallel robot (CDPR) with sagging cables. Solving the DK in real-time is not an issue provided that a guess of the solution is available. But difficulties arise when all DK solutions have to be determined (e.g. in the design phase of the CDPR). Continuation and interval analysis have been proposed to find the solutions but they are computer intensive. A preliminary investigation on the use of classical neural networks (NN) for the DK has shown that they were performing poorly. We present in this paper several methodological improvements that allows to get on average 99.95% of the exact DK solutions in about 5 seconds. Still this result is not completely satisfactory and we present possible axis to obtain better results in terms of exact results and multiple solutions.

Topological data analysis (TDA) is an area of data science that focuses on using invariants from algebraic topology to provide multiscale shape descriptors for geometric data sets such as point clouds. One of the most important such descriptors is {\em persistent homology}, which encodes the change in shape as a filtration parameter changes; a typical parameter is the feature scale. For many data sets, it is useful to simultaneously vary multiple filtration parameters, for example feature scale and density. While the theoretical properties of single parameter persistent homology are well understood, less is known about the multiparameter case. In particular, a central question is the problem of representing multiparameter persistent homology by elements of a vector space for integration with standard machine learning algorithms. Existing approaches to this problem either ignore most of the multiparameter information to reduce to the one-parameter case or are heuristic and potentially unstable in the face of noise. In this article, we introduce a new general representation framework that leverages recent results on {\em decompositions} of multiparameter persistent homology. This framework is rich in information, fast to compute, and encompasses previous approaches. Moreover, we establish theoretical stability guarantees under this framework as well as efficient algorithms for practical computation, making this framework an applicable and versatile tool for analyzing geometric and point cloud data. We validate our stability results and algorithms with numerical experiments that demonstrate statistical convergence, prediction accuracy, and fast running times on several real data sets.

Persistent homology (PH) provides topological descriptors for geometric data, such as weighted graphs, which are interpretable, stable to perturbations, and invariant under, e.g., relabeling. Most applications of PH focus on the one-parameter case -- where the descriptors summarize the changes in topology of data as it is filtered by a single quantity of interest -- and there is now a wide array of methods enabling the use of one-parameter PH descriptors in data science, which rely on the stable vectorization of these descriptors as elements of a Hilbert space. Although the multiparameter PH (MPH) of data that is filtered by several quantities of interest encodes much richer information than its one-parameter counterpart, the scarceness of stability results for MPH descriptors has so far limited the available options for the stable vectorization of MPH. In this paper, we aim to bring together the best of both worlds by showing how the interpretation of signed barcodes -- a recent family of MPH descriptors -- as signed measures leads to natural extensions of vectorization strategies from one parameter to multiple parameters. The resulting feature vectors are easy to define and to compute, and provably stable. While, as a proof of concept, we focus on simple choices of signed barcodes and vectorizations, we already see notable performance improvements when comparing our feature vectors to state-of-the-art topology-based methods on various types of data., 23 pages, 3 figures, 8 tables

The purpose of this thesis is to develop machine learning models that can leverage histology slides and clinical data to predict the outcome of immunotherapies against lung cancer. To this end, there are several challenges to overcome, such as the concurrent classification and localization of information within whole-slide images of large size, or the interpretability of the predictions made by the models. The thesis proposes several contributions to address such challenges.Chapter 2 introduces the concept of mixed supervision in histopathology. The purpose of this framework is to leverage several levels of supervision (i.e., global and local supervision) to make the model more efficient in both classification and localization tasks. Set on an attention-based deep learning model fit for global and local classification of tissue in whole-slide images, we show that it is possible to improve not only the model slide-level classification performance, but also and most importantly its ability to accurately locate regions of interest in the tissue, with only a few available local annotations.Chapter 3 extends the work presented in chapter 2, by consolidating the simultaneous classification and localization branches of the model with tailored loss functions that enforce attention distribution to follow the actual label distribution with respect to the available annotations. A slide sampling strategy is also proposed to strengthen the localization performance, and simplify the training procedure to have it fit in a single process.In Chapter 4, we present a multicentric lung cancer dataset dedicated to the prediction of immunotherapy response. We document the various steps followed to filter out low-quality samples, as well as undetermined outcomes, and we discuss the interpretation of what is a positive or a negative response with respect to the current gold standard clinical evaluations. Finally, we evaluate several models to directly predict the treatment response out of pathology slides, and discuss the caveats of the tested approaches.In Chapter 5, we switch from binary treatment response prediction to survival analysis, and use contrastive learning along with deep nonparametric clustering to generate a set of prognostic features in an unsupervised manner. We show that the obtained set of characteristics is a powerful indicator of survival, and that it maintains a good level of performance when picking one acquisition center as the test set. We also discuss the histological interpretation of the most prominent discovered clusters.To conclude, we address the remaining issues and challenges, and debate what future directions could be taken in order to tackle unanswered questions.; L’objectif de cette thèse est de développer des modèles d’apprentissage automatique capables d’exploiter des lames histologiques et des données cliniques pour prédire le résultat des traitements par immunothérapie contre le cancer du poumon. À cette fin, plusieurs défis doivent être relevés, tels que la classification et la localisation simultanées d’informations dans des images de lames entières de grande taille, ou l’interprétation des prédictions faites par les modèles. Dans ce qui suit, nous proposerons plusieurs contributions pour relever ces défis.Le chapitre 2 introduit le concept de supervision mélangée en histopathologie. L’objectif de cette méthode est de tirer parti de plusieurs niveaux de supervision (c’est-à-dire la supervision globale et locale) pour rendre le modèle plus efficace à la fois en classification et en localisation. Sur la base d’un modèle d’apprentissage profond basé sur l’attention et adapté à la classification globale et locale de tissu dans des coupes histologiques, nous montrons qu’il est possible d’améliorer non seulement les performances du modèle en matière de classification des lames, mais aussi et surtout sa capacité à localiser avec précision les régions d’intérêt dans le tissu disponible, lorsque seules quelques annotations disponibles.Le chapitre 3 étend le travail présenté dans le chapitre 2, en consolidant les branches de classification et de localisation simultanées du modèle avec des fonctions de coût adaptées qui contraignent la distribution de l’attention à suivre la distribution réelle des labels d’après les annotations disponibles. Une stratégie d’échantillonnage des images est également proposée pour renforcer les performances de localisation et simplifier la procédure d’apprentissage afin qu’elle s’inscrive dans un processus unique.Dans le chapitre 4, nous présentons un ensemble de données multicentriques sur le cancer du poumon dédié à la prédiction de la réponse à l’immunothérapie. Nous documentons les différentes étapes suivies pour éliminer les échantillons de faible qualité, ainsi que les cas indéterminés, et nous discutons de la définition de ce qu’est une réponse positive ou négative par rapport aux évaluations cliniques actuelles de référence. Enfin, nous évaluons plusieurs modèles permettant de prédire directement la réponse au traitement à partir de lames histologiques et discutons des écueils des approches envisagées. Dans le chapitre 5, nous passons de la prédiction binaire de la réponse au traitement à la prédiction de survie, et nous utilisons l’apprentissage par contraste ainsi que le regroupement non paramétrique profond pour générer un ensemble de caractéristiques pronostiques de manière non supervisée. Nous montrons que l’ensemble des caractéristiques obtenu est un puissant indicateur de survie et qu’il conserve un bon niveau de performance lorsque l’on choisit un seul centre comme ensemble de test. Nous discutons également de l’interprétation histologique faite du résultat de l’algorithme de regroupement, en particulier pour les groupes les plus corrélés à la survie.Pour conclure, nous abordons les questions et les défis en suspens, et nous discutons des orientations futures qui pourraient être prises afin de répondre aux questions restées sans réponse.

Motivated by the algorithmic study of 3-dimensional manifolds, we explore the structural relationship between the JSJ decomposition of a given 3-manifold and its triangulations. Building on work of Bachman, Derby-Talbot and Sedgwick, we show that a "sufficiently complicated" JSJ decomposition of a 3-manifold enforces a "complicated structure" for all of its triangulations. More concretely, we show that, under certain conditions, the treewidth (resp. pathwidth) of the graph that captures the incidences between the pieces of the JSJ decomposition of an irreducible, closed, orientable 3-manifold M yields a linear lower bound on its treewidth tw (M) (resp. pathwidth pw(M)), defined as the smallest treewidth (resp. pathwidth) of the dual graph of any triangulation of M. We present several applications of this result. We give the first example of an infinite family of bounded-treewidth 3-manifolds with unbounded pathwidth. We construct Haken 3-manifolds with arbitrarily large treewidth - previously the existence of such 3-manifolds was only known in the non-Haken case. We also show that the problem of providing a constant-factor approximation for the treewidth (resp. pathwidth) of bounded-degree graphs efficiently reduces to computing a constant-factor approximation for the treewidth (resp. pathwidth) of 3-manifolds., LIPIcs, Vol. 258, 39th International Symposium on Computational Geometry (SoCG 2023), pages 42:1-42:18

Flags are sequences of nested linear subspaces of increasing dimension. They belong to smooth manifolds generalizing Grassmannians and bring a richer multi-scale point of view to the traditional subspace methods in statistical analysis. Hence, there is an increasing interest in generalizing the formulae and statistical methods already developed for Grassmannians to flag manifolds. In particular, it is critical to compute accurately and efficiently the geodesic distance and the logarithm due to their fundamental importance in geometric statistics. However, there is no explicit expression known in the case of flags. In this work, we exploit the homogeneous quotient space structure of flag manifolds and rethink the geodesic endpoint problem as an alignment of orthogonal matrices on their equivalence classes. The relaxed problem with the Frobenius metric surprisingly enjoys an explicit solution. This is the key to modify a previously proposed algorithm. We show that our explicit alignment step brings drastic improvements in accuracy, speed and radius of convergence, in addition to overcoming the combinatorial issues raised by the non-connectedness of the equivalence classes.

The query of causality is of paramount importance in biomedical data analysis: assessing the causal relationships between the observed variables allows to improve our understanding of the tackled medical condition and better support decision-making. Torsade de Pointes (TdP) is an extremely serious drug-induced cardiac side effect, which can provoke ventricular fibrillation and lead to sudden death. TdP is related to abnormal repolarizations in single cells, and the minimum set of ion channels needed to correctly assess TdP risk is still an open question. Discovering causal relations between drug-induced ionic channels' perturbations could shed new light on the underlying mechanisms leading to TdP, and drive variable selection to improve TdP-risk assessment. In this work, we propose to apply the causal discovery method ICA-Linear Non-Gaussian Acyclic Model (ICA-LiNGAM) to uncover the relations across the 7 ion channels identified by the Comprehensive in vitro Proarrhythmia Assay (CiPA) initiative as potentially related to the induction of TdP: I Kr , I Na , I NaL , I CaL , I K1 , I Ks and I to. The obtained Bayesian causal network can be explored to infer the downstream impact of the ionic currents in the drug's safety label. We consider 109 drugs of known torsadogenic risk (51 unsafe) listed by CredibleMeds. We identify I Kr , I NaL and I CaL as the ones which directly affect TdP-risk assessment, and suggest that I Na perturbations could potentially have a high impact on proarrhythmic risk induction. Our causalitybased results were further confirmed by independently performing binary drug risk classification, which shows that the combination of the 3 selected ions maximizes the classification accuracy and specificity, outperforming state-ofthe-art approaches based on alternative ion channel combinations.

International audience; As online content continues to grow, so does the spread of online hate, especially on social media. Most research efforts conducted on the task of bullying participant role identification are directed towards social networks such as Twitter and Instagram. However, private instant messaging platforms and channels were pinpointed in recent studies as the most prominent grounds for cyberbullying, especially among teens. Since data collection from major social media platforms is strictly limited, very few studies have investigated this task in a multi-party setting. However, the recent release of resources mimicking online aggression situations that may occur among teens on private instant messaging platforms contributes to filling this gap. In this study, we introduce a full pipeline aiming at automating the identification of bullying participant roles (bully and victim) in multi-party chats. Leveraging pre-trained language models and different learning frameworks, we perform hateful content classification of exchanged messages according to a binary scheme (online hate or no online hate). Then,-from these bullying tracesbullying behavioural cues (repetition and intention to harm) are derived and formalised into a role scoring function. As a result, the proposed pipeline identifies the bully and the victim among chat participants. Evaluated against state-ofthe-art methods, the proposed pipeline achieves better performances considering all the datasets and roles to predict. In addition, the error analysis confirms that deriving bullying behavioural cues is beneficial to the task of participant role identification.

The real-world implementation of federated learning is complex and requires research and development actions at the crossroad between different domains ranging from data science, to software programming, networking, and security. While today several FL libraries are proposed to data scientists and users, most of these frameworks are not designed to find seamless application in medical use-cases, due to the specific challenges and requirements of working with medical data and hospital infrastructures. Moreover, governance, design principles, and security assumptions of these frameworks are generally not clearly illustrated, thus preventing the adoption in sensitive applications. Motivated by the current technological landscape of FL in healthcare, in this document we present Fed-BioMed: a research and development initiative aiming at translating federated learning (FL) into real-world medical research applications. We describe our design space, targeted users, domain constraints, and how these factors affect our current and future software architecture.

International audience; Phylogenetic PCA (p-PCA) is a version of PCA for observations that are leaf nodes of a phylogenetic tree. P-PCA accounts for the fact that such observations are not independent, due to shared evolutionary history. The method works on Euclidean data, but in evolutionary biology there is a need for applying it to data on manifolds, particularly shapes. We provide a generalization of p-PCA to data lying on Riemannian manifolds, called Tangent p-PCA. Tangent p-PCA thus makes it possible to perform dimension reduction on a data set of shapes, taking into account both the non-linear structure of the shape space as well as phylogenetic covariance. We show simulation results on the sphere, demonstrating well-behaved error distributions and fast convergence of estimators. Furthermore, we apply the method to a data set of mammal jaws, represented as points on a landmark manifold equipped with the LDDMM metric.

International audience; In this study, we explore the value of using a recently proposed multimodal learning method as an initialization for anomaly detection in abdominal ultrasound images. The method efficiently learns visual concepts from radiological reports using natural language supervision and constrastive learning. The underlying requirement of the method is simply the availability of image and textual descriptions pairs. However, in abdominal ultrasound examinations, radiological reports are associated with several images and describe all organs observed during the examination. To address this shortcoming, we automatically construct image and text pairs using 1) deep clustering for abdominal organ classification on ultrasound images and 2) natural language processing tools to extract the corresponding description on the report. We show that pre-training the model with these constructed pairs yields representations that better separate normal classes from abnormal ones on ultrasound images for the kidneys, compared to ImageNet-based representations, with a 10% improvement in macro-average accuracy.

International audience; This paper investigates when the importance weighting (IW) correction is needed to address covariate shift, a common situation in supervised learning where the input distributions of training and test data differ. Classic results show that the IW correction is needed when the model is parametric and misspecified. In contrast, recent results indicate that the IW correction may not be necessary when the model is nonparametric and well-specified. We examine the missing case in the literature where the model is nonparametric and misspecified, and show that the IW correction is needed for obtaining the best approximation of the true unknown function for the test distribution. We do this by analyzing IW-corrected kernel ridge regression, covering a variety of settings, including parametric and nonparametric models, well-specified and misspecified settings, and arbitrary weighting functions.

Federated learning allows for the training of machine learn- ing models on multiple decentralized local datasets without requiring explicit data exchange. However, data pre-processing, including strate- gies for handling missing data, remains a major bottleneck in real-world federated learning deployment, and is typically performed locally. This approach may be biased, since the subpopulations locally observed at each center may not be representative of the overall one. To address this issue, this paper first proposes a more consistent approach to data stan- dardization through a federated model. Additionally, we propose Fed- MIWAE, a federated version of the state-of-the-art imputation method MIWAE, a deep latent variable model for missing data imputation based on variational autoencoders. MIWAE has the great advantage of being easily trainable with classical federated aggregators. Furthermore, it is able to deal with MAR (Missing At Random) data, a more challenging missing-data mechanism than MCAR (Missing Completely At Random), where the missingness of a variable can depend on the observed ones. We evaluate our method on multi-modal medical imaging data and clinical scores from a simulated federated scenario with the ADNI dataset. We compare Fed-MIWAE with respect to classical imputation methods, ei- ther performed locally or in a centralized fashion. Fed-MIWAE allows to achieve imputation accuracy comparable with the best centralized method, even when local data distributions are highly heterogeneous. In addition, thanks to the variational nature of Fed-MIWAE, our method is designed to perform multiple imputation, allowing for the quantification of the imputation uncertainty in the federated scenario.

International audience; Isomanifolds are the generalization of isosurfaces to arbitrary dimension and codimension, i.e. submanifolds of $\mathbb{R}^d$ defined as the zero set of a multivariate multivalued smooth function $f: \mathbb{R}^d\rightarrow \mathbb{R}^{d-n}$, where $n$ is the intrinsic dimension of the manifold and we assume that $0$ is a regular value. A natural way to approximate a smooth isomanifold $\mathcal{M}=f^{-1}(0)$ is to consider its Piecewise-Linear (PL) approximation $\hat{\mathcal{M}}$ based on atriangulation $\mathcal{T}$ of the ambient space $\mathbb{R}^d$, whose longest edge has length $D$. In thispaper, we describe a simple algorithm to trace isomanifolds from a given starting point {on each connected component}. The algorithm works for arbitrary dimensions $n$ and $d$, and $D$. Our main result is that, when$f$ (or $\mathcal{M}$) has bounded complexity, the complexity of the algorithm is polynomial in $d$ and $\delta=1/D$ (and unavoidably exponential in $n$). Since it is known that for $\delta = \Omega (d^{2.5})$, $\hat{\mathcal{M}}$ is $O(D^2)$-close and isotopic to $\mathcal{M}$, our algorithm produces a faithful PL-approximation of isomanifolds of bounded complexity in time polynomial in $d$.Combining this algorithm with dimensionality reduction techniques, the dependency on $d$ in the size of $\mathcal{M}$ can be completely removed with high probability. We also show that the algorithm can handleisomanifolds with boundary and, more generally, isostratifolds. The algorithm for isomanifolds with boundary has been implemented and experimental results are reported, showing that it is practical and can handle cases that are far ahead of the state-of-the-art.

SUMMARY Secondary microseisms are the most energetic noise in continuous seismometer recordings. They are generated by interactions between ocean waves, including between gravity waves incident on and reflected from the coast. Coastal reflections of ocean waves leading to coastal microseismic sources are hard to estimate in various global numerical wave models, and independent quantification of these coastal sources through direct measurements can greatly improve these models. Here, we exploit a 41-km-long submarine optical fibre (OF) cable located offshore Toulon, France, using distributed acoustic sensing (DAS). We record both the amplitude and frequency of seafloor strains induced by ocean surface gravity waves, as well as secondary microseisms caused by the interaction of gravity waves incident and reflected from the coast. By leveraging the spatially distributed nature of DAS measurements, additional fundamental information is recovered such as the velocity and azimuth of the waves. We find that on average 30 per cent of the gravity waves are reflected at the coast generating local sources of secondary microseisms that manifest as Scholte waves. These local sources represent the most energetic contribution to the seismic noise recorded along the OF and by an onshore broad-band station located near the DAS interrogator. Furthermore, we estimate a coastal reflection coefficient of ocean surface gravity waves R2 of about 0.07, which provides improved constraints for seismic noise generation models. In addition, we show that new local sources of microseisms can be generated when gravity waves characteristics (azimuth and frequency content) change and lead to some delays between the OF cable and buoy recordings. These analyses pave the way for a wide use of DAS data to monitor ocean–solid earth interactions as they provide a wealth of information on the reflection of gravity waves, coastal microseismic sources, and new constraints for numerical models of microseismic noise.

International audience; Real-time globally consistent GPS tracking is critical for an accurate localization and is crucial for applications such as autonomous navigation or multi-robot mapping. However, under challenging environment conditions such as indoor/outdoor transitions, GPS signals are partially available or not consistent over time. In this paper, a real-time tracking system for continuously locating emergency response agents in challenging conditions is presented. A cooperative localization method based on Laser-Visual-Inertial (LVI) and GPS sensors is achieved by communicating optimization events between a LiDAR-Inertial-SLAM (LI-SLAM) and Visual-Inertial-SLAM (VI-SLAM) that operate simultaneously. The estimation of the pose assisted by multiple SLAM approaches provides the GPS localization of the agent when a stand-alone GPS fails. The system has been tested under the terms of the MALIN Challenge, which aims to globally localize agents across outdoor and indoor environments under challenging conditions (such as smoked rooms, stairs, indoor/outdoor transitions, repetitive patterns, extreme lighting changes) where it is well known that a stand-alone SLAM will not be enough to maintaining the localization. The system achieved Absolute Trajectory Error of 0.48%, with a pose update rate between 15 and 20 Hz. Furthermore, the system is able to build a global consistent 3D LiDAR Map that is post-processed to create a 3D reconstruction at different level of details.

International audience; Over the last years, Remote Sensing Images (RSI) analysis have started resorting to using deep neural networks to solve most of the commonly faced problems, such as detection, land cover classification or segmentation. As far as critical decision making can be based upon the results of RSI analysis, it is important to clearly identify and understand potential security threats occurring in those machine learning algorithms. Notably, it has recently been found that neural networks are particularly sensitive to carefully designed attacks, generally crafted given the full knowledge of the considered deep network. In this paper, we consider the more realistic but challenging case where one wants to generate such attacks in the case of a black-box neural network. In this case, only the prediction score of the network is accessible, given a specific input. Examples that lure away the network's prediction, while being perceptually similar to real images, are called natural or unrestricted adversarial examples. We present an original method to generate such examples, based on a variant of the Wasserstein Generative Adversarial Network. We demonstrate its effectiveness on natural adversarial hyper-spectral image generation and image modification for fooling a state-of-the-art detector. Among others, we also conduct a perceptual evaluation with human annotators to better assess the effectiveness of the proposed method.

In many areas such as computational biology, finance or social sciences, knowledge of an underlying graph explaining the interactions between agents is of paramount importance but still challenging. Considering that these interactions may be based on nonlinear relationships adds further complexity to the topology inference problem. Among the latest methods that respond to this need is a topology inference one proposed by the authors, which estimates a possibly directed adjacency matrix in an online manner. Contrasting with previous approaches based on linear models, the considered model is able to explain nonlinear interactions between the agents in a network. The novelty in the considered method is the use of a derivative-reproducing property to enforce network sparsity, while reproducing kernels are used to model the nonlinear interactions. The aim of this paper is to present a thorough convergence analysis of this method. The analysis is proven to be sane both in the mean and mean square sense. In addition, stability conditions are devised to ensure the convergence of the analyzed method., 24 pages, 11 of which of which are supplementary material; 4 figures

Noisy labels often occur in vision datasets, especially when they are obtained from crowdsourcing or Web scraping. We propose a new regularization method, which enables learning robust classifiers in presence of noisy data. To achieve this goal, we propose a new adversarial regularization {scheme} based on the Wasserstein distance. Using this distance allows taking into account specific relations between classes by leveraging the geometric properties of the labels space. {Our Wasserstein Adversarial Regularization (WAR) encodes a selective regularization, which promotes smoothness of the classifier between some classes, while preserving sufficient complexity of the decision boundary between others. We first discuss how and why adversarial regularization can be used in the context of noise and then show the effectiveness of our method on five datasets corrupted with noisy labels: in both benchmarks and real datasets, WAR outperforms the state-of-the-art competitors.

With the enthusiasm for digital twins in various domains, surface reconstruction from raw 3D point clouds is increasingly demanded while facing multifaceted challenges. Low cost data acquisition devices make it possible to obtain 3D scans easily and quickly, but may lead to defect-laden point clouds. In addition, users expect lightweight and controllable surface re- construction approaches with personalized criteria depending on their applications. This thesis aims at addressing these issues by contributing several approaches to the problem of surface reconstruction from two aspects: (1) point cloud comprehension and (2) surface reconstruction with a priori information.Concretely, we first propose the notion of progressive discrete domain for global implicit re- construction approaches that refines and optimizes a discrete 3D domain in accordance with both input and output, and to user-defined criteria. Based on such a domain discretization, we devise a progressive primitive-aware surface reconstruction approach with the capacity to refine the implicit function and its representation, in which the most ill-posed parts of the re- construction problem are postponed to later stages of the reconstruction, and where the fine geometric details are resolved after discovering the topology. Secondly, we contribute a deep learning-based approach that learns to detect and consolidate sharp feature points on raw 3D point clouds, whose results can be taken as additional inputs to consolidate sharp features for the previous reconstruction approach. Finally, we contribute a variational shape reconstruction method from unoriented point clouds, leveraging both quadric error metrics (QEM) and clus- tering through variational partitioning. The output triangle mesh is concise and anisotropic, with mesh vertices equidistributed along sharp creases.In summary, this thesis seeks effective surface reconstruction from a global and progressive perspective. By combining multiple priors and designing meaningful criteria, the contributed approaches can deal with various defects and multi-scale features.; Avec l’enthousiasme pour les jumeaux numériques dans divers domaines, la reconstruction de surfaces à partir de nuages de points 3D est de plus en plus nécessaire, tout en faisant face à des défis scientifiques multiformes. Les dispositifs d’acquisition de données à faible coût permettent d’obtenir facilement et rapidement des scans 3D, mais peuvent conduire à des nuages de points 3D avec défauts. De plus, les utilisateurs attendent des approches de reconstruction de surface légères et contrôlables avec des critères personnalisés en fonction de leurs applications. Cette thèse vise à aborder ces questions en apportant plusieurs approches au problème de la reconstruction de surface sous deux aspects : (1) l’analyse de nuages de points, et (2) la reconstruction de surface avec des informations a priori.Nous proposons d’abord une discrétisation progressive (grossier vers fin) du domaine pour les approches globales implicite, leur permettant de gérer divers défauts. Cette méthode est étendue pour être informée par des primitives détectées dans le nuage de points. Ensuite, nous présentons une approche basée sur l’apprentissage profond qui apprend à détecter et à consoli- der des points appartenant aux singularités (coins, arêtes vives) dans le nuage de points 3D. Ces point singuliers sont ensuite considérés comme une entrée supplémentaire à la méthode de re- construction progressive. Enfin, nous apportons une méthode de reconstruction variationnelle à partir de nuages de points non-orientés, tirant parti à la fois des métriques d’erreur quadra- tiques et du clustering via partitionnement variationnel. Le maillage triangulaire de sortie est concis et anisotrope, avec des sommets de maillage répartis sur les arêtes vives.En résumé, cette thèse vise à contribuer des approches efficientes pour la reconstruction de surfaces dans une perspective globale et progressive. En combinant plusieurs a priori et en prenant en compte différents critères, les méthodes proposées peuvent traiter divers défauts et détails à plusieurs échelles.

International audience; Recent studies have highlighted the importance to reach a fine-grained online hate speech characterisation to better understand how hate is conveyed, especially on social media. A key element in this scenario is the identification and characterisation of the hate speech target community, e.g., national, ethnic, religious minorities. In this paper, we propose a full pipeline relying on unsupervised methods to distinguish specific hate speech manifestations, i.e., targeted (group of) victim(s) and the protected characteristics (target-types) discriminated. Our contribution is threefold: (1) we leverage multiple data views to contrast different abusive behaviours; (2) we explore the use of clustering techniques to perform fine-grained hate speech target community detection, and (3) we address an in-depth content analysis of the generated hate speech target communities. Relying on multiple data views derived from multilingual pre-trained language models (i.e., multilingual BERT and multilingual Universal Sentence Encoder) and the Multi-view Spectral Clustering (MvSC) algorithm, the 69 experiments performed on the Multilingual Hate Speech dataset (MLMA) of tweets show that most of the configurations of the proposed pipeline significantly outperforms state-of-the-art clustering algorithms on French and English. Our experiments confirm the ability of the proposed approach to capture complex hate speech phenomena (i.e., intersections between victim-groups, target-types or both).

International audience; Symmetric Positive Definite (SPD) matrices are ubiquitous in data analysis under the form of covariance matrices or correlation matrices. Several O(n)-invariant Riemannian metrics were defined on the SPD cone, in particular the kernel metrics introduced by Hiai and Petz. The class of kernel metrics interpolates between many classical O(n)-invariant metrics and it satisfies key results of stability and completeness. However, it does not contain all the classical O(n)-invariant metrics. Therefore in this work, we investigate super-classes of kernel metrics and we study which key results remain true. We also introduce an additional key result called cometric-stability, a crucial property to implement geodesics with a Hamiltonian formulation. Our method to build intermediate embedded classes between O(n)-invariant metrics and kernel metrics is to give a characterization of the whole class of O(n)-invariant metrics on SPD matrices and to specify requirements on metrics one by one until we reach kernel metrics. As a secondary contribution, we synthesize the literature on the main O(n)-invariant metrics, we provide the complete formula of the sectional curvature of the affine-invariant metric and the formula of the geodesic parallel transport between commuting matrices for the Bures-Wasserstein metric.

International audience; The automatic generation of explanations to improve the transparency of machine predictions is a major challenge in Artificial Intelligence. Such explanations may also be effectively applied to other decision making processes where it is crucial to improve critical thinking in human beings. An example of that consists in the clinical cases proposed to medical residents together with a set of possible diseases to be diagnosed, where only one correct answer exists. The main goal is not to identify the correct answer, but to be able to explain why one is the correct answer and the others are not. In this paper, we propose a novel approach to generate argument-based natural language explanations for the correct and incorrect answers of standardized medical exams. By combining information extraction methods from heterogeneous medical knowledge bases, we propose an automatic approach where the symptoms relevant to the correct diagnosis are automatically extracted from the case, to build a natural language explanation. To do so, we annotated a new resource of 314 clinical cases, where 1843 different symptoms are identified. Results in retrieving and matching the relevant symptoms for the clinical cases to support the correct diagnosis and contrast incorrect ones outperform standard baselines.

International audience; Linked Multi-Agent Systems (MAS) is a vision where MAS are linked through their resources and representations, supporting knowledge augmentation for human and software agents on the Web and the linking and interoperability of different MAS. The goal of Linked MAS should not be to have one MAS platform or one MAS system on the Web but to allow us to turn the Web into an architecture supporting a network of interoperable MAS. The different statements submitted by participants of this Dagstuhl-Seminar 23081 show the variety of challenges to face and in this short position statement I will add a few ones to the list: the need for a viral first use case, the metadata specifically needed for hMAS, the need to think the insertion of hMAS in the initiatives for (re)decentralizing the Web and the need to make hMAS a usable extension of the Web.

In this note, we study the connection between Gray tensor product and suspension. We derive a characterization of weak equivalences as fully faithful and essentially surjective functors. We construct the $co$ duality, a weak involution that reverses the direction of even dimensional cells. We conclude by studying Grothendieck fibrations of complicial sets., Comment: This text is still in draft form. A new version should quickly replace this one

International audience; The large and still increasing popularity of deep learning clashes with a major limit of neural network architectures, that consists in their lack of capability in providing human-understandable motivations of their decisions. In situations in which the machine is expected to support the decision of human experts, providing a comprehensible explanation is a feature of crucial importance. The language used to communicate the explanations must be formal enough to be implementable in a machine and friendly enough to be understandable by a wide audience. In this paper, we propose a general approach to Explainable Artificial Intelligence in the case of neural architectures, showing how a mindful design of the networks leads to a family of interpretable deep learning models called Logic Explained Networks (LENs). LENs only require their inputs to be human-understandable predicates, and they provide explanations in terms of simple First-Order Logic (FOL) formulas involving such predicates. LENs are general enough to cover a large number of scenarios. Amongst them, we consider the case in which LENs are directly used as special classifiers with the capability of being explainable, or when they act as additional networks with the role of creating the conditions for making a black-box classifier explainable by FOL formulas. Despite supervised learning problems are mostly emphasized, we also show that LENs can learn and provide explanations in unsupervised learning settings. Experimental results on several datasets and tasks show that LENs may yield better classifications than established white-box models, such as decision trees and Bayesian rule lists, while providing more compact and meaningful explanations.

Mobile Access to the Web of Data is currently a real challenge in developing countries, mainly characterized by limited Internet connectivity and high penetration of mobile devices with the limited resources (such as cache and memory). In this paper, we survey and compare proposed solutions (such as models and architectures) that could contribute to solving this problem of mobile access to the Web of Data with intermittent Internet access. These solutions are discussed in relation to the underlying network architectures and data models considered. We present a conceptual study of peer-to-peer solutions based on gossip protocols dedicated to design the connected overlay networks. In addition, we provide a detailed analysis of client-server and data replication systems generally designed to ensure the local availability of data on the system. We conclude with some recommendations to achieve a connected architecture that provides mobile contributors with local access to the Web of data.

International audience; In recent years, a large number of RDF datasets have been built and published on the Web in fields as diverse as linguistics or life sciences, as well as general datasets such as DBpedia or Wikidata. The joint exploitation of these datasets requires specific knowledge about their content, access points, and commonalities. However, not all datasets contain a self-description, and not all access points can handle the complex queries used to generate such a description. In this article, we provide a standard-based approach to generate the description of a dataset. The generated descriptions as well as the process of their computation are expressed using standard vocabularies and languages. We implemented our approach into a framework, called IndeGx, where each indexing feature and its computation is collaboratively and declaratively defined in a GitHub repository. We have experimented IndeGx on a set of 339 RDF datasets with endpoints listed in public catalogs, over 8 months. The results show that we can collect, as much as possible, important characteristics of the datasets depending on their availability and capacities. The resulting index captures the commonalities, variety and disparity in the offered content and services and it provides an important support to any application designed to query RDF datasets.

Computing persistent homology using Gaussian kernels is useful in the domains of topological data analysis and machine learning as shown by Phillips, Wang and Zheng [SoCG 2015]. However, contrary to the case of computing persistent homology using the Euclidean distance or even the $k$-distance, it is not known how to compute the persistent homology of high dimensional data using Gaussian kernels. In this paper, we consider a power distance version of the Gaussian kernel distance (GKPD) given by Phillips, Wang and Zheng, and show that the persistent homology of the \v{C}ech filtration of $P$ computed using the GKPD is approximately preserved. For datasets in $d$-dimensional Euclidean space, under a relative error bound of $\varepsilon \in [0,1]$, we obtain a dimensionality of $(i)$ $O(\varepsilon^{-2}\log^2 n)$ for $n$-point datasets and $(ii)$ $O(D\varepsilon^{-2}\log (Dr/\varepsilon))$ for datasets having diameter $r$ (up to a scaling factor). We use two main ingredients. The first one is a new decomposition of the squared radii of \v{C}ech simplices using the kernel power distance, in terms of the pairwise GKPDs between the vertices, which we state and prove. The second one is the Random Fourier Features (RFF) map of Rahimi and Recht [NeurIPS 2007], as used by Chen and Phillips [ALT 2017].

This work focuses on unsupervised representation learning in person re-identification (ReID). Recent self-supervised contrastive learning methods learn invariance by maximizing the representation similarity between two augmented views of a same image. However, traditional data augmentation may bring to the fore undesirable distortions on identity features, which is not always favorable in id-sensitive ReID tasks. In this paper, we propose to replace traditional data augmentation with a generative adversarial network (GAN) that is targeted to generate augmented views for contrastive learning. A 3D mesh guided person image generator is proposed to disentangle a person image into id-related and id-unrelated features. Deviating from previous GAN-based ReID methods that only work in id-unrelated space (pose and camera style), we conduct GAN-based augmentation on both id-unrelated and id-related features. We further propose specific contrastive losses to help our network learn invariance from id-unrelated and id-related augmentations. By jointly training the generative and the contrastive modules, our method achieves new state-of-the-art unsupervised person ReID performance on mainstream large-scale benchmarks., Extension of conference paper arXiv:2012.09071. Accepted to TPAMI. Project page: https://github.com/chenhao2345/GCL-extended

International audience; Kleinjohann [1] and Bangert [2] extended the reach rch(S) from subsets S of Euclidean space to the reach rch M (S) of subsets S of Riemannian manifolds M, where M is smooth (we'll assume at least C^3). Bangert showed that sets of positive reach in Euclidean space and Riemannian manifolds are very similar. In this paper we introduce a slight variant of Kleinjohann's and Bangert's extension and quantify the similarity between sets of positive reach in Euclidean space and Riemannian manifolds in a new way: Given p ∈ M and q ∈ S, we bound the local feature size (a local version of the reach) of its lifting to the tangent space via the inverse exponential map (exp^{ −1}_p (S)) at q, assuming that rch M (S) and the geodesic distance d_ M (p, q) are bounded. These bounds are motivated by the importance of the reach and local feature size to manifold learning, topological inference, and triangulating manifolds and the fact that intrinsic approaches circumvent the curse of dimensionality.

The spatial resolution of images of living samples obtained by fluorescence microscopes is physically limited due to the diffraction of visible light, which makes the study of entities of size less than the diffraction barrier (around 200 nm in the x-y plane) very challenging. To overcome this limitation, several deconvolution and super-resolution techniques have been proposed. Within the framework of inverse problems, modern approaches in fluorescence microscopy reconstruct a super-resolved image from a temporal stack of frames by carefully designing suitable hand-crafted sparsity-promoting regularisers. Numerically, such approaches are solved by proximal gradient-based iterative schemes. Aiming at obtaining a reconstruction more adapted to sample geometries (e.g. thin filaments), we adopt a plug-and-play denoising approach with convergence guarantees and replace the proximity operator associated with the explicit image regulariser with an image denoiser (i.e. a pre-trained network) which, upon appropriate training, mimics the action of an implicit prior. To account for the independence of the fluctuations between molecules, the model relies on second-order statistics. The denoiser is then trained on covariance images coming from data representing sequences of fluctuating fluorescent molecules with filament structure. The method is evaluated on both simulated and real fluorescence microscopy images, showing its ability to correctly reconstruct filament structures with high values of peak signal-to-noise ratio (PSNR).

International audience; Given the size of digitized Whole Slide Images (WSIs), it is generally laborious and time-consuming for pathologists to exhaustively delineate objects within them, especially with datasets containing hundreds of slides to annotate. Most of the time, only slide-level labels are available, giving rise to the development of weakly-supervised models. However, it is often difficult to obtain from such models accurate object localization, e.g., patches with tumor cells in a tumor detection task, as they are mainly designed for slide-level classification. Using the attention-based deep Multiple Instance Learning (MIL) model as our base weakly-supervised model, we propose to use mixed supervision-i.e., the use of both slide-level and patch-level labels-to improve both the classification and the localization performances of the original model, using only a limited amount of patch-level labeled slides. In addition, we propose an attention loss term to regularize the attention between key instances, and a paired batch method to create balanced batches for the model. First, we show that the changes made to the model already improve its performance and interpretability in the weakly-supervised setting. Furthermore, when using only between 12 and 62% of the total available patch-level annotations, we can reach performance close to fully-supervised models on the tumor classification datasets DigestPath2019 and Camelyon16.

By interpreting the product of the Principal Component Analysis, that is the covariance matrix, as a sequence of nested subspaces naturally coming with weights according to the level of approximation they provide, we are able to embed all $d$--dimensional Grassmannians into a stratified space of covariance matrices. We observe that Grassmannians constitute the lowest dimensional skeleton of the stratification while it is possible to define a Riemaniann metric on the highest dimensional and dense stratum, such a metric being compatible with the global stratification. With such a Riemaniann metric at hand, it is possible to look for geodesics between two linear subspaces of different dimensions that do not go through higher dimensional linear subspaces as would euclidean geodesics. Building upon the proposed embedding of Grassmannians into the stratified space of covariance matrices, we generalize the concept of varifolds to what we call flagfolds in order to model multi-dimensional shapes.

We propose three variants of a nonlocal macroscopic pedestrian flow model accounting for anisotropic interactions between different groups and including the presence of walls or other obstacles in the walking domain. We compare the solutions behaviour in the case of bi-directional flows and othogonally crossing flows by numerical simulation based on high order finite difference schemes.

International audience; We introduce a reliable method to generate offset meshes from input triangle meshes or triangle soups. Our method proceeds in two steps. The first step performs a Dual Contouring method on the offset surface, operating on an adaptive octree that is refined in areas where the offset topology is complex. Our approach substantially reduces memory consumption and runtime compared to isosurfacing methods operating on uniform grids. The second step improves the output Dual Contouring mesh with an offset-aware remeshing algorithm to reduce the normal deviation between the mesh facets and the exact offset. This remeshing process reconstructs concave sharp features and approximates smooth shapes in convex areas up to a user-defined precision. We show the effectiveness and versatility of our method by applying it to a wide range of input meshes. We also benchmark our method on the Thingi10k dataset: watertight and topologically 2-manifold offset meshes are obtained for 100% of the cases.

In this paper we demonstrate how sub-Riemannian geometry can be used for manifold learning and surface reconstruction by combining local linear approximations of a point cloud to obtain lower dimensional bundles. Local approximations obtained by local PCAs are collected into a rank $k$ tangent subbundle on $\mathbb{R}^d$, $k

We propose a new class of models for variable clustering called Asymptotic Independent block (AI-block) models, which defines population-level clusters based on the independence of the maxima of a multivariate stationary mixing random process among clusters. This class of models is identifiable, meaning that there exists a maximal element with a partial order between partitions, allowing for statistical inference. We also present an algorithm for recovering the clusters of variables without specifying the number of clusters \emph{a priori}. Our work provides some theoritical insights into the consistency of our algorithm, demonstrating that under certain conditions it can effectively identify clusters in the data with a computational complexity that is polynomial in the dimension. This implies that groups can be learned nonparametrically in which block maxima of a dependent process are only sub-asymptotic., Comment: 50 pages, 6 figures

We extend the finite volume numerical scheme proposed by Hilliges and Weidlich [11] to second order traffic flow models consisting in 2 × 2 systems of non strictly hyperbolic conservation laws of Temple class. The scheme is shown to satisfy some maximum principle properties on the density. We provide numerical tests illustrating the behaviour at vacuum and the gain in computational time when dealing with optimization algorithms.

International audience; We introduce a deep latent recommender system named deepLTRS in order to provide users with high quality recommendations based on observed user ratings \textit{and} texts of product reviews. The underlying motivation is that, when a user scores only a few products, the texts used in the reviews represent a significant source of information, thereby enhancing the predictive ability of the model. Our approach adopts a variational auto-encoder (VAE) architecture as a deep generative latent model for an ordinal matrix encoding ratings and a document-term matrix encoding the reviews. Taking into account both matrices as model inputs, deepLTRS uses a neural network to capture the relationship between latent factors and latent topics. Moreover, a user-majoring encoder and a product-majoring encoder are constructed to jointly capture user and product preferences. Due to the specificity of the model structure, an original row-column alternated mini-batch optimization algorithm is proposed to deal with user-product dependencies and computational burden. Numerical experiments on simulated and real-world data sets demonstrate that deepLTRS outperforms the state-of-the-art, in particular in context of extreme data sparsity.

Every brain is unique, having its structural and functional organisation shaped by both genetic and environmental factors over the course of its development. Brain image studies tend to produce results by averaging across a group of subjects, under a common assumption that it is possible to subdivide the cortex into homogeneous areas while maintaining a correspondence across subjects. This paper questions such assumption: can the structural and functional properties of a specific region of an atlas be assumed to be the same across subjects? This question is addressed by looking at the network representation of the brain, with nodes corresponding to brain regions and edges to their structural relationships. We perform graph matching on a set of control patients and on parcellations of different granularity to understand which is the connectivity misalignment between regions. The graph matching is unsupervised andreveals interesting insight on local misalignment of brain regions across subjects.

This study simultaneously manipulates within-category (rule-based vs. similarity-based), between-category (blocked vs. interleaved), and across-blocks (constant vs. variable) orders to investigate how different types of presentation order interact with one another. With regard to within-category orders, stimuli were presented either in a “rule plus exceptions” fashion (in the rule-based order) or by maximizing the similarity between contiguous examples (in the similarity-based order). As for the between-category manipulation, categories were either blocked (in the blocked order) or alternated (in the interleaved order). Finally, the sequence of stimuli was either repeated (in the constant order) or varied (in the variable order) across blocks. This research offers a novel approach through both an individual and concurrent analysis of the studied factors, with the investigation of across-blocks manipulations being unprecedented. We found a significant interaction between within-category and across-blocks orders, as well as between between-category and across-blocks orders. In particular, the combination similarity-based + variable orders was the most detrimental, whereas the combination blocked + constant was the most beneficial. We also found a main effect of across-blocks manipulation, with faster learning in the constant order as compared to the variable one. With regard to the classification of novel stimuli, learners in the rule-based and interleaved orders showed generalization patterns that were more consistent with a specific rule-based strategy, as compared to learners in the similarity-based and blocked orders, respectively. This study shows that different types of order can interact in a subtle fashion and thus should not be considered in isolation.

International audience; Argumentation is used by people both internally, by evaluating arguments and counterarguments to make sense of a situation and take a decision, and externally, e.g., in a debate, by exchanging arguments to reach an agreement or to promote an individual position. In this context, the assessment of the quality of the arguments is of extreme importance, as it strongly influences the evaluation of the overall argumentation, impacting on the decision making process. The automatic assessment of the quality of natural language arguments is recently attracting interest in the Argument Mining field. However, the issue of automatically assessing the quality of an argumentation largely remains a challenging unsolved task. Our contribution is twofold: first, we present a novel resource of 402 student persuasive essays, where three main quality dimensions (i.e., cogency, rhetoric, and reasonableness) have been annotated, leading to 1908 arguments tagged with quality facets; second, we address this novel task of argumentation quality assessment proposing a novel neural architecture based on graph embeddings, that combines both the textual features of the natural language arguments and the overall argument graph, i.e., considering also the support and attack relations holding among the arguments. Results on the persuasive essays dataset outperform state-of-the-art and standard baselines' performance.

Background Presently, there is a wide variety of classification methods and deep neural network approaches in bioinformatics. Deep neural networks have proven their effectiveness for classification tasks, and have outperformed classical methods, but they suffer from a lack of interpretability. Therefore, these innovative methods are not appropriate for decision support systems in healthcare. Indeed, to allow clinicians to make informed and well thought out decisions, the algorithm should provide the main pieces of information used to compute the predicted diagnosis and/or prognosis, as well as a confidence score for this prediction. Methods Herein, we used a new supervised autoencoder (SAE) approach for classification of clinical metabolomic data. This new method has the advantage of providing a confidence score for each prediction thanks to a softmax classifier and a meaningful latent space visualization and to include a new efficient feature selection method, with a structured constraint, which allows for biologically interpretable results. Results Experimental results on three metabolomics datasets of clinical samples illustrate the effectiveness of our SAE and its confidence score. The supervised autoencoder provides an accurate localization of the patients in the latent space, and an efficient confidence score. Experiments show that the SAE outperforms classical methods (PLS-DA, Random Forests, SVM, and neural networks (NN)). Furthermore, the metabolites selected by the SAE were found to be biologically relevant. Conclusion In this paper, we describe a new efficient SAE method to support diagnostic or prognostic evaluation based on metabolomics analyses.