Alexandra Lefebvre (Sorbonne Université) 

11h - salle 38

TBA

Résumé: TBA

Alexandre Hippert-Ferrer (Centrale-Supelec) 

11h - salle Actia

Robust low-rank covariance matrix estimation with missing values and application to classification problems.

Résumé:  Missing values are inherent to real-world data sets. Statistical learning problems often require the estimation of parameters as the mean or the covariance matrix (CM). If the data is incomplete, new estimation methodologies need to be designed depending on the data distribution and the missingness pattern (i.e. the pattern describing which values are missing with respect to the observed data). This talk considers robust CM estimation when the data is incomplete. In this perspective, classical statistical estimation methodologies are usually built upon the Gaussian assumption, whereas existing robust estimation ones assume unstructured signal models. The former can be inaccurate in real-world data sets in which heterogeneity causes heavy-tail distributions, while the latter does not profit from the usual low-rank structure of the signal. Taking advantage of both worlds, a CM estimation procedure is designed on a robust (compound Gaussian) low-rank model by leveraging the observed-data likelihood function within an expectation-maximization (EM) algorithm. After a validation on simulated data sets with various missingness patterns, the interest the proposed procedure is shown for CM-based classification and clustering problems with incomplete data. Investigated examples generally show higher classification accuracies with a classifier based on robust estimation compared to the one based on Gaussian assumption and the one based on imputed data.

Jérome Stenger 

11h - salle ACTIA

 Quantification robuste de l'incertitude d'une mesure de risque issue d'un code de calcul

Résumé: La quantification d'incertitudes a pour but d'évaluer l'impact d'un manque de connaissance des paramètres d'entrées (considérés aléatoires) sur les résultats d'une expérience numérique. Dans ce travail, nous prenons en compte un second niveau d'incertitude qui affecte le choix du modèle probabiliste des paramètres d'entrées. Nous évaluons les bornes d'une quantité d'intérêt sur l'ensemble des mesures de probabilités uniquement définies par leur bornes et certains de leurs moments. Du fait du grand nombre de contraintes, l'optimisation numérique est complexe. Nous montrons que le problème d'optimisation peut se paramétriser sur les points extrémaux de cet espace de mesures de probabilité contraintes. De plus, nous proposons une nouvelle paramétrisation libre de contraintes basées sur les moments canoniques.

Sandrine Boulet (Sorbonne Université) 

11h - salle 31

Développement de méthodes Bayésiennes de sélection de variables pour intégrer l’information experte

Résumé:  La construction d’outils d’aide à la décision à destination des cliniciens nécessite de sélectionner, au préalable, les variables pertinentes pour modéliser la décision clinique. Pour cela, deux sources d’information peuvent être utilisées : les données observées recueillies dans les dossiers informatisés des patients, et l’expertise des médecins. Peu de méthodes permettent de combiner ces deux types d’information. L’originalité de mon travail est donc de proposer des stratégies pour prendre en compte ces deux sources de données. Notre application porte sur la modélisation de la décision médicale de modification des doses d’irinotécan en fonction des caractéristiques des patients et des effets indésirables dans le traitement du cancer colorectal métastatique.

Les méthodes développées reposent sur des méthodes de sélection de variables bayésiennes dans lesquelles l’expertise est intégrée sous forme de poids associés à chaque variable et élicités par les experts. Le premier modèle est une adaptation d’une méthode de sélection de variables bayésienne, la méthode de Stochastic Search Variable Selection (SSVS), pour laquelle l’information experte est introduite dans la distribution a priori de l’indicatrice de sélection de variable. Le modèle a été appelé Weights-Based SSVS (WBS). Le deuxième modèle est fondé sur la méthode du power prior (PP) et permet de combiner des données simulées à partir des paramètres experts et les observations. La sélection de variables est elle aussi assurée par un modèle SSVS. Nous avons réalisé des simulations afin d’analyser les performances du premier modèle par rapport à celles des modèles LASSO et SSVS. Pour le cas d’usage, les données de soin des patients ayant reçu un protocole incluant de l’irinotécan à une dose théorique de 180 mg/m2 sur un cycle d’une durée théorique de 14 jours ont été extraites des dossiers patients informatisés de l’Hôpital Européen Georges Pompidou à Paris. Pour prendre en compte l’expertise clinique, nous avons construit un questionnaire en collaboration avec les oncologues dans lequel ils quantifient l’importance de chaque caractéristique des patients et de chaque grade de chaque type de toxicités sur la réduction de dose par l’intermédiaire de poids élicités prenant leurs valeurs entre 0 et 100. Enfin, nous avons appliqué ces modèles à des données de patients traités par irinotécan pour un cancer colorectal.

Kamelia Daudel (Université of Oxford)

11h - salle 31

Infinite-dimensional alpha-divergence minimisation for Variational Inference

Résumé: Variational Inference methods have made it possible to build fast algorithms for posterior distribution approximation. Yet, the theoretical results and empirical performances of Variational Inference methods are often impacted by two factors : one, an inappropriate choice of the objective function appearing in the optimisation problem and two, a search space that is too restrictive to match the target at the end of the optimisation procedure. In this talk, we explore how we can remedy these two issues in order to develop improved Variational Inference methods. More precisely, we suggest selecting the Alpha-divergence as a more general class of objective functions and we propose to enlarge the search space by adding a prior on the variational parameters in the form of a probability measure.

Luke Kelly (Université Paris Dauphine)

11h - salle ACTIA

Lagged couplings for Markov chain Monte Carlo phylogenetic inference

Résumé: Phylogenetic inference is an intractable statistical problem on a complex sample space. Markov chain Monte Carlo methods are the primary tool for Bayesian phylogenetic inference, but it is challenging to construct efficient schemes to explore the associated posterior distribution or assess their performance. Existing approaches are unable to diagnose mixing or convergence of Markov schemes jointly across all components of a phylogenetic model. Building on recent work developing couplings of Markov chain Monte Carlo algorithms to diagnose convergence and construct unbiased estimators, we describe a procedure to couple Markov chains targeting a posterior distribution over a space of phylogenetic trees with branch lengths, scalar parameters and latent variables. We use these couplings to check mixing and convergence of Markov chains jointly across all components of the phylogenetic model; samples from our coupled chains may also be used to construct unbiased estimators.

Grégoire Clarté (University of Helsinki)

11h - salle ACTIA

Reconstruction phylogénétique depuis données matricielles : application à la linguistique historique

Résumé: Nous proposons de reconstruire, par inférence bayésienne, une phylogénie, en se basant sur des données matricielles. Nous appliquons cette méthode à l'histoire des langues des signes. Nous développons un modèle dans lequel lignes et colonnes sont corrélées ; ces données peuvent représenter des traits socio-culturels, phénotypiques, ou, comme dans notre cas, des données phonologiques. Nous montrons comment calculer la vraisemblance de ce modèle et proposons des méthodes numériques pour échantillonner depuis le posterior associé, basées sur les méthodes SMC associées à un tempering exotique. Les résultats sur données simulées sont satisfaisants, tandis que les résultats sur données réelles apportent des éléments de réponses aux questions des linguistes.

Antonello Lobianco (Agroparistech, BETA, Nancy)

11h - salle ACTIA

Using the Julia programming language to develop machine learning workflows and algorithms. The example of the Beta Machine Learning Toolkit (BetaML.jl)

Résumé: Julia is a relatively new high-level general purpose language with a focus on performances and data analysis. The seminar presents its key characteristics that make it particularly suitable in the scientific domain and in particular in the implementation of both custom machine learning algorithms and related workflows, when existing libraries are used. We will employ BetaML, a machine learning library developed at the "Bureau d'Économie Théorique et Appliquée" UMR, as an example of how simplicity (in the usage and in the code) doesn't need to trade off with performances.

Hugo Gangloff (Université Bretagne Sud)

11h - salle ACTIA

Deep pairwise and triplet Markov chains for unsupervised signal processing

Résumé: Probabilistic graphical models such as hidden Markov models have found many applications in unsupervised signal processing, such as part-of-speech tagging, image segmentation, genetic sequence analysis, etc. In this presentation, we focus on the pairwise and triplet Markov chain models which define very general frameworks that have been very little explored so far. While pairwise Markov chains strictly extend the direct dependencies that can be introduced by the model, triplet Markov chains additionally enable the introduction of much more complex probability distributions. However, such generalizations raise the questions of the choice of the probability distributions, their parametrization and the unsupervised parameter estimation in the complex models that can be built. We will explore these questions and propose answers: i) we use an auxiliary latent process to implicitly define complex probability distributions, ii) the parametrization issue is considered by embedding deep neural networks in the new models and iii) a general algorithm, based on a lower bound of the loglikelihood, is derived in order to perform unsupervised parameter estimation in these sequential models,. We show that the new models outperform the hidden Markov chains and their classical extensions usually considered in the literature.

Ricardo Carrizo (Paris II Assas)

11h - salle ACTIA

Development of a novel GWAS method for the detection of causal genes with population specific allelic effects

Kosuke Hamazaki (University of Tokyo)

11h - salle ACTIA

Development of a novel GWAS method for the detection of causal genes with population specific allelic effects

Résumé: GWAS (Genome-Wide Association Study) aims at detecting candidate genes (QTL) associated with a target trait via statistical testing. It is now widely used not only in humans but also in plant animal genetics and breeding. A classical GWAS starts with the constitution of a panel of individuals, usually gathered from different populations. This population structure needs to be accounted for in the analysis in order to efficiently control the Type I error rate of the QTL detection procedure. Many methods have been proposed to control the false positive error rate in large datasets with strong population structure; e.g., a mixed-effects model to correct the effects of family relatedness by regarding polygenetic effects as random effects. However, most methods assume the same QTL effect across populations, which is not always true in the real biological process. Rio et al. (PLOS Genetics, 2020) proposed a method to consider population-specific QTL effects by testing marker effects in each population separately with prior information on population membership for each individual. This information on the population structure, however, may not always be available, in which case the methodology of Rio et al. cannot be applied. We propose a novel method that does not require prior knowledge of the population structure. In the proposed model, we explicitly include an interaction term between a SNP of interest and the genetic background into the conventional GWAS model. The effect of the interaction term is also tested with the SNP effect of interest simultaneously. The proposed model, the SNPxGB model, can be justified because population-specific QTL effects can be regarded as the interaction between the QTLs and the genetic background. In this seminar, I will first illustrate the classical GWAS method with the mixed-effects model, including the influences of population structure on GWAS results. Then, I will introduce the model of Rio et al. with prior information on the population structure. Finally, I will explain the proposed model including the interaction term between SNPs and the genetic background. I will also introduce a model that combines the proposed model with a haplotype-based approach, the HBxGB model, in order to better control the Type I error rate compared to the proposed SNPxGB model. The different procedures will be compared on simulated data.

Edi Prifti (IRD)

11h - salle ACTIA

Vers une médecine de précision basée sur l’intelligence artificielle et l’intégration de données massives et hétérogènes

Résumé: Les données biomédicales sont complexes et en croissance constante. Leurs traitement, intégration et modélisation constituent des défis importants. L’objectif de mon projet de recherche est de contribuer à la recherche méthodologique, translationnelle et applicative, notamment en lien avec les pays du Sud. Ce projet de recherche est structuré en trois axes : 1) Algorithmes de passage à l'échelle pour la modélisation des microbiomes, 2) Découvertes de biomarqueurs et développement de tests diagnostiques, 3) Démocratisation de l'Intelligence Artificielle (IA) et des dispositifs médicaux connexes. Le premier axe consiste à contribuer à l’ouverture des trois verrous suivants. Verrou 1 : La détermination et la gestion de catalogues métagenomique précis de constitution de génomes. Il s’agit d’un problème complexe à résoudre par sa complexité algorithmique, mais aussi computationnelle. Verrou 2 : La compréhension des relations entre les différents composants du microbiome et construction d’abstractions qui les représentent le mieux. Verrou 3 : Étude de la dynamique et équilibre de ces relations en explorant l’approche par simulation multiagent. Le deuxième axe met le focus sur le développement d’approches méthodologiques interprétables pour identifier des signatures prédictives du microbiome. Mais aussi des aspects plus terre-à-terre visant à déterminer des protocoles permettant de passer des signatures à des modelés concrets utilisables en clinique. Le troisième axe du projet se situe au cœur des applications d’IA qui intègrent des données issues de capteurs médicaux connexes avec des données cliniques ou environnementales. En collaboration avec des collègues cliniciens. Dans l’ensemble, à travers collaborations existantes et futures, je souhaite proposer des solutions à des problèmes qui sont en lien avec la demande sociétale dans la recherche-action au Sud.

Achille Thin (CMAP Ecole Polytechnique)

11h - salle ACTIA

Monte Carlo Variational Auto Encoders

Résumé : Variational auto-encoders (VAE) are popular deep latent variable models which are trained by maximizing an Evidence Lower Bound (ELBO). To obtain tighter ELBO and hence better variational approximations, it has been proposed to use importance sampling to get a lower variance estimate of the evidence. However, importance sampling is known to perform poorly in high dimensions. While it has been suggested many times in the literature to use more sophisticated algorithms such as Annealed Importance Sampling (AIS) and its Sequential Importance Sampling (SIS) extensions, the potential benefits brought by these advanced techniques have never been realized for VAE: the AIS estimate cannot be easily differentiated, while SIS requires the specification of carefully chosen backward Markov kernels. In this work, we address both issues and demonstrate the performance of the resulting Monte Carlo VAEs on a variety of applications.

Marine Demangeot (LPSM, Sorbonne Université)

11h - salle ACTIA

Estimation of the extremal coefficient function based on a single spatial observation

Résumé : The extremal coefficient function is a bivariate measure of spatial dependence for stationary max-stable processes. It is usually estimated from time series, when the spatial object under study is observed through time (e.g. extreme precipitations, extreme temperatures, high concentrations of pollution in the air). However, in some cases, such types of data cannot be accessed: only one or just a few records are made available. This is the case, for instance, in mining resources estimation, soil contamination evaluation or any other applications where the phenomenon of interest either varies too slowly across time to hope for a decent time series, or is too expensive to sample from. This situation is rarely addressed in the spatial extremes community, contrary to geostatistics, which typically deals with such issues. A basic geostatistical tool is the so-called variogram, which is also a bivariate measure of spatial dependence. Considering the indicator variogram, above some threshold, of a stationary max-stable random field, we propose a new nonparametric estimator of the extremal coefficient function based on the variogram’s Nadaraya-Watson estimator. The latter has been studied by Garcia-Soidan et al. (2004) and Garcia-Soidan (2019); from their work, we derive asymptotic properties of our estimator when it is computed from a single spatial set of observations. Namely, under some assumptions, we show that it is consistent and asymptotically normal. These results are illustrated by numerical experiments and a comparison with the well-known F-madogram based estimator is performed. An application on a real dataset is also presented.

Fanny Mollandin (INRAE, GABI)

11h - salle ACTIA

Accounting for complex overlapping annotations as biological priors in genomic prediction models of complex traits

Résumé : The primary objective of genomic prediction is to use genomic variation, usually single nucleotide polymorphisms (SNPs), to predict complex phenotypes. In particular, genomic prediction models are widely used as an evaluation tool for genomic selection in plant and animal breeding. However, the prediction accuracies of many complex quantitative traits still have room for further improvement, due to factors such as marker density, the underlying genetic architecture, and population structure. Alongside this, there is an increasing accumulation of knowledge about the genome, including improved functional annotation and more widely available high-throughput molecular assays (i.e. omics data), providing a bridge from genome variation to phenotypes. Integrating this information into genomic prediction models could potentially lead to improved prediction accuracy and a better understanding of the underlying architecture of complex traits. Bayesian models provide a straightforward way to introduce known functional information into genomic prediction models through the use of prior distributions. In particular, BayesRC divides SNPs into disjoint annotation categories, allowing the proportion of QTLs to vary in each. Although BayesRC has shown promising results, it is limited by the non-overlapping nature of the annotations, which prevents SNPs from belonging to more than one functional list. As the number of potential annotation categories increases, this constraint will become a key limitation. To address this issue, we present two novel extensions of BayesRC to handle potentially overlapping annotations through either a cumulative or stochastic approach. Our approaches allow SNPs with multiple annotations to be respectively upweighted or preferentially assigned to the annotation that best characterizes them. We compare and evaluate these two proposed models with state-of-the-art Bayesian genomic prediction models on simulated and real data, with a simultaneous focus on prediction quality and QTL mapping accuracy.

Gilles Blanchard (LMO, Université Paris Saclay)

11h - salle ACTIA

TBA

Résumé : TBA

Mary Savino (MIA Paris, AgroParisTech)

11h - salle ACTIA

Statistical learning methods for the simulation of highly nonlinear problems for geological porous media and materials

Résumé : Computation simulation models using mathematical concepts and language are necessary to demonstrate the feasability of a project and in decision making processes. At Andra, (Agence Nationale pour la gestion des déchets radioactifs) simulation models are needed to prove the quality and the security of the Cigéo Project, a deep geological disposal facility for highly radioactive long-lived waste, before undertaking any construction work. However, these computational simulations can demand prohibitively large computational budgets and hinder fast decision-making. In this presentation, we propose a sequential data-driven method for dealing with equilibrium-based chemical simulations to build what is commonly called a surrogate model in order to circumvent these time-consuming and costly simulations. Our method is based on the idea that the function to estimate is a sample of a Gaussian Process (GP), which allows us to compute the global uncertainty of the function estimation. It can be seen as a specific machine learning (ML) approach called active learning, since it sequentially chooses the most relevant input data at which the function to estimate has to be evaluated. Our active learning method is first validated through numerical experiments based on halite precipitation and then applied to a complex chemical system commonly used in geoscience. This method is then applied to Reactive Transport Modeling.

Erme Anakok (MIA Paris, AgroParisTech)

11h30 - salle ACTIA

Magnetic and Mechanic Design Of a 14 Tesla MRI using Genetic Algorithms

Résumé : Les algorithmes génétiques sont des procédures d’optimisation s’inspirant du vivant : mutations, croisements et sélections sont utilisés afin d’optimiser des fonctions dans des grands espaces d’états. La phase de sélection peut prendre en compte différentes contraintes, avec des formes variables, afin de résoudre des problèmes d'optimisation sous contraintes. Un exemple où l’on est confronté à des contraintes de formes variées est le design d’une bobine principale pour IRM à 14 T. Il s’agit d’un problème complexe faisant intervenir différentes types de physique comme le magnétisme ou la mécanique, et comportant aussi des obligations quant à l’usinabilité de la machine. En reformulant ce problème de design en un problème d’optimisation sous contraintes, les algorithmes génétiques ont permis d'ouvrir la voie à une nouvelle philosophie du design dans le domaine des aimants supraconducteurs : la possibilité d’inclure toutes les étapes du design dans un même processus d’optimisation.

Vacances de Noël

Vacances de Noël

Anne Sabourin (LTCI, Télécom Paris, IPP)

11h - salle ACTIA

Tail inverse regression for dimension reduction with extreme response

Félix Cheysson (LPSM, Sorbonne Université)

11h - salle ACTIA

Evolution of groups at risk of death from Covid-19 using hospital data

Anass Aghbalou (LTCI, Télécom Paris, IPP)

11h - salle ACTIA

Validation croisée pour les événements rares

Perrine Lacroix (LMO, IPS2)

11h - salle ACTIA

Compromis entre risque prédictif et false discovery rate pour la régression linéaire gaussienne en grande dimension

Clément Chadebec (Université de Paris, INRIA, INSERM)

11h - salle ACTIA

Data Augmentation in High Dimensional Low Sample Size Setting with Geometry-Aware Variational Autoencoders

Baptiste Kerleguer (CMAP, Ecole polytechinque)

11h - salle ACTIA

Jour de la Toussaint

Vacances de la Toussaint

Liliane Bel (MIA Paris, AgroParisTech)

11h - salle ACTIA

Variable selection for spatial models

Thanh Mai Pham Ngoc (LMO, Université Paris Saclay)

11h - salle ACTIA

Adaptive estimation of nonparametric geometric graphs

Mathis Chagneux (MIA Paris, AgroParisTech)

11h - salle ACTIA

Macrolitter video counting on river banks with state space models for moving cameras

State of The R

11h - salle ACTIA

Retour de la semaine à Roscoff

Wencan Zhu (AgroParisTech & Sanofi)

11h - salle ACTIA

A variable selection approach for highly correlated predictors in high-dimensional data

Tâm Le Minh (MIA Paris, AgroParisTech)

11h - salle ACTIA

Comparaison de réseaux d'interaction écologiques au moyen de modèles probabilistes échangeables