AI methods powered by Deep Neural Networks (DNN) are so successful that some studies even claim they have surpassed human performance. Their complex black-box nature, lack of generalization out-of-distribution (OOD) and mistaken correlation for causation, which can have catastrophic consequences for data-driven decision-making limit their trustability and social acceptance.This has given rise to the quest for AI model robustness, interpretability, and explainability. Explainable AI (XAI) is a hot topic in the field of ML, which has even become a political and legal concern. Indeed, with the increasing use of AI in systems governing our society, important decisions based on the DNNs' predictions can be made with a great impact on human life. To ensure the AI systems stay in phasis with social interests, it is crucial to becoming computed accurate predictions trustable by empowering the AI model used by the decision maker with robustness, transparency, and explainability. XAI has a panel of goals; trustworthiness is the main and, one of which is causality. Answering causal questions from data is an open avenue to XAI. Causality offers an unexplored complement to ML, allowing it to go beyond model associations. Learning a causal model provides the mechanisms giving rise to the observed statistical dependencies, and allows both to model distribution shifts through interventions, and to raise the veil on the model process leading to a prediction. Thus, the OOD generalization is not limited to predictive performance or robustness but also the explainability and reliability of the model process across switching distributions. All these aspects make Causality for XAI and generalization a key concern. This project aims to develop an approach that benefits from recent advances in causality for tackling modern ML problems such as XAI and generalization for trustable informing decision-making process.

This project aims at setting up robust asynchronous domain decomposition methods for mechanical problems. Indeed not to depend on synchronism is a key element in order to warranty that computational methods perform well on highly solicited heterogeneous clusters involving thousands of cores; it is also a key ingredient in order to partially rid yourselves of the load balancing issue, which is practically impossible to address in a nonlinear framework with adaptive grids. The newly proposed methods will be applied to nonlinear quasi-static structure problems (such as generalized elasto-viscoplasticity) and air/ocean computations. The project also addresses the questions of acceleration, space and time-parallelism.

Le cancer est la principale cause de décès dans le monde. Ces derniers temps, les avancées dans le domaine de la médecine de précision ont modifié les normes de traitement du cancer en proposant de nouveaux plans de traitement personnalisés aux résultats prometteurs. Cette spécification du traitement aux besoins du patient s’appuie principalement sur l’évaluation de lames d’histopathologie, sur lesquelles des échantillons de tissus sont colorés chimiquement afin de fournir une lecture distincte de différents profils moléculaires tels que KRAS, HER2, Ki67 et autres. Des preuves supplémentaires soulignent l’importance de la distribution spatiale de ces phénotypes moléculaires pour les décisions de traitement. Dans le cadre de ce projet, nous avons pour objectif de concevoir et d’évaluer de nouveaux modèles d’intelligence artificielle permettant de quantifier automatiquement les informations spatiales associées au traitement, à partir d’images numériques de lames scannées. Plus précisément, nous investiguerons de nouveaux modèles à base de transformeurs visuels et de réseaux génératifs reposant sur de l’apprentissage profond pour en déduire des tests moléculaires à partir de lames H&E, de manière supervisée et non supervisée. En outre, nous étudierons comment cette organisation spatiale des profils moléculaires peut être corrélée avec les résultats du traitement à travers une modélisation par graphe dans un paradigme d’apprentissage profond. Nous avons la profonde conviction que les résultats de ce projet auront un effet positif sur l’oncologie clinique, et donc sur l’espérance de vie des patients.

Developing a multiscale model for plants, capable of managing complex plant response to environmental conditions and its underlying genetic diversity, is a major issue in agronomy and biology. The Resource Balance Analysis framework is promising to integrate the finest scales (genes) to the individual scale, by refining the organ description in existing functional-structural plant models and by taking advantage of new omics data, such as quantitative proteomics. Dealing with multicellular organisms such as the plant opens mathematical and computational challenges for building, calibrating and simulating such multiscale models in steady-state and dynamical conditions. This project aims to first address these mathematical challenges, develop and experimentally validate the first multiscale model of the plant Arabidopsis thaliana. Second, the model will help biologists study the plant's resource allocation strategy under normal and limiting environmental conditions, primarily using data sets generated by one of the most robust plant phenotyping systems. The model will be used to generate new biological knowledge or hypotheses (e.g., cellular functions affected in plant plasticity to nutritional stress). This clearly represents a significant advance in the context of plant adaptation, and to decipher the specific responses of other A. thaliana genotypes to complex environmental conditions. The modeling framework will address important challenges such as the integration of heterogeneous and multiscale data (from omics to phenotypic traits), genotype-phenotype relationships in complex environments, and the integration of genetic diversity into modeling.

Agriculture has currently to tackle new challenges, largely due to the need to increase global food supply under the declining availability of soil and water resources and increasing threats from climate change. It has to face main changes and to adapt to new conditions, in particular environmental ones. To better handle this adaptation, it is necessary to better understand several key notions such as genetic variability and interactions between the plant and its environment. In this context, predictive approaches relying on ecophysiology and genetic knowledge, as well as mathematical modeling are very promising. The Stat4Plant project aims at developing new statistical methodologies and new algorithmic tools for modeling and analyzing genotypic variability and interaction between plant and its environment in a context of climate change. The project consortium gathers scientists in modeling and applied statistics with large experience in interdisciplinary collaborations in plant sciences and biologists with strong expertise in phenotype-genotype relations. The project is structured in four main research axes, supported by strong collaborations between statisticians and biologists and motivated by practical questions linked with biological dataset. The first axis aims at developing new methods for identifying key biological processes driving plant development lying behind the observed genotypic variability. These works combine mechanistic ecophysiologic modeling highly nonlinear of plant development, statistical mixed effects modeling for genotypic variability and statistical testing procedures, in particular adapted to small data samples, to identify genotype-dependent parameters. These approaches will allow to better understand genotype by environment interactions and to identify new tools for varietal selection. The second axis is dedicated to joint modeling of a time of interest such as flowering time or harvest time and of a phenotypic dynamical trait depending on time such as biomass or pest presence. The considered joint models combine survival models with random effects and covariates of high dimension and nonlinear mixed effects models for the dynamical trait. The objective is to identify the relevant covariates, to estimate the parameters and to predict the time of interest. These methods will allow for example to better predict flowering time or optimal harvest time. The third axis aims at developing new methods for identifying among a high number of covariates those who are the most influent for a phenotypic trait of interest, solely or jointly with a time of interest. Nonlinear mixed effects models combining mechanistic models of plant development and genetic models integrating a high number of genetic covariates will be used to model genotypic variability of the trait of interest. New covariates selection methods adapted to the nonlinear context will be developed. These methods will allow to identify the main genetic factors influencing the trait. Finally, the last axis aims at building new criteria for varietal selection, integrating randomness of environmental conditions and targeting simultaneously several objectives, such as maximizing yield and minimizing yield variability. New methodologies for optimizing these criteria will be developed. Such criteria will be new tools for decision support system in agriculture.