Journée de l’axe DAC sur la causalité et la quantification d’incertitude

L’axe « Données, Apprentissage, Connaissances » de la fédération Normastic organise le jeudi 26 juin 2025 une journée sur la causalité et la quantification d’incertitudes. Elle est notamment ouverte à tous les membres du GREYC et du LITIS (collègues, doctorants, post-doctorants, master…).

Cette journée aura lieu à l’université de Caen Normandie sur le site du campus 2, UFR des Sciences, bâtiment Sciences 3, salle S3-351.

Programme :

9h15 : accueil café

9h30: Alessandro Leite : introduction à la journée (slides)

9h45 : tutoriel de Vincent Blot (Capgemini, CNRS, LISN, Université Paris-Saclay) et Valentin Laurent  (Quantmetry – Part of Capgemini Invent, Paris) : MAPIE – Model Agnostic Prediction Interval Estimator Lien notebook.

11h : pause

11h15 : tutoriel de Nicolas Voisine (Orange Labs) : Khiops Uplift – A Library for Causal Inference

Abstract: Khiops is an open source machine learning tool designed for mining large multi-table databases. Khiops is based on a unique Bayesian approach that has attracted academic interest with more than 20 publications on topics such as variable selection, classification, decision trees, co-clustering and upLift Modeling. It provides a predictive measure of variable importance using discretisation models for numerical data and value clustering for categorical data. The proposed classification/regression model is a naive Bayesian classifier incorporating variable selection and weight learning. In the case of multi-table databases, it provides propositionalisation by automatically constructing aggregates. Khiops is adapted to the analysis of large databases with millions of individuals, tens of thousands of variables and hundreds of millions of records in secondary tables. It is available on many environments, both from a Python library and via a user interface.

12h30: déjeuner

13h45 : Shuyu Dong (L2S, CentraleSupélec, CNRS, Université Paris-Saclay) : « Large-scale causal structure learning: challenges and new methods« .

Abstract: Learning causal structures from observational data is a fundamental yet challenging problem. In this talk I will first present a few different causal structure learning methods running on a single machine. Then I will present our distributed approach named DCILP, which is a new divide-and-conquer-based method. The divide phase proceeds by identifying the Markov blanket MB(X_i) of each variable, and solving the subproblems restricted to each MB in parallel. The novelty of DCILP lies in the conquer phase, which tackles the problem of aggregating the local causal graphs from the divide phase. We show that this aggregation task can be formulated as an integer linear programming (ILP) problem which is delegated to an ILP solver, and that the resulting algorithm demonstrates significant improvements in scalability with satisfactory learning accuracy.

14h15 : Louis Hernandez (LITIS, INSA Rouen, Craft AI) : « Causal Inference and Large Language Models: opportunities, methods, and challenges« .

Abstract: Large language models (LLMs) are increasingly used in scientific discovery, decision support, and knowledge extraction. At the same time, causal inference provides principled tools for reasoning about the results of actions and counterfactual questions. This talk will explore the emerging intersection between causal inference and large language models. Through recent examples and ongoing research, it will highlight both the promises and limitations of combining LLMs with causal inference and outline open questions for future work in this rapidly evolving field.

14h45 : Giorgio Morales (GREYC, Université de Caen Normandie) : « Managing Uncertainty in Regression Neural Networks: From Prediction Intervals to Adaptive Sampling« .

Abstract: Understanding and managing uncertainty is a critical aspect of deploying regression neural network models in real-world scientific and engineering applications. This presentation introduces two novel contributions aimed at improving uncertainty quantification and guiding data acquisition under uncertainty. The first is DualAQD, a dual-network architecture for generating high-quality prediction intervals (PIs). DualAQD integrates a custom loss function that minimizes interval width while ensuring coverage constraints, striking a balance between tightness and reliability of uncertainty estimates. It consistently outperforms existing PI-generation techniques in both interval efficiency and prediction accuracy across diverse datasets.
Building on DualAQD’s uncertainty modeling, we present ASPINN, an adaptive sampling strategy designed for data-scarce environments where measurement collection is costly or constrained. ASPINN addresses this by focusing on epistemic uncertainty reduction in regression problems, using NN-generated PIs to guide adaptive data acquisition to strategically select new data points that most reduce model uncertainty. By incorporating a Gaussian Process surrogate to support batch sampling, ASPINN balances informativeness and diversity in acquisition decisions. Empirical evaluations show that ASPINN achieves faster convergence and greater uncertainty reduction compared to leading alternatives. Together, these methods offer a robust framework for uncertainty-aware learning in regression tasks.

15h15 : Su Ruan (AIMS, Université de Rouen Normandie) :  « Uncertainty Quantification in Deep Evidential Fusion for Medical Imaging Applications« .

Abstract : In medical imaging, integrating information from multiple sources, such as different imaging modalities, is essential for achieving accurate diagnoses and robust clinical decisions. This presentation addresses the challenge of uncertainty quantification in deep learning-based fusion methods. We introduce a framework called Deep Evidential Fusion, which typically operates at the late fusion stage. It produces not only improved fused predictions but also reliable uncertainty estimates, without relying on computationally intensive techniques like Monte Carlo dropout. We will demonstrate the effectiveness of this approach across various medical imaging tasks, showing enhanced predictive performance and better-calibrated uncertainty compared to conventional fusion strategies.

15h45 : pause

16h : Echanges sur la causalité et la quantification d’incertitude.

Fin vers 16h30

Inscription (nécessaire pour l’organisation) : https://framaforms.org/registration-1748241556

Comme pour toute journée d’animation de la fédération, les frais de déplacement, repas et pauses café des membres du GREYC et du LITIS sont pris en charge par la fédération.