Title: Analyse de Formes pour la Compréhension du Comportement Humain

Abstract: L’analyse visuelle du mouvement humain, un domaine majeur de la vision par ordinateur, vise à détecter, suivre et identifier les individus, ainsi qu’à interpréter leurs comportements à partir de séquences d’images. Les formes du corps et du visage sont représentées dans un espace invariant à certaines transformations indésirables (échelle, translation, rotation, reparamétrisation).

La contribution principale est un cadre unifié basé sur des variétés multiples pour analyser différents types de données, avec des applications allant de la reconnaissance d’actions à l’estimation de biométries douces et à l’analyse des expressions faciales. Les landmarks (squelette et visage) sont modélisés dans l’espace de Kendall, et une méthode de codage parcimonieux intrinsèque avec apprentissage de dictionnaire y est proposée, comparée à une approche extrinsèque.

Ensuite, les courbes faciales sont modélisées sur une variété infinie, et des déformations géodésiques entre visages 3D sont exploitées pour la reconnaissance biométrique et des expressions. Enfin, un cadre pour les surfaces 3D paramétrées est introduit, avec une méthode innovante inspirée de la théorie de jauge pour calculer les géodésiques sans reparamétrisation.

Les résultats expérimentaux montrent l’efficacité de ce cadre pour la compréhension des comportements humains.

Title: Intelligent Approaches for Software Effort Estimation: Challenges and Advances 

Abstract: Software effort estimation remains one of the most critical and challenging tasks in software project management, as inaccurate predictions can lead to cost overruns, schedule delays, and resource misallocation. Recent advances in intelligent computing have opened new opportunities to improve estimation accuracy, robustness, and adaptability across heterogeneous project datasets. This presentation explores three emerging directions in intelligent software effort estimation. First, it highlights the role of attention-based autoencoders combined with ensemble learning to automatically extract meaningful latent representations from project data and enhance predictive performance through model diversity. Second, it addresses the importance of data harmonization for software cost estimation, emphasizing how preprocessing, feature alignment, and cross-dataset consistency can reduce noise and improve the transferability of estimation models. Third, it discusses meta-learning-based approaches, including Reptile-ACO, as promising solutions for rapidly adapting effort estimation models to new environments with limited training data while benefiting from prior experience across tasks. By combining representation learning, data quality improvement, and adaptive optimization, these intelligent approaches contribute to more accurate, scalable, and generalizable software effort estimation frameworks. The presentation concludes by outlining key challenges, recent advances, and future research opportunities in this evolving field. 

Title: Bayesian Geodesic Regression on Grassmann Manifolds Leveraging Lagrangian Hamiltonian Monte Carlo 

Abstract: Hamiltonian Monte Carlo (HMC) improves traditional Markov Chain Monte Carlo (MCMC) efficiency by mitigating random walk behavior. Lagrangian Hamiltonian Monte Carlo (LHMC) further enhances sampling by shifting the dynamics from Riemannian Hamiltonian to Lagrangian mechanics. This paper leverages LHMC to introduce a new Bayesian regression model for Grassmann manifolds, with parameters estimated via Grassmann LHMC (GLHMC). GLHMC effectively utilizes the manifold’s Riemannian geometry to achieve high-fidelity statistical inference. We show that solving the Lagrangian dynamics is equivalent to following the geodesic flow, which enables efficient computation. We analyze the asymptotic properties of the resulting posterior distribution, providing strong theoretical guarantees for the inference. 

Title: Optimization Methods for Smooth Curve Fitting on Riemannian Manifolds 

Abstract: Many data analysis problems involve observations that naturally lie on nonlinear spaces rather than on Euclidean ones. This gives rise to challenging optimization problems on Riemannian manifolds, both from theoretical and computational perspectives. We study the construction of smooth curves on a Riemannian manifold that interpolate or approximate data observed at prescribed time instants while preserving regularity properties along the trajectory. Our approach relies on geometric tools adapted to manifold-valued data, with particular emphasis on Bézier-type constructions and related recursive procedures. Building on these ingredients, we formulate suitable optimization models and derive practical algorithms for their solution. 

Title: Aitchison Geometry and Optimization Algorithms on the Probability Simplex

Abstract: Optimization over the probability simplex is a fundamental problem that arises in various fields, such as machine learning and portfolio optimization. Many methods have been developed for this problem, including projected gradient descent, mirror descent, and exponentiated gradient. In this paper, we propose an alternative approach based on Aitchison geometry, in which we use the centered log-ratio (clr) transformation to map the simplex onto a hyperplane. We introduce the Dynamic Aitchison Flow (DAF), a novel optimization framework that formulates gradient descent natively within the clr latent space of the Aitchison geometry. Experimental results demonstrate that the proposed method achieves faster convergence, as evidenced by a more rapid decrease of the objective function compared to existing approaches.

Title: Autour du Score de Brier en Analyse de Survie 

Abstract: Le score de Brier est une mesure très utilisée en épidémiologie pour évaluer les performances des modèles de survie. Dans cet exposé, nous expliquons comment cette mesure est définie, comme elle s'interprète, et comment elle a été adaptée au contexte, courant en analyse de survie, de la censure à droite. En particulier, nous revenons sur le principe de la méthode IPCW (pondération par l'inverse de la probabilité de censure). Dans une seconde partie de l'exposé, nous construisons un score de Brier adapté au cas où l'information relative à la censure est manquante pour certains individus de l'étude. Plusieurs méthodes sont envisagées : calibration par régression, imputation multiple, méthode doublement robuste. 

Title: Generative AI & LLM: Learning Perspectives 

Abstract: This presentation explores the transformative role of Generative Artificial Intelligence and Large Language Models (LLMs) in modern learning environments. It provides an overview of the fundamental concepts behind generative models, highlighting how systems such as transformer-based architectures enable machines to generate human-like text, images, and code. The session examines the impact of LLMs on practical applications, challenges, and limitations. Finally, the presentation offers perspectives on how learners, educators, and researchers can effectively leverage generative AI tools to enhance knowledge acquisition, creativity, and productivity in academic and professional contexts. 

Title: Hyperparameters Selection Problems in Nonparametric Trend Estimation 

Abstract: A central challenge in nonparametric estimation lies in the selection of tuning parameters that govern the trade-off between bias and variance. In contrast to parametric models, nonparametric approache (such as kernel smoothing, spline regression, and local polynomial estimation) depend critically on hyperparameters, including bandwidths, smoothing penalties, and the number of basis functions.

An inappropriate choice of these parameters can lead either to overfitting, where noise is mistaken for signal, or to oversmoothing, where important structural features of the underlying trend are obscured. Classical selection procedures, such as cross-validation, plug-in methods, and information criteria, provide practical solutions, yet they come with specific theoretical guarantees and inherent limitations.

This talk focuses on the hyperparameter selection problem in the context of stationary dependent observations, where standard methods may fail to adequately account for dependence structures.

This is joint work with K. Benhenni and D. A. Girard.

Title: Dependent Default Times with Mixed Hitting-time Models

Abstract: Mixed hitting-time models allow the analysis of competing risks through optimal stopping decisions interpreted as crossing times of latent Lévy processes with heterogeneous thresholds. In this paper, we consider a bivariate time model with dependent default, where observation times are subject to censoring and share a common latent process given by a Lévy subordinator. We establish the identifiability of the model and propose different estimators for the marginal distributions and the joint survival distribution. We establish their asymptotic properties and evaluate the finite-sample performance of our results through a simulation study on synthetic data, followed by an application using real data.

Title: Model Selection by Cross-validation in an Expectile Linear Regression

Abstract: For linear models that may have asymmetric errors, we study variable selection by cross-validation. The data are split into training and validation sets, with the number of observations in the validation set much larger than in the training set. For the model coefficients, the expectile or adaptive LASSO expectile estimators are calculated on the training set. These estimators will be used to calculate the cross-validation mean score (CVS) on the validation set. We show that the model that minimizes CVS is consistent in two cases: when the number of explanatory variables is fixed or when it depends on the number of observations. Monte Carlo simulations confirm the theoretical results and demonstrate the superiority of our estimation method compared to two others in the literature. The usefulness of the CV expectile model selection technique is illustrated by applying it to real data sets.