François Grolleau

This talk examines two critical aspects of data-driven medicine: personalized treatment strategies using causal inference and the robust evaluation of large language models (LLMs) for clinical applications.

In the first part, we present a novel approach to personalized medicine, applying causal statistical learning to observational data to develop individualized treatment rules. We focus on optimizing the timing of renal replacement therapy initiation in acute kidney injury, demonstrating: (i) the estimation and validation of an optimal dynamic strategy, and (ii) a comprehensive framework for evaluating individualized rules using observational data.

In the second part, we tackle the challenge of evaluating LLMs in medicine, focusing on the generation of hospital course summaries. Current evaluation methods are often either unscalable (physician-led) or untrustworthy for clinical settings (LLM-as-a-judge). We propose a rubric-based approach to LLM evaluation that combines the scalability of automated methods with the trustworthiness demanded by medical applications, paving the way for responsible deployment of LLMs in healthcare.

Data-driven medicine, LLMs, clinical applications, LabéliséHallesIA

Inria Montpellier, St-Priest Campus, Building 5, Room 01/124
Machine Learning in Montpellier, Theory & Practice

Data.table is one of the most efficient open-source in-memory data manipulation packages available today. First released to CRAN by Matt Dowle in 2006, it continues to grow in popularity, and now over 1500 other CRAN packages depend on data.table. This talk will start with data reading from CSV, discuss basic and advanced data manipulation topics, and finally will end with a discussion about how you can contribute to data.table.
https://github.com/tdhock/2023-10-LatinR-data.table?tab=readme-ov-file#source-files-for-latinr-datatable-tutorial-slides

LabéliséHallesIA

Inria Montpellier, St-Priest Campus, Building 5, Room 02/022
Machine Learning in Montpellier, Theory & Practice

Machine learning classifiers often produce probabilistic predictions that are critical for accurate and interpretable decision-making in various domains. The quality of these predictions is generally evaluated with proper losses like cross-entropy, which decompose into two components: calibration error assesses general under/overconfidence, while refinement error measures the ability to distinguish different classes.
In this paper, we provide theoretical and empirical evidence that these two errors are not minimized simultaneously during training. Selecting the best training epoch based on validation loss thus leads to a compromise point that is suboptimal for both calibration error and, mostimportantly, refinement error. To address this, we introduce a new metric for early stopping and hyperparameter tuning that makes it possible to minimize refinement error during training. The calibration error is minimized after training, using standard techniques. Our method integrates seamlessly with any architecture and consistently improves performance across diverse classification tasks.

LabéliséHallesIA

Campus St Priest
Rebecca Farina

Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage. Our approach is inspired by modern conformal inference literature, in that, unlike classical frameworks, we obviate the need for a well-specified parametric or semi-parametric survival model to accomplish our goal. In contrast to existing conformal prediction methods for survival data, which restrict censoring to be of Type I, whereby potential censoring times are assumed to be fully observed on all units in both training and validation samples, we consider the more common right-censoring setting in which either only the censoring time or only the event time of primary interest is directly observed, whichever comes first. Under a standard conditional independence assumption between the potential survival and censoring times given covariates, we propose and analyze two methods to construct valid and efficient lower predictive bounds for the survival time of a future observation. The proposed methods build upon modern semiparametric efficiency theory for censored data, in that the first approach incorporates inverse-probability-of-censoring weighting (IPCW), while the second approach is based on augmented-inverse-probability-of-censoring weighting (AIPCW). For both methods, we formally establish asymptotic coverage guarantees, and demonstrate both via theory and empirical experiments that AIPCW substantially improves efficiency over IPCW in the sense that its coverage error bound is of second-order mixed bias type, that is doubly robust, and therefore guaranteed to be asymptotically negligible relative to the coverage error of IPCW.

LabéliséHallesIA

Campus St Priest
Machine Learning in Montpellier, Theory & Practice
Jim Smit

Despite several randomised controlled trials (RCTs) on the use of adjuvant treatment with corticosteroids in patients with community-acquired pneumonia (CAP), the effect of this intervention on mortality remains controversial. We aimed to evaluate heterogeneity of treatment effect (HTE) of adjuvant treatment with corticosteroids on 30-day mortality in patients with CAP.

LabéliséHallesIA

Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice
Tiffany Ding (UC Berckley)

In the first part of the talk, I will present some reflections on the purpose of prediction sets and the role that statistics can play in forming useful prediction sets. In particular, I will discuss how prediction sets fit into a decision making pipeline and the different kinds of decisions one may make using a prediction set. In the second part of the talk, I will describe a particular statistically motivated set-generating procedure for the classification setting called clustered conformal prediction, which gives all classes an equal chance of being correctly included in the prediction set (“class-conditional coverage”). This procedure can be useful in situations where it is important to identify instances of all classes, even the rare ones. We demonstrate the performance of this method on ImageNet and other image classification datasets.

LabéliséHallesIA

Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice
Jacob Feitelberg

We study the problem of distributional matrix completion: Given a sparsely observed matrix of empirical distributions, we seek to impute the true distributions associated with both observed and unobserved matrix entries. This is a generalization of traditional matrix completion where the observations per matrix entry are scalar-valued. To do so, we utilize tools from optimal transport to generalize the nearest neighbors method to the distributional setting. Under a suitable latent factor model on probability distributions, we establish that our method recovers the distributions in the Wasserstein metric. We demonstrate through simulations that our method (i) provides better distributional estimates for an entry compared to using observed samples for that entry alone, (ii) yields accurate estimates of distributional quantities such as standard deviation and value-at-risk, and (iii) inherently supports heteroscedastic distributions. In addition, we demonstrate our method on a real-world quarterly earnings predictions dataset. We also prove novel asymptotic results for Wasserstein barycenters over one-dimensional distributions.

LabéliséHallesIA

Inria Montpellier, St-Priest Campus, Building 5, Room 02/022
Machine Learning in Montpellier, Theory & Practice
Toby Hocking

Data.table is one of the most efficient open-source in-memory data manipulation packages available today. First released to CRAN by Matt Dowle in 2006, it continues to grow in popularity, and now over 1500 other CRAN packages depend on data.table. This talk will start with data reading from CSV, discuss basic and advanced data manipulation topics, and finally will end with a discussion about how you can contribute to data.table.

https://github.com/tdhock/2023-10-LatinR-data.table?tab=readme-ov-file#source-files-for-latinr-datatable-tutorial-slides

LabéliséHallesIA

Room Nadir, Maison de la Télédétection (500 rue Jean François Breton)
Machine Learning in Montpellier, Theory & Practice - Lars Van der Laan

We introduce efficient plug-in (EP) learning, a novel framework for the estimation of heterogeneous causal contrasts, such as the conditional average treatment effect and conditional relative risk. The EP-learning framework enjoys the same oracle-efficiency as Neyman-orthogonal learning strategies, such as DR-learning and R-learning, while addressing some of their primary drawbacks, including that (i) their practical applicability can be hindered by loss function non-convexity; and (ii) they may suffer from poor performance and instability due to inverse probability weighting and pseudo-outcomes that violate bounds. To avoid these drawbacks, EP-learner constructs an efficient plug-in estimator of the population risk function for the causal contrast, thereby inheriting the stability and robustness properties of plug-in estimation strategies like T-learning. Under reasonable conditions, EP-learners based on empirical risk minimization are oracle-efficient, exhibiting asymptotic equivalence to the minimizer of an oracle-efficient one-step debiased estimator of the population risk function. In simulation experiments, we illustrate that EP-learners of the conditional average treatment effect and conditional relative risk outperform state-of-the-art competitors, including T-learner, R-learner, and DR-learner. Open-source implementations of the proposed methods are available in our R package hte3.

LabéliséHallesIA

Room Nadir, Maison de la Télédétection (500 rue Jean François Breton)
Machine Learning in Montpellier, Theory & Practice - Martin Atzmüller, Scientific Director at DFKI and Full Professor at Osnabrück University (Germany)

In many applications, modeling complex data is of utmost importance, requiring the use of advanced machine learning models and approaches. However, in many domains users require insight into models and/or their decisions, which is not necessarily provided by the respective models per se. Explainable and interpretable learning approaches can facilitate such insights for making sense of models and decisions. The talk presents examples of such approaches in complex modeling domains, including interpretable as well as explainable deep-learning-based methods, and a neuro-symbolic architecture including domain knowledge for facilitating explainability.

LabéliséHallesIA

Inria Montpellier, St-Priest Campus, Building 5, Room 02/124
Machine Learning in Montpellier, Theory & Practice - Vincent Mallet

Learning on the 3D structure of biomolecules resulted in several
breakthroughs, for instance the de-novo design of potent binders or
functional luciferases, eventually leading to a Nobel Prize. Most of
this work relies on modeling the structure of protein as
graphs, with well-established tools.

IA&Experts, LabéliséHallesIA

Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice - Olivier Cappé (CNRS / ENS / Université PSL)

Performative learning addresses the increasingly pervasive situations in which algorithmic decisions may induce changes in the data distribution as a consequence of their public deployment. We propose a novel view in which these performative effects are modelled as push-forward measures. This general framework encompasses existing models and enables novel performative gradient estimation methods, leading to more efficient and scalable learning strategies. For distribution shifts, unlike previous models which require full specification of the data distribution, we only assume knowledge of the shift operator that represents the performative changes. Focusing on classification with a linear-in-parameters performative effect, we prove the convexity of the performative risk under a new set of assumptions. We also establish a connection with adversarially robust classification by reformulating the minimization of the performative risk as a min-max variational problem.

IA&Experts

Salle 104 du bâtiment 11 dit le château, 2 Place Pierre Viala Campus La Gaillarde

Le séminaire commencera par un café offert par les Halles de l'IA. Au programme :
(1) News du W3C en 10 minutes, par Pierre-Antoine CHAMPIN, W3C / INRIA / Université de Lyon
(2) Propagation de propriétés entre entités contextuellement identiques par Pierre Henri PARIS, Universite Paris-Saclay :

Inria Montpellier, St-Priest Campus, Building 5, Room 02/022
Machine Learning in Montpellier, Theory & Practice

In decentralized machine learning, different devices communicate in a peer-to-peer manner to collaboratively learn from each other's data. Such approaches are vulnerable to misbehaving (or Byzantine) devices. We introduce F-RG, a general framework for building robust decentralized algorithms with guarantees arising from robust-sum-like aggregation rules F. We then investigate the notion of breakdown point, and show an upper bound on the number of adversaries that decentralized algorithms can tolerate. We introduce a practical robust aggregation rule, coined CSours, such that CSours-RG has a near-optimal breakdown. Other choices of aggregation rules lead to existing algorithms such as ClippedGossip or NNA. We give experimental evidence to validate the effectiveness of CSours-RG and highlight the gap with NNA, in particular against a novel attack tailored to decentralized communications.

MachineLearning, LabéliséHallesIA, IA&Expert

Machine Learning in Montpellier, Theory & Practice - Pedro Valdeira

Vertical federated learning trains models from feature-partitioned datasets across multiple clients, who collaborate without sharing their local data. Standard approaches assume that all feature partitions are available during both training and inference. Yet, in practice, this assumption rarely holds, as for many samples only a subset of the clients observe their partition. However, not utilizing incomplete samples during training harms generalization, and not supporting them during inference limits the utility of the model.

Vertical Federated Learning, Missing Feature Blocks, LabéliséHallesIA

Salle 104 du bâtiment 11 dit le château, 2 Place Pierre Viala Campus La Gaillarde, 34000 Montpellier
SEmantic web SeminAr MontpEllier

Le séminaire commencera par un café offert par les Halles de l'IA. Au programme :
(1) News du W3C en 10 minutes, par Pierre-Antoine CHAMPIN, W3C / INRIA / Université de Lyon
(2) FAIRiser les données du réseau d’observatoires de la zone critique Français (OZCAR) : cas d’utilisation de l’ontologie I-ADOPT par Charly COUSSOT, IRD Université Grenoble Auvergne Rhône Alpes

LabéliséHallesIA

Salle SC 003, Bâtiment 31, Polytech, Campus Triolet
Équipe Montpellier BioInfo

C’est l’occasion pour les étudiants de Master 1 et 2 en bioinfo de présenter l’avancée de leur stage, en vue de leur soutenance. L'objectif : s’entraîner à l’oral, partager ses travaux et recevoir des retours constructifs dans une ambiance bienveillant. Quelques sujets au programme : développement de l'outil INpalm pour prédire les émissions d’azote, étude de la frustration locale des protéines et analyse de l’impact du génome de référence sur le mapping et l’hétérozygotie en génomique et bien d’autres encore ! Venez nombreux pour encourager les présentateurs et découvrir la diversité des projets.

Master1et2, BioInfo, Stage, LabéliséHallesIA

Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice - Pedro Valdeira

Julien Patras (IRD / Marbec) will give a talk : Seabirds serve as key bioindicators of ocean health, and with modern tracking technologies like GPS tags, we now have rich datasets capturing their offshore movements. These trajectory data enable us to study behavioral states and identify ecologically significant foraging zones. However, modeling how environmental variables (like wind, sea surface temperature, or currents) shape these trajectories remains largely unexplored. Trajectory simulation using generative artificial intelligence has shown great promise in fields like autonomous driving and human motion modeling. Inspired by these advances, we introduce a novel approach in movement ecology by adapting DiffTraj (Zhu et al., 2023),a pretrained diffusion model originally developed for vehicle trajectories. We then adapt and fine-tune the model on seabird movement data, so that it supports multi-species, multi-site simulation, with the aim of environmental conditioned generation, and improving state-of-the-art seabird trajectory simulation.

Seabirds, Trajectories, Environmental variables, LabéliséHallesIA

Salle 104 du bâtiment 11 dit le château, 2 Place Pierre Viala Campus La Gaillarde, 34000 Montpellier
SEmantic web SeminAr MontpEllier (MISTEA) - Vincent ARMAND (IRD Montpellier)

Thanks to open science initiatives, thousands of standardised datasets on Earth System compartments are now available on the web. In particular, we have widely used ISO 19115 to encode spatiotemporal aspects of Earth System observations. However, this standard does not specify the multiple dimensions of observations, including the features of interest, the observable properties, and the provenance. As a result, researchers interested in Earth System multi-disciplinary studies may miss meaningful datasets when querying independently domain-specific data portals.

IA & Experts

Campus Triolet, Bat 36, SC36.08
Montpellier BioInfo - présenté par Iago Bonnici

R, Python, Perl, Java, C++, Julia, Rust.. Mais qu’est-ce qui les relie vraiment ? Voici quelques repères pour s’orienter et trouver le bon outil au bon usage.

Salle de séminaire du Bât 4 - Campus Saint Priest
LIRMM - Prof. Stefan Dietze

Dans le cadre de l’axe transverse IA – Science des données, le LIRMM a le plaisir d’accueillir le Prof. Stefan Dietze(Heinrich-Heine-Universität Düsseldorf, GESIS, HeiCAD, L3S). Résumé : Au cours des dernières décennies, les sciences sociales ont progressivement intégré de nouvelles sources de données issues du web et des réseaux sociaux, ainsi que des méthodes computationnelles de plus en plus complexes, portées par l’essor de l’IA et du traitement automatique du langage naturel. Ces évolutions ouvrent de nouvelles perspectives, mais soulèvent également des défis majeurs en termes de transparence, de reproductibilité et de qualité des données

IA & Experts

Amphi Moreau, B2. Campus St Priest
Machine Learning in Montpellier, Theory & Practice - Nicolas Papernot (Univ. Toronto)

Pour cette session exceptionnelle du séminaire ML-MTP, nous aurons l'honneur d'accueillir Nicolas Papernot, chercheur de renommée mondiale en IA de confiance et expert des questions de sécurité et de vie privée. Il a été l'un des pionniers de plusieurs axes de recherche majeurs, notamment les attaques adversariales sur les réseaux de neurones et le machine unlearning. Il est titulaire d'une Inria International Chair (2025–2027) associée à l'équipe PreMeDICaL.

IA & Experts

CIRAD Campus Lavalette - Amphi Jacques Alliot - Bâtiment 4
Machine Learning in Montpellier, Theory & Practice - Étienne Laliberté (Université de Montréal)
‎
Dans le cadre des échanges Inria–IVADO, nous aurons le plaisir d’accueillir à Montpellier du 3 au 14 novembre 2025, Étienne Laliberté, professeur titulaire au Département des sciences biologiques de l'Université de Montréal et membre académique associé au Mila (Institut québécois d'intelligence artificielle).
‎
IA et Experts

Building 5, 02.124, Campus St Priest
Machine Learning in Montpellier, Theory & Practice - Matthieu de Castelbajac (Univ. Montpellier)
‎

Citizen science records are a valuable source of biodiversity data, and even more essential to help track mobile marine species like jellyfish. However, these records can be highly uncertain, containing many potential errors and biases. They are typically validated by experts, which is impractical at scale. Although deep learning methods for automatic validation have shown promising results, they fail to account for the uncertainty present in both the input data and their predictions. Here, we present a semi-automated method to support record validation at scale while providing strong statistical guarantees, including for highly uncertain citizen science records.
‎

IA et Experts

Inria Montpellier, St-Priest Campus, Building 5, Room 02/124
Machine Learning in Montpellier, Theory et Practice - Emmy Fang & Arielle Zhang

Machine Learning (ML) models typically benefit from access to large and diverse datasets across different parties. However, privacy concerns often prohibit direct data sharing or centralized data aggregation. To address this, various collaborative training frameworks have been proposed that enable joint model training while adhering to differential privacy (DP), the current golden standard privacy definition. In this talk, I will provide an overview of existing privacy-preserving collaborative ML frameworks, highlighting their core techniques and their limitations.

IA et Experts

Salle 104 du bâtiment 11 dit le château, 2 Place Pierre Viala Campus La Gaillarde, 34000 Montpellier
SEmantic web SeminAr MontpEllier (MISTEA) - Nathalie Abadie

Les graphes de connaissances constituent un modèle de données privilégié pour représenter des entités - géographiques - qui évoluent au cours du temps. Plusieurs travaux ont proposé des ontologies et des méthodes de peuplement pour représenter l'évolution d'unités administratives ou statistiques à partir de jeux de données millésimés décrivant les états successifs de ces entités géographiques

IA et Experts