Workshop — November 3–4, 2026
Program
Two days of plenary sessions, workshops, round tables, and poster presentations on transfer learning and foundation models.
Day 1 — Transfer Learning - Tuesday, November 3, 2026
Welcome and coffee
Participant registration, welcome coffee.
Transfer Learning. Theory, methods and applications
Mathilde Mougeot
It is well known the greater the volume of data the better the performance of a model. However, for given applications, access to large databases to train a model is not necessarily easy. The aim of transfer learning is to take advantage of basic model sources or pre-existing data in order to adapt a high-performance model to a given target application for which sufficient data is not directly available. This course will present various transfer learning methods (instance-based, feature-based, parameter-based) that have been introduced in the literature depending on the information available for the source and the target. Different approaches will be presented depending on the class of models used: models based on neural networks, models based on decision trees, models based on Gaussian processes, etc. Concrete application cases will also be presented in which transfer learning has been used to obtain very high-performance machine learning models, even when dedicated data are scarce or even non-existent.
Transfer Learning. Practical workshop
Mathilde Mougeot
Part of the concepts presented before will be illustrated from a practical point of view using the ADAPT library, which offers free access to a large majority of transfer learning algorithms presented in the literature. These algorithms will be used both on simulated and real databases. By taking into account all the former elements (presentation of the methods, illustration on concrete cases and operational practice on simple examples), this course offers all the ingredients necessary to tackle problems of transfer learning on your own using new data.
Lunch
Transfer learning for Conditional Average Treatment Effect (CATE) estimation.
Véronique Maume-Deschamps
The talk presents a joint work with Bérénice-Alexia Jocteur and Pierre Ribereau. It will focus on model shift and transfer learning applied for a causal forest namely HTERF. This causal forest aims to estimate the Conditional Average Treatment Effect (CATE). The approach considered is the offset method presented by Wang (2016) adapted to a causal context. This method relies on the use of intermediate models in order to estimate the offset between source and target distributions. Our main result is a bound on the CATE error of HTERF on target depending on the error of the intermediate models.
Langevin-Informed Transfer Learning: Replacing Target Samples by Black-Box Feedback
Vladimir R. Kostic
Many scientific and machine learning systems from molecular dynamics to diffusion models and beyond are governed by stochastic dynamics with low-dimensional structure evolving on slow timescales. However target trajectories used to identify and interpret such dynamics are often inaccessible: only biased or static samples that explore the underlying manifold are available. In this talk we present Langevin-Informed Transfer Learning (LITL) a novel framework for recovering target Langevin dynamics from biased source samples using only black-box feedback. LITL learns the leading spectral structure of the target infinitesimal generator and the projected drift through Dirichlet representation learning, thus enabling kinetic reconstruction in spectral form and slow-manifold gradient field estimation. In particular, we will discuss a spherical variant well suited to steering normalized latent representations commonly used in learning systems toward desired objectives. While we will cover finite-sample guarantees for eigenvalue eigenfunction and projected drift estimation in Sobolev norms, we will also show empirically how LITL recovers physical transition timescales from biased molecular simulations, now it builds kinetic structure from static samples of generative models and how it reconstructs spherical symmetries of physical systems. Finally, we will demonstrate how it enables post-hoc latent steering of trained neural networks under black-box feedback.
Adaptation to data shift without labels : Methods, benchmarks and domain adaptation on biomedical signals.
Rémi Flamary
Domain Adaptation (DA) is a fundamental Machine Learning challenge that aim at adapting supervised predictors to new data in the absence of labels on the new data. We will first present the problem and classical approaches that have been proposed to tackle data shift. Then we will introduce a multimodal benchmark of those methods and discuss the results, highlighting the limits of the current approaches in practice. Finally, we will focus on a specific modality, electroencephalograms (EEGs) and the problem of sleep stage classification, and present an approach based on optimal transport that allows to model, and use the specificity of the different subjects for better personalized predictions.
Poster session + discussions
Presentation of selected posters, informal discussions with the authors over refreshments.
Day 2 — Time Series Foundation Models - Wednesday, November 4, 2026
Welcome and coffee
Participant registration, welcome coffee.
Introduction to Deep Learning for Time Series: Towards Better Generalization at Inference
Adrien Petralia (Mistral AI), Etienne Le Naour (EDF)
This introductory talk provides the background needed to understand the recent emergence of foundation models for time series. It begins with a concise overview of the deep learning methods that have shaped the field over the past decade, covering key architectures and their applications to forecasting, classification, imputation, regression and anomaly detection. It then discusses the limitations of task- and dataset-specific training, motivating the need for models that learn more general and transferable temporal representations. This evolution naturally leads to the paradigm of foundation models, which will be explored in depth throughout the subsequent talks of the workshop.
TabICLv2: A Better, Faster, Scalable, and Open Tabular Foundation Model
David Holzmüller or Gaël Varoquaux (INRIA Soda)
Foundation models are rapidly transforming tabular machine learning, demonstrating the potential of large-scale pretraining and in-context learning for predictive tasks. This talk presents TabICLv2, a new open foundation model for tabular data that advances the state of the art in both performance and scalability. The presentation highlights the key ideas behind the model, including innovations in synthetic data generation, model architecture, and pretraining, and discusses how these advances enable stronger generalization across datasets while making foundation models more practical for real-world applications.
Lunch
TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning
Tahar Nabil(EDF), Etienne Le Naour (EDF)
While most time series foundation models have focused on forecasting, many real-world applications require reasoning over irregularly and partially observed data. This talk presents TS-ICL, a foundation model that unifies forecasting and imputation within a single in-context learning framework. By casting time series problems as timestamp-aligned regression and leveraging large-scale synthetic pretraining, TS-ICL learns flexible temporal representations that generalize across tasks and observation settings. The presentation discusses the core principles behind the model and illustrates how in-context learning can enable more versatile and robust time series foundation models.
CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data
Shifeng Xie (Université Paris Cité)
The success of time series foundation models largely relies on large-scale pretraining datasets, whose collection and curation remain costly and challenging. This talk presents CauKer, a synthetic data generation framework designed to pretrain classification foundation models on diverse and causally coherent time series. By combining Gaussian Process kernels with Structural Causal Models, CauKer generates realistic synthetic data that enables effective pretraining across different architectures. The presentation discusses how synthetic data can serve as a scalable alternative to real-world collections and sheds light on the scaling behavior of time series foundation models.
Coffee break
Retrieval Over Training: Similarity Search-Based Model Selection for Time Series Anomaly Detection.
Roberto Stanzione (INRIA Valda)
Selecting effective anomaly detection models remains a major challenge in time series analysis, as no single method consistently performs well across diverse datasets and application scenarios. This talk presents RAMSAD, a retrieval-based framework that addresses this challenge through similarity search rather than additional model training. By retrieving previously observed time series with similar characteristics, RAMSAD leverages their performance profiles to recommend suitable anomaly detectors for new instances. The presentation discusses how retrieval-based approaches can improve model selection, particularly in out-of-distribution settings, and provide a scalable alternative to traditional adaptation strategies.
Industry Talk
Raphaël Nedellec (Decathlon)
Forecasting in operation at Decathlon