1-3 Apr 2026 Grenoble (France)

Program > Program and invited speakers

Tutorial (bring your laptop!)

Diffusion models (April 1st, morning)

Invited speakers

Topic #1 - Causality (April 1st, afternoon)

Topic #2 - Generative models (April 2nd, morning)

The talk of Stéphane Girard has been canceled due to unforeseen circumstances. We apologize for the inconvenience.

Topic #3 - Non-convex optimization (April 2nd, afternoon)

Topic #4 - Sequential learning and reinforcement (April 3rd, morning)

 

Talk Titles and Abstracts

Gabriel Victorino Cardoso: Introduction to sampling from score-based generative models

In this talk I will introduce score-based generative models in it's most diverse formulations, which each yield different sampling strategies. Each sampling strategy will be implemented in Python (Pytorch) with pre-trained networks.

Krikamol Muandet: When Shift Happens — Confounding is To Blame

When shift happens, who is to blame? Prior work has observed that classical methods such as empirical risk minimisation (ERM) and boosting can outperform specialised approaches in out-of-distribution (OOD) generalisation. In this talk, I argue that much of this phenomenon can be traced to confounding [1]. In real-world tabular settings, distribution shifts are often driven by changes in unobserved confounders, simultaneously violating standard assumptions so that covariate, conditional, and label shifts may arise together. This perspective helps explain why many sophisticated OOD methods underperform relative to classical approaches like boosting. But how can we account for shifts in variables that are inherently unobservable, i.e., unknown unknowns? I will further argue that boosting methods possess an implicit mechanism that induces a form of epistemic humility, enabling them to remain robust even under such unobserved shifts [2].
 
References:
[1] When Shift Happens — Confounding Is To Blame, ICLR2026 (To Appear, https://arxiv.org/abs/2505.21422)
[2] Boosting for Predictive Sufficiency, ICLR2026 (To Appear, https://cispa.de/en/research/publications/104607-boosting-for-predictive-sufficiency)

Julie Josse: Personalized Care Through Causal & Federated Learning: From Data to Decisions

In this talk, we revisit a fundamental question: how can we reliably measure the causal effect of an intervention or treatment on an outcome? While this may appear straightforward, the problem is not trivial even when outcomes are simple and univariate. Re-examining this question invites us to rethink one of the pillars of evidence-based medicine: meta-analysis, traditionally considered the highest level of clinical evidence. Yet, modern medical research increasingly relies on heterogeneous sources of information—from randomized controlled trials to real-world data—raising new methodological and practical challenges for how evidence should be combined.
We present recent developments in causal meta-analysis, a framework designed to estimate and aggregate treatment effects across diverse data sources, including aggregated study summaries, randomized trials, and observational real-world evidence.  By combining causal inference with federated learning, this paradigm enables the estimation and aggregation of treatment effects across decentralized data silos while preserving privacy and avoiding the need to share individual-level data. Together, these advances outline a new paradigm for evidence synthesis—one that is causal, privacy-preserving, and scalable—and that moves us closer to translating heterogeneous data into interpretable decisions for personalized care. 

Alexander Reisach: The Case for Time in Causal DAGs

We make the case for incorporating a notion of time into causal directed acyclic graphs (DAGs). We demonstrate that nontemporal causal DAGs are ambiguous and obstruct justification of the acyclicity assumption. Assuming that causes precede effects, causal relationships are relative to the time order, and causal DAGs require temporal qualification. We propose a formalization via composite causal variables that refer to quantities at one or multiple time points. We emphasize that the acyclicity assumption requires different justifications depending on whether the time order allows cycles. We discuss and illustrate implications for the interpretation and applicability of DAGs as causal models.

Ségolène Martin: Flow Matching Meets Denoising: A Plug-and-Play Approach to Inverse Problems

In this talk, we explore the connections between flow matching and denoising, and show how these links can be leveraged to solve imaging inverse problems such as super-resolution and inpainting. We begin with a brief introduction to flow matching generative models, and recall the classical formulation of inverse problems, along with standard non-generative approaches, with a focus on the Plug-and-Play (PnP) framework.

We then introduce PnP-Flow, a method that bridges flow matching and PnP. Our approach constructs a time-dependent denoiser from a pre-trained flow matching model, and integrates it into an iterative scheme combining data-fidelity updates, reprojections onto the flow trajectory, and denoising steps. The resulting algorithm is efficient, memory-friendly, and achieves strong performance across a range of inverse problems.

The second part of the talk focuses on how to train effective generative denoisers and provides practical guidelines. In particular, we investigate the impact of time weighting in the loss, as well as the parameterization of the neural network (e.g., predicting velocity, noise, or the denoised signal). Although these formulations are theoretically equivalent under perfect training, we show that they lead to markedly different empirical performance.

Eric Vanden Eijnden - Generative modeling with flows and diffusions

Generative models based on dynamical transport have recently led to significant advances in unsupervised learning. At mathematical level, these models are primarily designed around the construction of a map between two probability distributions that transform samples from the first into samples from the second.  While these methods were first introduced in the context of image generation, they have found a wide range of applications, including in scientific computing where they offer interesting ways to reconsider complex problems once thought intractable because of the curse of dimensionality. In this talk, I will discuss the mathematical underpinning of generative models based on flows and diffusions, and show how a better understanding of their inner workings can help improve their design. These results indicate how to structure the transport to best reach complex target distributions while maintaining computational efficiency, both at learning and sampling stages.  

Constance Douwes : Energy and environmental impacts of deep learning

In recent years, artificial intelligence based on deep learning models has become increasingly widespread across a wide range of applications, achieving ever-greater performance with ever-increasing computational cost. This growth drives higher energy consumption and carbon emissions, running counter the emission reduction pathways recommended by the IPCC. In this presentation, I will present and explore the environmental impacts associated with developing, training, and deploying deep learning models using the Life-Cycle Assessment methodology. I aim to provide useful methods for estimating the impacts and to raise awareness of the environmental challenges posed by the rapid growth of AI, thereby encouraging more efficient and responsible practices.

Nicolas Boumal: Smooth, globally Polyak-Łojasiewicz functions are nonlinear least-squares

PŁ functions abound in the literature, especially in nonconvex optimization. When they are also smooth, they become surprisingly simple---with an exotic twist. I'll try for us to have an interactive session about the structure of those functions and of their sets of minimizers.

Joint work with Christopher Criscitiello and Quentin Rebjock.

Nelly Pustelnik: Nonconvex proximal optimisation for solving inverse problems

This presentation explores nonconvex optimization challenges in solving nonlinear inverse problems, where nonlinearity may arise either in the data fidelity term or in the regularization penalty. We will discuss various algorithmic strategies and their associated convergence guarantees. Special focus will be given to the discrete Mumford-Shah model for joint image restoration and contour detection, as well as the recovery of DNA replication dynamics. These applications highlight the importance of tailored optimization approaches to address the complexities of nonlinear inverse problems.

Le Quoc Tung: VC-type Dimension Bounds for Nonconvex Bilevel Optimization in Hyperparameter Tuning

Data-driven algorithm design automates hyper-parameter tuning, but its statistical foundations remain limited because model performance can depend on hyperparameters in implicit and non-smooth ways. Existing guarantees focus on the simple case of a one-dimensional (scalar) hyperparameter. This leaves the practically important, multi-dimensional hyperparameter tuning setting unresolved. We address this open question by establishing the first general framework for establishing generalization guarantees for tuning multi-dimensional hyperparameters in data-driven settings. Our approach strengthens the generalization guarantee framework for semi-algebraic function classes by exploiting tools from real algebraic geometry, yielding sharper, more broadly applicable guarantees. We then extend the analysis to hyperparameter tuning using the validation loss under minimal assumptions, and derive improved bounds when additional structure is available. Finally, we demonstrate the scope of the framework with new learnability results, including data-driven weighted group lasso.

Chloe Rouyer: Sequential Learning for Non-parametric Regression

We consider a non-parametric regression problem in which a learner is given a budget of n samples to best approximate an unknown function f. The learner has full control over how to sample the query points x_1, …, x_n, and for each sample it receives a noisy observation of f(x_i).  In this talk, we investigate whether and how a sequential approach, meaning that the learner can query points and receive observations one by one rather than all at once, can speed up learning and provide refined guarantees compared to a fixed design. After identifying which classes of functions benefit from the sequential framework, we propose a simple algorithm that can adapt to this structure. 

Emilie Kaufman: Posterior Sampling Beyond Regret Minimization

Thompson Sampling, or Posterior Sampling, is a popular principle for sequential decision making, which has mostly been used in the so-called regret minimization paradigm, when the agent's goal is to maximize rewards while learning. In this talk, I will discuss variants of Posterior Sampling for pure exploration tasks in multi-armed bandit models. In bandit pure exploration there is no incentive on maximizing rewards but the agent should identify quickly and accurately the answer to some questions about the unknown arms' distributions. 
Loading... Loading...