If you are interested in presenting at the seminar, please contact us at ml-seminar@polytechnique.fr.

Upcoming Seminars

Thursday, November 12, 2026 (11h-12h) (SGS)

• Adrien Bolland: Reinforcement learning for emergency voltage control and blackout prevention in electrical grids
  Abstract

  Voltage instability has returned to the center of power-system operation. Such instabilities have contributed to the blackouts of Athens in 2004 and of the Iberian Peninsula in 2025. They occur when, after a disturbance, restoring the operating point demands more than the grid can supply or absorb, and voltages drift over minutes, down or up, beyond their limits. This presentation reviews the control schemes in place and the emergency controls proposed in the literature, and what remains open. We then discuss the emergency control of on-load tap changers in the face of contingencies. We formulate the problem as a chance-constrained decentralized POMDP, in which differential-algebraic equations with random disturbances describe the power system and safety is stated as probabilities of voltage excursion and collapse. We describe the learning challenges it raises, arising from slow simulations, rare events and a combinatorial action space, and the approaches that address them. We show preliminary results, in which controllers trained on a simulated grid avert collapses and exhibit different trade-offs between security and control effort, and close with the questions the problem leaves open to mathematics and machine learning.

Thursday, November 19, 2026 (11h-12h) (SGM)

• Andreas Habring: To be announced

Thursday, December 3, 2026 (11h-12h) (SGS)

• Théo Lacombe: Learning Monge Maps with Constrained Drifting Models
  Abstract

  The estimation of optimal transport maps (a.k.a. Monge maps) is a central problem in optimal transport literature. Recent observations reveal that the flow map of the Wasserstein gradient flow of the relative entropy closely approximates—though does not exactly equal—the Monge map between a given source distribution and a Gaussian target. In this work, we demonstrate how the evolution equation governing this flow map can be corrected to form a constrained gradient flow that provably converges to the true Monge map. When the maps are parametrized as gradients of convex models (e.g. ICNN), we show that the resulting optimization scheme can be interpreted as a L2-natural gradient descent in the parameter space, and that this approach is further connected to the recently introduced drifting generative models when coupled with a Gauss--Newton update. On toy experiments, we empirically illustrate the clear advantage of using that L2-natural gradient instead of the Euclidean for estimating Monge maps in this framework. Joint works with T. Dumont and F.-X. Vialard.

Past Seminars

Group Meeting Tuesday, October 20, 2026 (11h-12h) (SGM)

• Gayane Tarturyan: A Comprehensive View of Fairness through Distributional Stability
  Abstract

  We view fairness as a property of distributional stability. Rather than assessing a predictor under a fixed data distribution, we study how its predictions change under perturbations that modify the composition of protected groups. A predictor is fair if it remains stable under such shifts. Under this perspective, several classical notions of fairness arise as stability with respect to specific perturbations, with the associated unfairness gap given by a Lipschitz constant of a prediction-rate functional. This formulation also yields guarantees that hold uniformly over a range of demographic compositions at test time, without requiring knowledge of the deployment distribution. It leads to a learning procedure based on convex combinations of reweighted predictors, formulated as a second-order cone program, for which we establish generalization bounds. Experiments on standard benchmarks illustrate the approach. Joint work with Charlotte Laclau and Stephan Clémençon.
• Marie Generalli-Lince: Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings
  Abstract

  Fairness-accuracy trade-offs are a central concern in the deployment of fairness-aware machine learning methods. When sensitive attributes are unavailable at inference time–the so called unawareness setting, principled methods for obtaining accurate predictions under relaxed fairness constraints are largely missing. In this work, we address this gap by formulating regression under a demographic parity penalty as an optimal transport problem. Our framework, unifies both the aware and unaware settings and characterizes optimal prediction functions via optimal transport maps, under both squared Wasserstein-2 and Total Variation penalties. These results reveal that the choice of penalty reflects fundamentally different fairness philosophies: the Wasserstein penalty induces a smooth, population-wide compromise, while Total Variation enforces parity for a subset of individuals.

Tuesday, October 13, 2026 (11h-12h) (SGS)

• Konstantinos Vlachas: Value-Driven Digital Twinning via Generative Learning and Adaptive Reduced-Order Modeling
  Abstract

  Infrastructure operation and maintenance relies on many-query predictive and inverse tasks, which hinge on computational models that are both accurate and tractable in real time. However, the underlying physical systems are governed by nonlinear, high-dimensional dynamics and operate under stochastic and evolving conditions that often push their response beyond the assumptions under which models were built. Critically, static models, which are often limited in interpretability, brittle under extrapolation, and tied to their training conditions, cannot remain adequate across the full life cycle of a system, even when equipped with passive data assimilation. This work addresses this gap by formulating digital twinning as tracking an evolving system within a space of models. The resulting framework rests on three interconnected components: (i) reduced physics baselines that keep these tasks tractable and interpretable for PDE-governed dynamical systems, (ii) generative learning that builds on these baselines to construct families of models across operating regimes and extends them as the system evolves or unexpected events occur, and (iii) a value-driven strategy that quantifies the utility of candidate models under uncertainty, so that model adaptation is guided by the task at hand, such as damage assessment or maintenance planning. I will illustrate these ideas through my work on forward and inverse problems and outline how they form the basis of my SNSF Postdoc.Mobility project at ENSAM.

Group Meeting Thursday, June 18, 2026 (10h-12h) (SGM)

• Julien Fageot: Byzantine Machine Learning: MultiKrum and an optimal notion of robustness
  Abstract

  Aggregation rules are the cornerstone of distributed (or federated) learning in the presence of adversaries, under the so-called Byzantine threat model. They are also interesting mathematical objects from the point of view of robust mean estimation. The Krum aggregation rule has been extensively studied, and endowed with formal robustness and convergence guarantees. Yet, MultiKrum, a natural extension of Krum, is often preferred in practice for its superior empirical performance, even though no theoretical guarantees were available until now. In this work, we provide the first proof that MultiKrum is a robust aggregation rule, and bound its robustness coefficient. To do so, we introduce , the optimal robustness coefficient of an aggregation rule, which quantifies the accuracy of mean estimation in the presence of adversaries in a tighter manner compared with previously adopted notions of robustness. We then construct an upper and a lower bound on MultiKrum's robustness coefficient. We also characterize MultiKrum's breakdown point and show that it is optimal. As a by-product, we also improve on the best-known bounds on Krum's robustness coefficient. We show that MultiKrum's bounds are never worse than Krum's, and better in realistic regimes. We illustrate this analysis by an experimental investigation on the quality of our upper and lower bounds.
• Martin Beaufils: BRoADflip: Backdooring Robust Aggregators via Sub-Differentiability with Clean-Image Poisoning
  Abstract

  Among existing threats to machine learning, label-only poisoning is particularly stealthy: by modifying only training labels, an attacker can significantly alter a model's behavior without changing the input features, e.g. with clean-image poisoning. The state-of-the-art FLIP method shows that backdoors can be induced using label poisoning alone. In this work, we first show that the effectiveness of FLIP degrades across datasets, model architectures, and in the presence of robust aggregation defenses. Moreover, we identify an inconsistency between its formulation and its implementation, which we resolve. Our contributions are threefold: (i) we resolve a formulation inconsistency in FLIP and extend it to federated learning with BRoADflip; (ii) we establish convergence of the label optimization problem to Clarke critical points under mild regularity assumptions; and (iii) we design an aggregator-aware trigger mechanism for label-only backdoor attacks. In the same setup for which FLIP was designed, i.e., against the Mean aggregator, only label flips suffice for BRoADflip to reach attack success rate (ASR) without degrading the overall performance, whereas FLIP requires to reach only ASR and incurs a drop. Moreover, we show that BRoADflip consistently bypasses state of the art robust aggregation defenses (e.g., Krum and MultiKrum) while flipping as little as and of labels, respectively, whereas FLIP fails. To the best of our knowledge, this is the first work to formulate the attacker's optimization as stochastic subgradient problem over a nonsmooth trajectory-matching objective, and to prove its convergence. From an empirical angle, this is also the first successful label-only backdoor attack against sub-differentiable, robust aggregation rules.
• Adrien Majka: The Strong, weak and benign Goodhart’s law. An independence-free and paradigm-agnostic formalisation
  Abstract

  Goodhart's law is a famous adage in policy-making stating that ``When a measure becomes a target, it ceases to be a good measure''. As machine learning models and the optimization capacity to train them grow, increasing empirical evidence reinforced the belief in the validity of this law, without however formalizing it. Recently, a few attempts were made to formalize Goodhart's law, either by categorizing variants of it, or by looking at how optimizing a proxy metric affects the optimization of an intended goal. In this work, we relax the simplifying independence assumption, made in previous works, and the assumption on the learning paradigm made in most of them, to study the effect of the coupling between the proxy metric and the intended goal on Goodhart's law. Our results show that in the case where the tail of the goal strictly dominates the tail of the discrepancy, the dependence does not change the nature of Goodhart's effect. However, in the light tailed goal and heavy tailed discrepancy case, we provide an example where over-optimization occurs at a rate inversely proportional to the heavy-tailedness of the discrepancy between the goal and the metric. To the best of our knowledge, this is the first formalization of Goodhart's law that relieved both the assumption on independence and that does not assume knowledge of the learning paradigm.

Tuesday, June 2, 2026 (14h-15h) (SGS)

• Boris Bonev: Geometric Machine Learning for ML Probabilistic Weather Prediction
  Abstract

  Machine learning has recently enabled fast and accurate surrogate models for numerical weather and climate prediction. However, many current architectures struggle on extreme events, introduce spurious artifacts or instabilities when they ignore the spherical geometry of the Earth, and are often tightly coupled to a particular input grid or resolution, limiting flexibility for downstream applications. These shortcomings highlight a mismatch between black‑box models and more traditional numerical methods, which encode underlying physical and geometric properties. Neural operators address part of this gap by learning solution operators of partial differential equations in a grid‑agnostic manner, yet standard Fourier Neural Operators (FNOs) break down on the sphere, where the flat‑geometry Fourier transform induces artifacts and excessive dissipation. This talk presents geometric neural operators that respect the symmetries and topology of the sphere, with a particular focus on applications to the chaotic dynamics of Earth's atmosphere. The Spherical Fourier Neural Operator (SFNO) is formulated via Driscoll–Healy's convolution theorem, linking spherical harmonic transforms to group convolutions on the sphere and yielding rotationally equivariant, grid‑invariant models capable of stable, year‑long autoregressive rollouts with physically plausible dynamics. Building on insights from hyperbolic PDE solvers, a principled framework is introduced for incorporating localized integral operators using quadrature‑based discrete–continuous convolutions on both Euclidean and spherical geometries to capture sharp fronts and local extremes. To complement convolutional approaches, a generalized attention mechanism for spherical domains is developed, allowing Transformer‑style architectures to natively process data on the two‑dimensional sphere. This discretization‑agnostic spherical attention incorporates numerical quadrature weights into a continuous spherical formulation, producing geometrically faithful attention that is approximately rotationally equivariant. The practical impact of these ideas is demonstrated with FourCastNet 3, a scalable probabilistic weather forecasting system cast as a hidden Markov model built from spherical signal‑processing primitives. FourCastNet 3 is trained on 1000+ GPUs using flexible domain-parallelism paradigms inspired from traditional HPC methods. The resulting method matches or exceeds leading conventional ensemble systems and state‑of‑the‑art diffusion models while delivering 8–60× faster forecasts, positioning it as a robust foundation for next‑generation weather and climate modeling and downscaling systems.

Wednesday, May 20, 2026 (11h-12h) (SGS)

• Ana Stoica: Policy Design in Long-Run Welfare Dynamics
  Abstract

  Improving social welfare is a complex challenge requiring policymakers to optimize objectives across multiple time horizons. Evaluating the impact of such policies presents a fundamental challenge, as those that appear suboptimal in the short run may yield significant long-term benefits. We tackle this challenge by analyzing the long-term dynamics of two prominent policy frameworks: Rawlsian policies, which prioritize those with the greatest need, and utilitarian policies, which maximize immediate welfare gains. Conventional wisdom suggests these policies are at odds, as Rawlsian policies are assumed to come at the cost of reducing the average social welfare, which their utilitarian counterparts directly optimize. We challenge this assumption by analyzing these policies in a sequential decision-making framework where individuals' welfare levels stochastically decay over time, and policymakers can intervene to prevent this decay. Under reasonable assumptions, we prove that interventions following Rawlsian policies can outperform utilitarian policies in the long run, even when the latter dominate in the short run. We characterize the exact conditions under which Rawlsian policies can outperform utilitarian policies. Our results underscore the necessity of considering long-term horizons in designing and evaluating welfare policies; the true efficacy of even well-established policies may only emerge over time. Beyond this, our work reveals novel technical insights into the theory of stochastic processes. This is joint work with Jiduan Wu, Rediet Abebe, and Moritz Hardt, and was presented at ICLR'25.

Tuesday, May 12, 2026 (11h30-12h30) (SGS)

• Jason Altschuler: Negative Stepsizes Make Gradient-Descent-Ascent Converge
  Abstract

  Solving min-max problems is a central question in optimization, games, learning, and controls. Arguably the most natural algorithm is Gradient-Descent-Ascent (GDA), however since the 1970s, conventional wisdom has argued that it fails to converge even on simple problems. This failure spurred the extensive literature on modifying GDA with extragradients, optimism, momentum, anchoring, etc. In contrast, we show that GDA converges in its original form by simply using a judicious choice of stepsizes. The key innovation is the proposal of unconventional stepsize schedules that are time-varying, asymmetric, and (most surprisingly) periodically negative. We show that all three properties are necessary for convergence, and that altogether this enables GDA to converge on the classical counterexamples (e.g., unconstrained convex-concave problems). The core intuition is that although negative stepsizes make backward progress, they de-synchronize the min/max variables (overcoming the cycling issue of GDA) and lead to a slingshot phenomenon in which the forward progress in the other iterations is overwhelmingly larger. This results in fast overall convergence. Geometrically, the slingshot dynamics leverage the non-reversibility of gradient flow: positive/negative steps cancel to first order, yielding a second-order net movement in a new direction that leads to convergence and is otherwise impossible for GDA to move in. Joint work with Henry Shugart.

Group Meeting Thursday, April 16, 2026 (10h-12h) (SGM)

• Clément Bonet: Introduction to First-order Optimization Methods in Wasserstein Space.
  Abstract

  As the problem of minimizing functionals on the Wasserstein space encompasses many applications in machine learning, different optimization algorithms on Rd have received their counterpart analog on the Wasserstein space. We will first focus on lifting the mirror descent algorithm. This algorithm has been introduced to better capture the geometry of the function to minimize and is provably convergent under appropriate (namely relative) smoothness and convexity conditions. Adapting these notions to the Wasserstein space, we prove guarantees of convergence of some Wasserstein-gradient-based discrete-time schemes for new pairings of objective functionals and regularizers. Then, we will focus on lifting the Convex-Concave procedure which is tailored to deal with Difference-of-Convex functionals.
• Anna Korba: A Computable Measure of Suboptimality for Entropy-Regularised Variational Objectives.
  Abstract

  Several emerging post-Bayesian methods target a probability distribution for which an entropy-regularised variational objective is minimised. This increased flexibility introduces a computational challenge, as one loses access to an explicit unnormalised density for the target. To mitigate this difficulty, we introduce a novel measure of suboptimality called ‘gradient discrepancy’, and in particular a ‘kernel’ gradient discrepancy (KGD) that can be explicitly computed. In the standard Bayesian context, KGD coincides with the kernel Stein discrepancy (KSD), and we obtain a novel characterisation of KSD as measuring the size of a variational gradient. Outside this familiar setting, KGD enables novel sampling algorithms to be developed and compared, even when unnormalised densities cannot be obtained. To illustrate this point several novel algorithms are proposed and studied, including a natural generalisation of Stein variational gradient descent, with applications to mean-field neural networks and predictively oriented posteriors presented. On the theoretical side, our principal contribution is to establish sufficient conditions for desirable properties of KGD, such as continuity and convergence control. From https://arxiv.org/abs/2509.10393.
• Pierre-Cyril Aubin-Frankowski: A gradient flow for every c(x, y) cost: EVI-inspired convexity.
  Abstract

  How to go beyond the square distance d^2 in optimization algorithms and flows in metric spaces? Rooted in cross-differences, the convergence theory to the continuous flow is investigated is based on a (discrete) evolution variational inequality (EVI) which enjoys similar properties to the EVI with d^2 regularizer. This provides a theoretical framework for studying splitting schemes beyond the usual implicit Euler in gradient flows. This talk is based on the works https://arxiv.org/abs/2305.04917 with Flavien Léger (INRIA Paris), and https://arxiv.org/abs/2505.00559 (Journal of Functional Analysis 2026) with Giacomo Sodini and Ulisse Stefanelli (Uni Vienna).

Wednesday, April 15, 2026 (13h-14h) (SGS)

• Jean-Baptiste Fermanian: Class conditional conformal prediction for multiple inputs by p-value aggregation
  Abstract

  Conformal prediction methods are statistical tools designed to quantify uncertainty and generate predictive sets with guaranteed coverage probabilities. This work introduces a refinement to these methods for classification tasks, specifically tailored for scenarios where multiple observations (multi-inputs) of a single instance are available at prediction time. Our approach is particularly motivated by applications in citizen science, where multiple images of the same plant or animal are captured by individuals. Our method integrates the information from each observation into conformal prediction, enabling a reduction in the size of the predicted label set while preserving the required class-conditional coverage guarantee. The approach is based on the aggregation of conformal p-values computed from each observation of a multi-input. By exploiting the exact distribution of these p-values, we propose a general aggregation framework using an abstract scoring function, encompassing many classical statistical tools. Knowledge of this distribution also enables refined versions of standard strategies, such as majority voting. We evaluate our method on simulated and real data, with a particular focus on Pl@ntNet, a prominent citizen science platform that facilitates the collection and identification of plant species through user-submitted images.

Wednesday, April 8, 2026 (13h-14h) (SGS)

• Louis-Pierre Chaintron: ResNets of All Shapes and Sizes: Quantitative Large-Scale Theory of Training Dynamics
  Abstract

  We study the convergence of the training dynamics of residual neural networks (ResNets) towards their joint infinite depth–width limit. We focus on ResNets with two-layer perceptron blocks, whose shape is determined by the depth L, hidden width M, and embedding dimension D, and we adopt the residual scaling O(√D/√(LM) ) recently identified as necessary for local feature learning. We show that after a bounded number of training steps, the error between the finite ResNet and its infinite-size limit is O(1/L + √D/√(LM) + 1/√D), and numerical experiments suggest that this bound is tight in the early training phase. From a probabilistic viewpoint, the D→∞ limit amounts to a mean-field limit over the coordinates of the embedding, where some interactions scale in 1/√D contrary to the usual 1/D setting. Our analysis is a rigorous and quantitative instance of the Dynamical Mean Field Theory (DMFT) from statistical physics; it combines propagation of chaos arguments with the cavity method at a functional level.

Wednesday, March 25, 2026 (14h-15h) (SGS)

• Timothy Johnston: Bayesian Adversarial Privacy
  Abstract

  Theoretical and applied research into privacy encompasses an incredibly broad swathe of differing approaches, emphasis and aims. In this talk I shall discuss a new quantitative notion of privacy that is both contextual and specific. Our definition relies on concepts inherent to standard Bayesian decision theory, while departing from it in several important respects. In particular, the party controlling the release of sensitive information should make disclosure decisions from the prior viewpoint, rather than conditional on the data, even when the data is itself observed. I shall also discuss toy examples and computational methods, to highlight the specificities of our framework.

Wednesday, March 18, 2026 (14h-15h) (SGS)

• Marc Jourdan: Advances in Pure Exploration in Bandits: Non-Asymptotic and Private.
  Abstract

  In pure exploration problems for stochastic multi-armed bandits, the goal is to answer a question about a set of unknown distributions (for example, the efficacy of a treatment) by strategically sampling from them, while providing guarantees on the returned answer. The archetypal example is the best arm identification problem, where the task is to find the arm with the largest mean. Top Two algorithms, which select the next arm to sample from among a leader and a challenger, have received significant attention in recent years due to their simplicity and interpretability. In this talk, I will present recent advances on two complementary aspects of pure exploration: achieving non-asymptotic guarantees and ensuring differential privacy. First, we propose a Top Two algorithm which has an asymptotically optimal expected sample complexity, and also provides anytime guarantees on the probability of misidentifying a sufficiently good arm. Second, we show how the Top Two principle can be combined with differential privacy mechanisms, leading to algorithms that preserve near-optimal efficiency while ensuring privacy guarantees. These results not only deepen our theoretical understanding but also enable more practical and privacy-aware bandit algorithms.

Tuesday, March 17, 2026 (10h30-12h30) (SGS)

• Ye Zhu: Dynamic and Structural Sampling for Interpretable and Versatile Control in Multimodal Generation
  Abstract

  Generative models are revolutionizing daily life through applications such as image and audio synthesis, while also enabling breakthroughs in scientific discovery. Despite their huge practical successes, the interpretability of modern generative models remains relatively underexplored. In this talk, I will present one line of my recent work that investigates the intrinsic dynamics and latent geometric structures of generative models. By drawing on both probabilistic and physical perspectives, we aim to demonstrate how these insights can be harnessed during the sampling stage to guide and control pre-trained multimodal models in fine-grained scenarios. This enables versatile downstream multimodal applications, including image semantic editing, text-image guided data customization, controllable enhancement of low-level visual attributes, text-guided acoustic masking, and text-to-image diversity enhancement. Through a series of real-world applications, we hope to bridge the gap between theoretical insights and the reliable deployment of multimodal generative models in complex, real-world environments.
• Elise Bayraktar: Efficient estimation of jump parameters for stochastic differential equations driven by Lévy processes
  Abstract

  In this talk, we consider the high-frequency estimation of the jump parameters of a stochastic differential equation driven by a Lévy process. More precisely, we are interested in the efficient estimation of scaling and jump activity parameters in the presence of a Brownian motion and a jump component. We first study efficiency for the prototype Lévy process. By studying the behavior of the density of the process in small time, we prove that the LAN property holds for the joint estimation of the diffusion, scaling and jump activity parameters. We next consider a stochastic differential equation driven both by a Brownian Motion and a locally stable pure-jump Lévy process. Using a quasi-likelihood estimation method, we exhibit an estimator that attains the optimal rate of convergence previously identified.

Group Meeting Wednesday, February 24, 2026 (10h-12h) (SGM)

• Badr Moufad: When Test-Time Guidance Is Enough: Training-Free Posterior Sampling with Diffusion Priors
• Benjamin Dupuis: Score matching gap and generalization properties of diffusion models
• Aymeric Capitaine: From Prediction to Decision in Dynamic Strategic Environments

Monday, February 16, 2026 (11h-12h) (SGS)

• Eloi Tanguy: Computing Barycentres of Measures for Generic Transport Costs
  Abstract

  Wasserstein barycentres represent average distributions between multiple probability measures for the Wasserstein distance. The numerical computation of Wasserstein barycentres is notoriously challenging. A common approach is to use Sinkhorn iterations, where an entropic regularisation term is introduced to make the problem more manageable. Another approach involves using fixed-point methods, akin to those employed for computing Fréchet means on manifolds. The convergence of such methods for 2-Wasserstein barycentres, specifically with a quadratic cost function and absolutely continuous measures, was studied by Alvarez-Esteban et al.. In this paper, we delve into the main ideas behind this fixed-point method and explore how it can be generalised to accommodate more diverse transport costs and generic probability measures, thereby extending its applicability to a broader range of problems. We show convergence results for this approach and illustrate its numerical behaviour on several barycentre problems.

Group Meeting Wednesday, January 22, 2026 (10h-12h) (SGM)

• Rémi Flamary: Introduction to Optimal Transport and Gromov-Wasserstein
• Paul Krzakala: The quest for the GRAph Level autoEncoder (GRALE)
• Sonia Mazelet: Unsupervised Learning for Optimal Transport plan prediction between unbalanced graphs
• Thibaut Germain: A Spectral-Grassmann Wasserstein metric for operator representations of dynamical systems

Group Meeting Wednesday, November 12, 2025 (10h-12h) (SGM)

• Paul Mangold: Federated Learning: a Tale of Heterogeneity
• Lucas Versini: Analysis of Decentralized SGD: a Markov Chain Perspective
• Safwan Labbi: Federated Model-Based Reinforcement Learning
• Lorenzo Mancini: Personalized Federated Reinforcement Learning