Activités scientifiques

Séminaires

Le programme et les archives du séminaire du CSD : des exposés de recherche en sciences des données et en intelligence artificielle.

La salle Dussane à l’ENS, Paris

Pour l’année 2026–2027, les séminaires ont habituellement lieu le jeudi de 11 h à 12 h. Les colloquiums du mardi ont habituellement lieu de 11 h à 12 h. Consultez chaque événement pour les exceptions. Tous les horaires sont à l’heure de Paris.

Prochains séminaires

11:00 · Room W (salle par défaut)
À venirÀ confirmerSéminaire

Nicolas Schreuder (CNRS / École normale supérieure, Paris)

-

11:00 · Room W (salle par défaut)
À venirSéminaire

Jerome Garnier Brun (Bocconi University)

-

11:00 · Room W (salle par défaut)
À venirColloquium

Marc Lelarge (Inria / École normale supérieure, Paris)

-

11:00 · Room W (salle par défaut)
À venirSéminaire

Xiaoming Yuan (University of Hong Kong)

-

11:00 · Room W (salle par défaut)
À venirColloquium

Pierre Gentine (Columbia University)

-

11:00 · Room W (salle par défaut)
À venirÀ confirmerSéminaire

Antonin de Courcel (École normale supérieure, Paris)

Effective models in high dimensional unstable regimes: optimization dynamics at the edge of stability

11:00 · Room W (salle par défaut)
À venirColloquium

Alexander Mathis (EPFL)

-

14:00 · Room W (salle par défaut)
À venirÀ confirmerSéminaire

Steven L. Brunton (University of Washington)

-

CSD · ENS–PSL

Archives des séminaires

Connexion organisateur

127 exposés

11:00
Séminaire

Samuel Vaiter (CNRS / Université Côte d’Azur)

(Trying to) Understand Steering in Language Models

11:00
Groupe de travail

Vladimir Petrovic; Damien Rouchousse

Fast coresets for the Sinkhorn divergence / Discrete Diffusion Design: An Asymptotic Error Analysis Approach

Exposés de cette séance

Vladimir Petrovic · Fast coresets for the Sinkhorn divergence

Damien Rouchousse · Discrete Diffusion Design: An Asymptotic Error Analysis Approach

11:00
Groupe de travail

Guy Gilboa (Technion – Israel Institute of Technology)

How to Encode World Knowledge?

12:00
Séminaire

Florent Krzakala (EPFL)

Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning

12:00
Séminaire

Emmanuel Sérié; Gabriel Peyré

LLM for coding tutorial

Note sur les archives

The document marks Gabriel Peyré’s participation as provisional.

12:00
Séminaire

Constant Bourdrez; Hanna Benarroch (École normale supérieure, Paris)

Learning To Sample From Diffusion Models Via Inverse Reinforcement Learning / Certified Per-instance Unlearning using Individual Sensitivity Bounds

Exposés de cette séance

Constant Bourdrez · Learning To Sample From Diffusion Models Via Inverse Reinforcement Learning

Hanna Benarroch · Certified Per-instance Unlearning using Individual Sensitivity Bounds

12:00
Séminaire

Théo Lacombe (Université Gustave Eiffel)

Learning Monge Maps with Constrained Drifting Models

Colloquium

Rebecca Willet (University of Chicago)

How do simple rotations affect the implicit bias of Adam?

Séminaire

Fabian Schaipp (Inria / École normale supérieure, Paris)

The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training

Séminaire

Jiayi Li (Max Planck Institute of Molecular Cell Biology and Genetics)

A Basin-Selection Perspective on Grokking via Singular Learning Theory

Séminaire

Valérie Castin; Gauthier Thurin (École normale supérieure, Paris)

Balanced Low-Rank Adaptation: Removing Parameter Invariance to Accelerate Convergence / Convergence Rates for Distribution Matching with Sliced Optimal Transport

Exposés de cette séance

Valérie Castin · Balanced Low-Rank Adaptation: Removing Parameter Invariance to Accelerate Convergence

Gauthier Thurin · Convergence Rates for Distribution Matching with Sliced Optimal Transport

AnnuléSéminaire

Eugène Ndiaye (Apple)

Beyond Uncertainty Sets: Leveraging Optimal Transport to Extend Conformal Predictive Distribution to Multivariate Settings

Séminaire

Batiste Le bars (Inria Lille)

Conformal Prediction for Uncertainty Quantification in Machine Learning: Recent Advances

Colloquium

Michael Chertkov (University of Arizona)

Samples That Cooperate, Samples That Remember: Two Exactly Solvable Bridge Diffusions

Séminaire

Tony Silveti-Falls (CentraleSupélec / Université Paris-Saclay)

Training Neural Networks at Any Scale

Séminaire

François Malgouyres (Institut de Mathématiques de Toulouse)

Geometry-induced regularization and identifiability of deep ReLU networks

Séminaire

Eugène Ndiaye (Apple)

-

Colloquium

Julien Mairal (Inria Grenoble)

Physical Models and Machine Learning for Scientific Imaging

Groupe de travail

Leonardo Defilippis; Luigi Fogliani (École normale supérieure, Paris)

Scaling Laws and Spectra of Shallow Neural Networks in the Feature Learning Regime / Annealing in variational inference mitigates mode collapse: A theoretical study on Gaussian mixtures

Exposés de cette séance

Leonardo Defilippis · Scaling Laws and Spectra of Shallow Neural Networks in the Feature Learning Regime

Luigi Fogliani · Annealing in variational inference mitigates mode collapse: A theoretical study on Gaussian mixtures

Séminaire

Thibaut Germain (École polytechnique)

A Spectral-Grassmann Wasserstein metric for operator representations of dynamical systems

Colloquium

Alessandro Laio (SISSA)

Identifying informative distance measures in high-dimensional feature spaces

Groupe de travail

Nina Vesseron; Johannes Hertrich (Nina Vesseron: CREST / ENSAE Paris; Johannes Hertrich: Université Paris Dauphine-PSL / Inria / ENS)

On a Neural Implementation of Brenier's Polar Factorization / On the Relation between Rectified Flows and Optimal Transport

Exposés de cette séance

Nina Vesseron · On a Neural Implementation of Brenier's Polar Factorization

Johannes Hertrich · On the Relation between Rectified Flows and Optimal Transport

Séminaire

Ashok Vardhan (Télécom Paris)

From Markov to Laplace: A Markovian Tale of Large Language Models

Résumé

Large Language Models based on transformers and state-space models (Mamba) are at the forefront of recent breakthroughs in a variety of disciplines, including natural languages. Despite their remarkable empirical success, our theoretical understanding of how these models learn and represent sequential structure, and how in-context learning (ICL) capabilities emerge on such tasks, remains limited. This raises a fundamental question: How do modern sequence models learn from sequential data? To address this question, in this talk I will present our new framework for a principled theoretical and empirical analysis of LLMs via Markov chains. The key idea underpinning our approach is the modeling of sequential input data as a Markov process, inspired by the Markovianity of natural languages. We utilize this framework to systematically study the interplay between the Markov order and model depth, revealing sharp contrasts between transformers and state-space models. In particular, our analysis reveals the curious phenomena that (i) even a single-layer Mamba efficiently learns the in-context Laplacian smoothing estimator, whereas transformers require two layers, and (ii) while a single layer transformer cannot efficiently represent a conditional $k$-gram model, a two layer transformer can represent it for any Markov order $k \geq 1$. Together, our results provide the first formal connection between Mamba and optimal statistical estimators, and yield the tightest known characterization of the interplay between transformer depth and Markov order for ICL. Deepening our understanding of LLMs and their ICL capabilities, we believe our framework provides a new avenue for a principled study of LLMs with plenty of interesting open questions abound, which I will discuss in the end.

Groupe de travail

Eddie Aamari (CNRS / École normale supérieure, Paris)

-

Séminaire

Marine Le Morvan (Inria)

A Tabular Foundation Model for In-Context Learning on Large Data

Colloquium

Nathan Srebro (Toyota Technological Institute at Chicago)

Learning to Answer from Correct Demonstrations

Séminaire

Pierre Marion (EPFL)

Phase Diagram of Dropout for Two-Layer Neural Networks in the Mean-Field Regime

Groupe de travail

Etienne Lempereur; Raphaël Urfin

Moment Guided Diffusion for Maximum Entropy Generation / Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training

Exposés de cette séance

Etienne Lempereur · Moment Guided Diffusion for Maximum Entropy Generation

Raphaël Urfin · Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training

Séminaire

Jérôme Bolte (Toulouse School of Economics / École normale supérieure, Paris)

Optimization Perspectives on the Differentiation of Programs

Séminaire

Stephen Zhang (University of Melbourne)

Flow matching for (generalized) Schrödinger bridges and dynamical optimal transport

Colloquium

Étienne Ollion (CNRS / École polytechnique)

-

Séminaire

Hugo Cui (Harvard University)

-

Groupe de travail

Samuel Boïté; Léa Bohbot

-

Groupe de travail

Arie Wortsman; Benjamin Dupuis

-

Colloquium

Eva Dyer (University of Pennsylvania)

Large-scale pretraining on neural data allows for transfer across individuals, tasks and species

Groupe de travail

Arthur Pellegrino; Nikos Tsilivis

-

Séminaire

Emile Pierret (Université Paris Cité / École normale supérieure, Paris)

-

Groupe de travail

Davide Carbone; Nathanael Cuvelle-Magar

-

Séminaire

Julie Delon (École normale supérieure, Paris)

Computing Barycentres of Measures for Generic Transport Costs

Séminaire

Luiz Chamon (École polytechnique)

Learning under requirements

Groupe de travail

Kruno; Luca del Bono

-

Colloquium

David Louapre (Ubisoft)

-

Colloquium

Eric Vanden-Eijnden (New York University)

-

Séminaire

Grant Rotskoff (Stanford University)

-

Séminaire

Loucas-Pillaud Vivien (École des Ponts ParisTech)

-

Groupe de travail

Beatrice Achilli (Bocconi University)

-

Groupe de travail

Olivier Cappé; Gabriel Peyré (CNRS / École normale supérieure, Paris)

Modern Transformers and GRPO

Séminaire

Denny Wu (New York University / Flatiron Institute)

-

Séminaire

Demba Ba (Harvard University)

Ten years of playing in deep haystacks: a taxonomy of interpretability research

Séminaire

Alex Cayco Gajic (École normale supérieure, Paris)

-

Séminaire

Yann Fleureau (Project Numina)

Projet Numina: IA appliquée au raisonnement et à la formalisation mathématiques

Séminaire

Antonio Ocello (École polytechnique)

-

Séminaire

Clément Bonet (CREST / ENSAE Paris)

-

Colloquium

Fanny Yang (ETH Zürich)

-

Séminaire

Maxence Noble (École polytechnique)

-

Séminaire

Alessandro de Palma (Inria Paris)

-

Colloquium

Michele Ceriotti (EPFL)

-

Groupe de travail

Antonin Brossollet; Etienne Lempereur

-

Groupe de travail

Gauthier Thurin; Leonardo Defilippis (CNRS / École normale supérieure, Paris)

-

Groupe de travail

Olivier Cappé (CNRS / École normale supérieure, Paris)

-

Groupe de travail

Emanuele Francazi; Alexis Aymé

-

Colloquium

Jean-Remi King (CNRS / École normale supérieure, Paris / Meta AI)

-

Groupe de travail

Carl Allen (École normale supérieure, Paris)

-

Séminaire

Elisabetta Cornacchia (Inria / École normale supérieure, Paris)

-

Colloquium

Luca Biferale (University of Rome Tor Vergata)

-

Séminaire

Nicolas Schreuder (LIGM)

Fairness in machine learning : a study of the Demographic Parity constraint

Note sur les archives

The document’s “Nicolas Scwhreuder” spelling was reconciled with the matching date and full name in the old public archive.

Groupe de travail

Louis Grenioux; Christoph Schönle

-

Groupe de travail

Alexandre Vérine; Jérémie Kalfon

-

Séminaire

Gabriele Sicuro (Bologna)

The random assignment problem

Groupe de travail

Tony Bonnaire (École normale supérieure, Paris / Université PSL)

-

Séminaire

Vassilis Papadopoulos (EPFL)

Arrows of Time for Large Language Models

Séminaire

Matthieu Wyart (EPFL)

Learning hierarchical representations with deep nets and large language models

Séminaire

Carl Allen (ENS)

Variational Classification

Groupe de travail

Maarten de Hoop (Rice University)

Lecture on neural operators

Note sur les archives

The planning document records two lectures, 20 June and 2 July. The speaker’s “Martin” spelling is reconciled with the document’s Maarten de Hoop entry and the CSD visiting-professor list.

Colloquium

Andrea Montanari (Stanford University)

-

Groupe de travail

Maarten de Hoop (RICE)

Lecture on neural operators

Note sur les archives

The planning document records two lectures, 20 June and 2 July. The speaker’s “Martin” spelling is reconciled with the document’s Maarten de Hoop entry and the CSD visiting-professor list.

Séminaire

Raphael Berthier (Sorbonne)

The two timescale regime for the analysis of neural networks

Séminaire

Ehsan Elhamifar (Northeastern)

Understanding Complex Procedural Videos

Groupe de travail

Nina Vesseron; Francisco Andrade

On a Neural Implementation of Brenier's Polar Factorization / Sparsistency for Inverse Optimal Transport

Exposés de cette séance

Nina Vesseron · On a Neural Implementation of Brenier's Polar Factorization

Francisco Andrade · Sparsistency for Inverse Optimal Transport

Séminaire

Kimia Nadjahi (ENS)

Slicing Mutual Information Generalization Bounds for Neural Networks

Colloquium

M. Jordan

-

Note sur les archives

The document records the speaker as “M Jordan”; the name has not been expanded.

Groupe de travail

Kimia Nadjahi; Samuel Hurault

-

Séminaire

Malick

-

Note sur les archives

Only the speaker’s surname or first name “Malick” is recorded in the document.

Colloquium

Lenaic Chizat (EPFL)

-

Séminaire

Ulugbek Kamilov (Washington University in St. Louis)

Computational Imaging: Restoration Deep Networks as Image Priors

Colloquium

Antoine Georges (Collège de France / Flatiron Institute)

-

12:00
Groupe de travail

Florent Krzakala (EPFL)

How do neural nets learn?

Séminaire

Aymeric Dieuleveut (École polytechnique)

-

Colloquium

Noah Smith (University of Washington)

-

Séminaire

Cyril Letrouit (Orsay)

Two mathematical perspectives on Transformers

Groupe de travail

Maarten de Hoop (Rice)

Advances in the design and analysis of neural operators

Séminaire

Francis Bach (INRIA)

An alternative view of denoising diffusion models

Séminaire

Lenka Zdeborová (EPFL)

Backtracking Dynamical Cavity method

Groupe de travail

Lenka Zdeborová (EPFL); Backtracking Dynamical Cavity method

-

Groupe de travail

Gabriel

Tutorial on Autodiff

Note sur les archives

Only the speaker’s first name is recorded in the document.

Séminaire

Mathieu Desbrun (Inria / École polytechnique)

Blue Noise Sampling

Séminaire

Pierre-Alexandre Mattei (Inria Sophia Antipolis)

Are ensembles getting better All the time ?

Groupe de travail

Leonardo Defilippis (CNRS / École normale supérieure, Paris)

High-dimensional estimation under heavy-tails, with applications to robust regression

Note sur les archives

The document’s “Leonardo Defillipis” spelling was reconciled with its other entries and the public archive.

Séminaire

Antoine Maillard (ETH Zürich)

Fitting ellipsoids to random points

Groupe de travail

Etienne Lempereur; Edwige Cyffers

Learning Physics with Wavelet Renormalisation Group / Differential Privacy in Decentralized Learning

Exposés de cette séance

Etienne Lempereur · Learning Physics with Wavelet Renormalisation Group

Edwige Cyffers · Differential Privacy in Decentralized Learning

Groupe de travail

Nathanael Cuvelle-Magar; Raja Giryes

-

12:30
Séminaire

Andrea Montanari (Stanford University)

Sampling via diffusion processes : rigorous guarantees, hardness, disorder chaos

Groupe de travail

Michael Sander (CNRS / École normale supérieure, Paris)

-

Colloquium

Julia Kempe (New York University)

-

Groupe de travail

Valérie Castin

Theory of ML journal club

Séminaire

Soledad Villar (Johns Hopkins University)

Machine learning and invariant theory

Séminaire

Antonio Silveti-Falls (CentraleSupélec / Université Paris-Saclay)

Differentiating Nonsmooth Solutions to Parametric Monotone Inclusion Problems

Séminaire

Francois Lanusse (CNRS / CEA Paris-Saclay)

Deep Generative Models for Hybrid Physical/Data-Driven Bayesian Inference

Séminaire

Pierre Ablin (Apple)

Training neural networks with orthogonal weights

Séminaire

Arthur Mensch

Flamingo : a Visual Language Model for Few-Shot Learning

Séminaire

Stéphanie Allassonnière (Université Paris Cité / École polytechnique)

Data-Augmentation in High Dimensional Low Sample Size Setting Using a Geometry-Based Variational Autoencoder

Séminaire

James Thornton

Diffusion / Field based Reconstruction and Generation for 3D Shapes

Séminaire

Gaël Varoquaux (Inria)

Embeddings to learn on messy relational data

Séminaire

Hervé Jégou

Learning image representations with coarse, instance-level and image-level supervision

Séminaire

François Charton (Meta AI / FAIR Paris)

Deep learning for symbolic maths

Note sur les archives

Date retained exactly as published. This entry appears after December 2021 in the source list; its year may need archival verification.

Séminaire

Mathieu Wyart (EPFL)

What makes data learnable by deep learning

Note sur les archives

Date retained exactly as published. This entry appears after December 2021 in the source list; its year may need archival verification.

Séminaire

Rémi Monasson (CNRS / École normale supérieure, Paris)

Restricted Boltzmann Machines revisited : from sampling to design

Note sur les archives

Date retained exactly as published. This entry appears after December 2021 in the source list; its year may need archival verification.

Séminaire

Bruno Loureiro (CNRS / École normale supérieure, Paris)

Phase diagram of Stochastic Gradient Descent in high-dimensional two-layer neural networks

Séminaire

Raja Giryes (Tel Aviv University)

Sampling based analysis of neural network generalization and extrapolation

Séminaire

Francis Bach (Inria / École normale supérieure, Paris)

The quest for adaptivity

Séminaire

Rachel Bawden (Inria)

Low-resource MT : few-shot learning and historical language normalisation

Séminaire

Thomas Moreau (Inria Saclay)

Convolutional Sparse Coding for Electromagnetic Brain Signals

Séminaire

Marylou Gabrié (École polytechnique)

Enhancing Sampling with Learning

Séminaire

Charles Martin (Calculation Consulting)

A Semi-Empirical Model for the Generalization Capacity of Deep Neural Networks

Séminaire

Marc Lelarge (Inria / École normale supérieure, Paris)

Exploiting Graph Invariants in Deep Learning

Séminaire

Sebastian Goldt (SISSA)

The interplay of data structure and learning dynamics in simple neural networks

Séminaire

Dieuwke Hupkes - On locality (Facebook AI Research)

globality, consistency and compositionality in neural machine translation

Les archives conservent les informations publiées sur l’ancien site. Les titres, horaires et salles manquants n’ont pas été reconstitués.

Télécharger la base des séminaires (JSON)

Liste de diffusion du CSD

Recevez les annonces de séminaires, cours et événements du centre.

Rejoindre la liste de diffusion