Nicolas Schreuder (CNRS / École normale supérieure, Paris)
-

Scientific activities
The CSD seminar programme and archive: research talks in data science and artificial intelligence.
For 2026–2027, seminars usually take place on Thursdays, 11:00–12:00. Tuesday colloquia usually take place at 11:00–12:00. Check each event for exceptions. All times are Paris local time.
-
-
-
-
-
Effective models in high dimensional unstable regimes: optimization dynamics at the edge of stability
-
-
CSD · ENS–PSL
127 talks
(Trying to) Understand Steering in Language Models
Fast coresets for the Sinkhorn divergence / Discrete Diffusion Design: An Asymptotic Error Analysis Approach
Vladimir Petrovic · Fast coresets for the Sinkhorn divergence
Damien Rouchousse · Discrete Diffusion Design: An Asymptotic Error Analysis Approach
How to Encode World Knowledge?
Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning
LLM for coding tutorial
The document marks Gabriel Peyré’s participation as provisional.
Learning To Sample From Diffusion Models Via Inverse Reinforcement Learning / Certified Per-instance Unlearning using Individual Sensitivity Bounds
Constant Bourdrez · Learning To Sample From Diffusion Models Via Inverse Reinforcement Learning
Hanna Benarroch · Certified Per-instance Unlearning using Individual Sensitivity Bounds
Learning Monge Maps with Constrained Drifting Models
How do simple rotations affect the implicit bias of Adam?
The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training
A Basin-Selection Perspective on Grokking via Singular Learning Theory
Balanced Low-Rank Adaptation: Removing Parameter Invariance to Accelerate Convergence / Convergence Rates for Distribution Matching with Sliced Optimal Transport
Valérie Castin · Balanced Low-Rank Adaptation: Removing Parameter Invariance to Accelerate Convergence
Gauthier Thurin · Convergence Rates for Distribution Matching with Sliced Optimal Transport
Beyond Uncertainty Sets: Leveraging Optimal Transport to Extend Conformal Predictive Distribution to Multivariate Settings
Conformal Prediction for Uncertainty Quantification in Machine Learning: Recent Advances
Samples That Cooperate, Samples That Remember: Two Exactly Solvable Bridge Diffusions
Training Neural Networks at Any Scale
Geometry-induced regularization and identifiability of deep ReLU networks
-
Physical Models and Machine Learning for Scientific Imaging
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
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
A Spectral-Grassmann Wasserstein metric for operator representations of dynamical systems
Identifying informative distance measures in high-dimensional feature spaces
On a Neural Implementation of Brenier's Polar Factorization / On the Relation between Rectified Flows and Optimal Transport
Nina Vesseron · On a Neural Implementation of Brenier's Polar Factorization
Johannes Hertrich · On the Relation between Rectified Flows and Optimal Transport
From Markov to Laplace: A Markovian Tale of Large Language Models
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.
-
A Tabular Foundation Model for In-Context Learning on Large Data
Learning to Answer from Correct Demonstrations
Phase Diagram of Dropout for Two-Layer Neural Networks in the Mean-Field Regime
Moment Guided Diffusion for Maximum Entropy Generation / Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training
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
Optimization Perspectives on the Differentiation of Programs
Flow matching for (generalized) Schrödinger bridges and dynamical optimal transport
-
-
-
-
Large-scale pretraining on neural data allows for transfer across individuals, tasks and species
-
-
-
Computing Barycentres of Measures for Generic Transport Costs
Learning under requirements
-
-
-
-
-
-
Modern Transformers and GRPO
-
Ten years of playing in deep haystacks: a taxonomy of interpretability research
-
Projet Numina: IA appliquée au raisonnement et à la formalisation mathématiques
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Fairness in machine learning : a study of the Demographic Parity constraint
The document’s “Nicolas Scwhreuder” spelling was reconciled with the matching date and full name in the old public archive.
-
-
The random assignment problem
-
Arrows of Time for Large Language Models
Learning hierarchical representations with deep nets and large language models
Variational Classification
Lecture on neural operators
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.
-
Lecture on neural operators
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.
The two timescale regime for the analysis of neural networks
Understanding Complex Procedural Videos
On a Neural Implementation of Brenier's Polar Factorization / Sparsistency for Inverse Optimal Transport
Nina Vesseron · On a Neural Implementation of Brenier's Polar Factorization
Francisco Andrade · Sparsistency for Inverse Optimal Transport
Slicing Mutual Information Generalization Bounds for Neural Networks
-
The document records the speaker as “M Jordan”; the name has not been expanded.
-
-
Only the speaker’s surname or first name “Malick” is recorded in the document.
-
Computational Imaging: Restoration Deep Networks as Image Priors
-
How do neural nets learn?
-
-
Two mathematical perspectives on Transformers
Advances in the design and analysis of neural operators
An alternative view of denoising diffusion models
Backtracking Dynamical Cavity method
-
Tutorial on Autodiff
Only the speaker’s first name is recorded in the document.
Blue Noise Sampling
Are ensembles getting better All the time ?
High-dimensional estimation under heavy-tails, with applications to robust regression
The document’s “Leonardo Defillipis” spelling was reconciled with its other entries and the public archive.
Fitting ellipsoids to random points
Learning Physics with Wavelet Renormalisation Group / Differential Privacy in Decentralized Learning
Etienne Lempereur · Learning Physics with Wavelet Renormalisation Group
Edwige Cyffers · Differential Privacy in Decentralized Learning
-
Sampling via diffusion processes : rigorous guarantees, hardness, disorder chaos
-
-
Theory of ML journal club
Machine learning and invariant theory
Differentiating Nonsmooth Solutions to Parametric Monotone Inclusion Problems
Deep Generative Models for Hybrid Physical/Data-Driven Bayesian Inference
Training neural networks with orthogonal weights
Flamingo : a Visual Language Model for Few-Shot Learning
Data-Augmentation in High Dimensional Low Sample Size Setting Using a Geometry-Based Variational Autoencoder
Diffusion / Field based Reconstruction and Generation for 3D Shapes
Embeddings to learn on messy relational data
Learning image representations with coarse, instance-level and image-level supervision
Deep learning for symbolic maths
Date retained exactly as published. This entry appears after December 2021 in the source list; its year may need archival verification.
What makes data learnable by deep learning
Date retained exactly as published. This entry appears after December 2021 in the source list; its year may need archival verification.
Restricted Boltzmann Machines revisited : from sampling to design
Date retained exactly as published. This entry appears after December 2021 in the source list; its year may need archival verification.
Phase diagram of Stochastic Gradient Descent in high-dimensional two-layer neural networks
Sampling based analysis of neural network generalization and extrapolation
The quest for adaptivity
Low-resource MT : few-shot learning and historical language normalisation
Convolutional Sparse Coding for Electromagnetic Brain Signals
Enhancing Sampling with Learning
A Semi-Empirical Model for the Generalization Capacity of Deep Neural Networks
Exploiting Graph Invariants in Deep Learning
The interplay of data structure and learning dynamics in simple neural networks
globality, consistency and compositionality in neural machine translation
Historical entries retain the information published in the old archive. Missing titles, times, and rooms have not been reconstructed.
Download the seminar database (JSON)Receive announcements of seminars, courses, and events at the centre.