Generative modelling

In the wake of models like DALL·E and ChatGPT, generative models have had a massive impact on text and image applications. The goal of this class is to present their mathematical and algorithmic foundations.

Teachers for 2026

Quentin Bertrand (CR Inria, MALICE team)
Ségolène Martin (CR Inria, OCKHAM team)
Mathurin Massias (CR Inria, OCKHAM team)

For questions, write to all three teachers, using the pattern firstname.lastname@inria.fr. Many thanks to Rémi Emonet who taught the class in 2025 and created most of the current material.

Validation

5 weekly quizzes on class questions + 3 Python labs.
2 h written exam at the end of the class (on 22/10/26)
Paper presentation and extension of a selected research article and the associated code. We expect an in-depth analysis..

Prerequisites

Probability (densities, change of variable formula)
Linear algebra (PSD matrices, eigenvalue decomposition, spectral theorem)
Calculus (gradient, Hessian, Jacobian, chain rule, ordinary differential equations)

A pdf recap is available here.

Tentative Schedule

Date	Description	Material
09/09	Introduction to generative models, maximum likelihood, Mixtures of Gaussian; slides, pdf
10/09/	Bayes, (Variational) Autoencoders
16/09	Lab 1 (AE, MoG)
17/09	GAN/WGAN
23/09	Lab 2 (GAN)
24/09	Flow matching
30/09	Lab 3 (FM)
01/10	Diffusion I
07/10	Diffusion II	Arnaud Doucet invited talk at NeurIPS 2024
08/10	Model evaluation, Optimal transport
14/10	Discrete diffusion
15/10	Introduction to sequence modelling, tokenizer, autoregressive model, bigram
21/10	Attention and Transformers + Lab 4
22/10	Written exam
11/11	Project Presentations 1/2
12/11	Project Presentations 2/2