GFlowNets: Sampling as sequential decision making - IFT 6760B A26

Course description

Generative flow networks, also known as GFlowNets or simply GFN, are a class of generative machine learning models that perform amortised probabilistic inference and are trained to sample from unnormalised distributions as sequential decision making. This course will cover the fundamental aspects of GFlowNets, starting from a motivation and introduction to the method, and progressing towards more advanced concepts and applications, as well as the connection to other generative models, reinforcement learning, sampling and probabilistic inference methods. The course will combine theory with project work, with a special emphasis on the application of GFlowNets for scientific discovery.

Les réseaux de flux génératifs, également connus sous le nom de GFlowNets ou simplement GFN, sont une classe de modèles d’apprentissage automatique génératifs, basée sur l’apprentissage par renforcement, qui permettent l’échantillonnage et l’inférence probabiliste amortie à travers la prise de décision séquentielle. Ce cours couvre les aspects fondamentaux de GFlowNets, à partir d’une motivation et d’une introduction à la méthode, en progressant vers des concepts et applications plus avancés, ainsi que le lien avec d’autres modèles génératifs, apprentissage par renforcement et méthodes d’inférence probabiliste. Le cours combine la théorie avec le travail de projet, avec un accent particulier sur l’application des GFlowNets pour la découverte scientifique.

Course outline

This seminar course consists of three blocks:

  1. Lectures by the instructor and invited speakers
  2. Presentations by students about relevant papers
  3. Project work and presentations by students

The lectures will cover the following topics:

  • Introduction and motivation
  • Brief review of requisite background
  • Context and original formulation of GFlowNets
  • Main concepts and theoretical results
  • Modern mathematical re-formulations
  • Relevant loss functions
  • Continuous GFlowNets
  • Training and evaluation guidelines
  • Connections with reinforcement learning
  • Connections with diffusion models and variational inference
  • Multi-objective GFlowNets
  • Conditional GFlowNets
  • Active learning with GFlowNets
  • Applications in drug discovery
  • Applications in materials discovery

Evaluation criteria

The evaluation will be based on four aspects:

  • Project work: 30 %
    • Students will form teams and work on research-like projects.
    • Projects may focus on extending, analysing or reproducing theoretical or practical aspects of GFlowNets.
    • Projects will be evaluated based on a conference-like paper, a presentation and possibly personal interviews.
  • Paper presentations: 30 %
    • Students will select a relevant paper and present it for the rest of the group either individually or in small teams.
    • Presentations will be followed by a discussion in which everyone can participate.
    • The evaluation will consider both the presentation as well as the participation in the discussions.
  • Quizzes: 30 %
    • A few quizzes will have to be completed by students throughout the lectures block of the course.
    • The quizzes will be based on the content of the lectures and suggested additional material.
  • Short literature review: 10 %
    • Students will perform a literature review on a particular GFlowNet aspect or related topic.
    • The literature review will be summarised into a 1–2 pages report, which may be reused for the project paper too.

Prerequisites

As a prerequisite to register for this course, students must have successfully completed the following courses:

Additionally, it is recommended to have taken (or take in parallel) the following courses:

Comme exigences d’inscription à ce cours, les étudiant.e.s doivent avoir réussi :

En outre, il est recommandé d’avoir suivi (ou de suivre en parallèle) les cours suivants :

Why such prerequisites?

In order to cover the specifics of generative flow networks (GFlowNets), we have to depart from a position of familiarity with the fundamental concepts of machine learning as well as deep learning. Additionally, we will be using the language of probability, statistics, linear algebra and calculus. A good reference for the contents that are expected to be familiar with is Part I of the Deep Learning book, by Goodfellow, Bengio and Courville. Finally, the course includes project work, which requires familiarity with Python and the typical machine learning libraries, such numpy, pandas, PyTorch, etc. You can check the resources section below for additional references.

Policy about the use of AI agents and chatbots

The use of AI agents, assistants and chatbots is generally discouraged for the development of the course work, and it is explicitly not allowed for certain tasks such as generating the text of the written assignments and non-negligible amounts of code.

The details of the policy can be found in a dedicated page that describes what is allowed and what is not, as well as additional motivation for the policy and bibliographic references.

Useful links

Session d’automne 2026

Les cours ont lieu :

  • Les lundis, 15h30–17h15 (ET)
  • Les jeudis, 10h30–12h15 (ET)

Resources

Introduction to GFlowNets

Machine learning and deep learning review

Machine learning in practice