Generative Flow Networks: Theory and Applications to Structure Learning
Tristan Deleu · 2025
Read Generative Flow Networks: Theory and Applications to Structure Learning by Tristan Deleu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-05498.txt.
Introduction
Tristan Deleu's thesis tackles a core problem in causal learning: many different causal models can explain the same observations equally well. Picking one at random risks unsafe decisions if reality differs. The Bayesian answer keeps a spread of belief over candidate models, shown as directed graphs without cycles.
The thesis introduces Generative Flow Networks, models that build complex objects piece by piece while sampling them in proportion to their reward. The first half lays out their math and their links to variational inference and reinforcement learning. The second half shows they can approximate the belief spread over causal graphs, and over the settings of each graph's mechanisms, from both passive and experimental data.
Catalog record
| CREATOR | Tristan Deleu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 740 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-05498