The What, Why and How of Generative Flow Networks
Published in Artificial Intelligence.
GFlowNets were introduced at NeurIPS in 2021 by Emmanuel Bengio and co-authors. GFlowNets are a deep learning technique for “building objects” at a frequency proportional to the expected reward of those objects in an environment.
The motivating example in their first paper is the discovery of new chemical structures (or I’ll also refer to them as “molecules”). Chemical structures, or molecules, are “compositional” in the sense that they consist of discrete building blocks (atoms). A chemical structure can be tested for a quality of interest, such as antibiotic activity. Some molecules are strong antibiotics — when tested against bacteria in a laboratory, most bacteria die, and a measurement is obtained that we can think of as a large “reward”. Most molecules do not kill bacteria, and so, return a small “reward” value from a lab test.
https://towardsdatascience.com/the-what-why-and-how-of-generative-flow-networks-4fb3cd309af0