Unifying Generative Models with GFlowNets and Beyond

TMLR Paper1505 Authors

Abstract:

There are many frameworks for generative modeling, each often presented with their own specific training algorithms and inference methods.
Here, we demonstrate the connections between existing generative models and the recently introduced GFlowNet framework (Bengio, et al.), a probabilistic inference machine which treats sampling as a decision-making process. This analysis sheds light on their overlapping traits and provides a unifying viewpoint through the lens of learning with Markovian trajectories.
Our framework provides a means for unifying training and inference algorithms, and provides a route to shine a unifying light over many generative models.
Beyond this, we provide a practical and experimentally verified recipe for improving generative modeling with insights from the GFlowNet perspective.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

Major update includes
- rephrase the way of presentation of our main claims in abstract, introduction, and conclusion to avoid overclaim
- improve the writing of GFlowNet backgrounds for more general readers
- improve the two setting descriptions (which was originally in Section 4) and move to Section 2 for better logical soundness
- add discussion about algorithms / propositions on multiple places in Section 3
- add experimental details in the experiment section (previous Section 5, now Section 4)

All changes are marked in blue fonts.

Assigned Action Editor: ~antonio_vergari2

Submission Number: 1505