GFlowNets and variational inference

TL;DR

This paper explores the connections between variational inference and GFlowNets, highlighting their similarities, differences, and the advantages of GFlowNets in off-policy training and modeling diverse distributions.

Contribution

It establishes the theoretical equivalence between certain VI algorithms and GFlowNets, and discusses the practical benefits of GFlowNets over VI in training and diversity modeling.

Findings

VI algorithms are equivalent to special cases of GFlowNets in expected gradient terms.

GFlowNets are more suitable for off-policy training without high gradient variance.

GFlowNets better capture diversity in multimodal distributions.

Abstract

This paper builds bridges between two families of probabilistic algorithms: (hierarchical) variational inference (VI), which is typically used to model distributions over continuous spaces, and generative flow networks (GFlowNets), which have been used for distributions over discrete structures such as graphs. We demonstrate that, in certain cases, VI algorithms are equivalent to special cases of GFlowNets in the sense of equality of expected gradients of their learning objectives. We then point out the differences between the two families and show how these differences emerge experimentally. Notably, GFlowNets, which borrow ideas from reinforcement learning, are more amenable than VI to off-policy training without the cost of high gradient variance induced by importance sampling. We argue that this property of GFlowNets can provide advantages for capturing diversity in multimodal target distributions.