Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

GFlowNets and variational inference

About

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.

Nikolay Malkin, Salem Lahlou, Tristan Deleu, Xu Ji, Edward Hu, Katie Everett, Dinghuai Zhang, Yoshua Bengio• 2022

Related benchmarks

TaskDatasetResultRank
Target Distribution SamplingFunnel 10D
Sinkhorn Distance127.6
29
Sampling on discretised synthetic densitiesManywell d = 32
Sinkhorn Dist.29.57
15
Amortised SamplingMoS d = 50
Sinkhorn Cost2.13e+3
13
Amortised SamplingGMM40 d = 50
Sinkhorn Distance3.90e+3
12
Amortised SamplingRobot4 d = 10
Sinkhorn Distance1.72
12
Amortised SamplingManyWell d = 64
MMD0.243
10
Amortised SamplingGMM40 d = 2
Sinkhorn Distance607.3
7
Amortised SamplingGMM40 d=5
Sinkhorn Distance3.11e+3
7
biological sequence designTFbind8
ELBO12.272
6
Chemical sequence designQM9
ELBO21.591
6
Showing 10 of 12 rows

Other info

Follow for update