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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Target Distribution Sampling | Funnel 10D | Sinkhorn Distance127.6 | 29 | |
| Sampling on discretised synthetic densities | Manywell d = 32 | Sinkhorn Dist.29.57 | 15 | |
| Amortised Sampling | MoS d = 50 | Sinkhorn Cost2.13e+3 | 13 | |
| Amortised Sampling | GMM40 d = 50 | Sinkhorn Distance3.90e+3 | 12 | |
| Amortised Sampling | Robot4 d = 10 | Sinkhorn Distance1.72 | 12 | |
| Amortised Sampling | ManyWell d = 64 | MMD0.243 | 10 | |
| Amortised Sampling | GMM40 d = 2 | Sinkhorn Distance607.3 | 7 | |
| Amortised Sampling | GMM40 d=5 | Sinkhorn Distance3.11e+3 | 7 | |
| biological sequence design | TFbind8 | ELBO12.272 | 6 | |
| Chemical sequence design | QM9 | ELBO21.591 | 6 |