GFlowNets and variational inference
This work bridges probabilistic algorithms for continuous and discrete spaces, potentially improving training efficiency and diversity in multimodal modeling, though it is incremental in linking existing methods.
The paper connects variational inference (VI) and generative flow networks (GFlowNets), showing that VI algorithms are equivalent to special cases of GFlowNets in terms of gradient expectations, and demonstrates experimentally that GFlowNets offer advantages like lower gradient variance and better diversity capture in multimodal distributions.
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.