Particle GFlowNets: Rethinking Generative Marginalization Models
This research provides a new perspective on generative models, potentially leading to faster and more efficient training for complex tasks.
This paper introduces Particle GFlowNets, a new model that rethinks generative marginalization models. It shows that MaMs are equivalent to GFlowNets and extends MaMs' sampling strategy to non-autoregressive processes. An automatic criterion for full-state rejuvenation is described, derived from the Gelman-Rubin statistic, which speeds up learning convergence. Experiments show the method markedly accelerates training in large combinatorial spaces.
Generative Marginalization Models (MaMs) have been recently introduced as efficient neural sampling models for any-order autoregressive modelling of discrete distributions. By learning both the marginal and conditional probabilities of a persistent-block Gibbs sampler, MaMs enable fast posterior evaluation with a single neural network forward pass. While prior work has considered MaMs to be distinct from Generative Flow Networks (GFlowNets), a well-established paradigm for inference in discrete stochastic models, we show that they are equivalent. Then, we also extend MaMs' sampling strategy to non-autoregressive generative processes. In particular, we describe an automatic criterion for full-state rejuvenation of the Gibbs sampler, derived from the Gelman-Rubin statistic, which plays a key role in speeding up learning convergence. Our experiments show that our method, called Particle GFlowNets, markedly accelerates training in large combinatorial spaces.