Streaming, Distributed, and Asynchronous Amortized Inference

We address the problem of amortized inference over a compositional finite space in a non-localized environment, i.e., when data are observed in a distributed, streaming, or mixed (asynchronous) fashion. This setting comprises applications in causal discovery, phylogenetic inference, natural language processing, and other problems. In particular, we focus on Generative Flow Networks (GFlowNets), an emergent family of deep generative models that cast amortized inference as finding a balanced flow assignment in a flow network. To accomplish this, a GFlowNet parameterizes the flow function as a neural network and optimizes its parameters via stochastic gradient descent. Drawing on this, we make both practical and theoretical contributions. On the practical side, we introduce three algorithms — Streaming Bayes (SB) GFlowNets, Embarrassingly Parallel (EP) GFlowNets, and Subgraph Asynchronous Learning (SAL) — along with efficient gradient estimators that significantly accelerate GFlowNet training when compared against traditional approaches. Also, we develop the first computationally amenable and sound metric for assessing the correctness of a trained GFlowNet. From a theoretical perspective, we delineate the limitations and present the first non-vacuous generalization guarantees for the learning of GFlowNets. All in all, our work paves the road for a better understanding, usability, and fair assessment of amortized inference algorithms. This extended abstract provides an overview of our research, which was published at the proceedings of ICML [da Silva et al. 2024c], NeurIPS [da Silva et al. 2024a, da Silva et al. 2024b], and ICLR [da Silva et al. 2025b, da Silva et al. 2025a].