Researchers establish a theoretical connection between Generative Flow Networks (GFlowNets) and optimal transport theory, demonstrating that minimum-flow GFlowNets reduce to Kantorovich optimal transport problems. This framework enables GFlowNets to learn optimal transport plans on large graphs through neural parameterization, with experimental validation confirming alignment with exact solvers.
This research bridges two important areas of machine learning by formalizing the relationship between GFlowNets and optimal transport theory. GFlowNets have emerged as a powerful framework for sampling complex structured objects through trajectory-based sampling in directed graphs, while optimal transport provides a mathematically rigorous foundation for comparing and transforming probability distributions. The key contribution is showing that when GFlowNets are constrained to minimum-flow conditions with fixed initial distributions, they naturally encode optimal transport plans, effectively transforming source distributions into target distributions with theoretical guarantees.
The theoretical insight matters because it provides interpretability to GFlowNet behavior while simultaneously offering a scalable approach to solving optimal transport problems on large graphs. Previously, computing optimal transport plans required dedicated specialized solvers that struggled with computational complexity on large-scale problems. By leveraging neural parameterization and edge flow representations, this approach enables practitioners to tackle previously intractable optimal transport instances.
For the machine learning community, this development expands GFlowNets' applicability beyond their original scope. Researchers working on structured generation, sampling problems, and distribution matching gain a mathematically principled framework with optimal transport guarantees. The experimental validation demonstrating agreement with exact solvers strengthens confidence in the approach's correctness. This work particularly benefits applications requiring efficient sampling from complex distributions over structured spaces, such as molecular generation, graph sampling, and combinatorial optimization problems where understanding the underlying transport mechanism provides both computational and theoretical advantages.
- →GFlowNets with minimum-flow constraints encode optimal transport plans with theoretical guarantees.
- →Neural parameterization enables scaling optimal transport solutions to large graphs previously unsolvable by exact methods.
- →The framework connects two previously separate mathematical areas, providing new interpretability for GFlowNet behavior.
- →Experimental validation confirms learned transport plans match exact optimal transport solvers.
- →Applications span molecular generation, graph sampling, and structured object generation problems.