Signal from Structure: Exploiting Submodular Upper Bounds in Generative Flow Networks
This addresses efficiency in generative modeling for tasks with submodular rewards, such as candidate generation, but is incremental as it builds on existing GFN methods.
The paper tackled the problem of training Generative Flow Networks (GFlowNets) more efficiently by exploiting submodular structure in rewards, resulting in SUBo-GFN generating orders of magnitude more training data than classical GFNs for the same number of reward queries.
Generative Flow Networks (GFlowNets; GFNs) are a class of generative models that learn to sample compositional objects proportionally to their a priori unknown value, their reward. We focus on the case where the reward has a specified, actionable structure, namely that it is submodular. We show submodularity can be harnessed to retrieve upper bounds on the reward of compositional objects that have not yet been observed. We provide in-depth analyses of the probability of such bounds occurring, as well as how many unobserved compositional objects can be covered by a bound. Following the Optimism in the Face of Uncertainty principle, we then introduce SUBo-GFN, which uses the submodular upper bounds to train a GFN. We show that SUBo-GFN generates orders of magnitude more training data than classical GFNs for the same number of queries to the reward function. We demonstrate the effectiveness of SUBo-GFN in terms of distribution matching and high-quality candidate generation on synthetic and real-world submodular tasks.