利用子模结构提升生成模型采样效率，大幅减少奖励函数调用次数。
Signal from Structure: Exploiting Submodular Upper Bounds in Generative Flow Networks
- 基于子模性构造未观测对象的奖励上界，指导高效采样。
- 相同查询次数下，生成数据量比传统方法多一个数量级。
- 适合需要高样本效率的组合优化与生成任务，如结构化数据生成。
生成流网络（GFlowNets, GFNs）是一类学习按未知奖励值比例采样组合对象的生成模型。本文聚焦于奖励具有可行动结构的情形——即奖励为子模函数。我们证明子模性可用于推导尚未观测到的组合对象的奖励上界。深入分析了此类上界出现的概率及可覆盖的未观测对象数量。基于‘面对不确定性保持乐观’原则，提出SUBo-GFN，利用子模上界训练GFN。实验表明，相同奖励函数查询次数下，SUBo-GFN生成的数据量比经典GFN高出一个数量级。在合成与真实世界的子模任务中，该方法在分布匹配和高质量候选生成方面均表现出色。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。