Sivaraman Balakrishnan 在 arXiv 上发表的一篇新论文探讨了在生成模型中估计传输图的统计局限性。该研究引入了一个 minimax 框架，用于严格定义学习任何有效传输图（而不仅仅是最优传输图）的任务。该框架为扩散模型和流匹配等方法建立了样本复杂度下界，表明在标准假设下，估计任何有效传输图在统计上与估计最优传输图一样困难。该论文还强调了在某些稳定性假设不满足时，偏离最优传输图可以导致更准确的学习。 AI
影响 为理解扩散模型和流匹配等现代生成模型的统计效率提供了理论框架。
排序理由 关于生成模型理论的学术论文。[lever_c_demoted from research: ic=1 ai=1.0]
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2606.30574v1 Announce Type: cross Abstract: Many modern generative modeling methods, including diffusion models, normalizing flows, and flow matching, estimate transport maps or plans between distributions without explicitly targeting an optimal transport (OT) map. In appli…
Many modern generative modeling methods, including diffusion models, normalizing flows, and flow matching, estimate transport maps or plans between distributions without explicitly targeting an optimal transport (OT) map. In applications like generative modeling, the transport co…