Reproduce “Convergence of flow-based generative models via proximal gradient descent in Wasserstein space” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiuyuan Cheng, Jianfeng Lu, Yixin Tan, Yao Xie

This research provides theoretical guarantees for JKO-scheme-based progressive flow models, demonstrating exponential convergence of the forward process in Wasserstein-2 distance and KL divergence. It shows that the reverse process can generate data matching the true distribution within O("")KL divergence and O("") total variation distance after a logarithmic number of steps, and quantifies the impact of inversion errors.

Key insights

  • The progressive training of JKO flow models can be rigorously understood as an approximate discrete-time Wasserstein gradient flow, offering a strong theoretical foundation.
  • High accuracy in data generation (O("") KL divergence and O("") TV distance) can be achieved with a relatively shallow network depth, specifically N steps, where N is logarithmic in the desired error (Nlog(1/")).
  • Inversion errors in the reverse process accumulate exponentially, requiring significantly tighter control ( be a much higher-order power of ") compared to per-step learning errors to maintain high-quality generated samples.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.