Neural Sampling with Reweighted Normalizing Flows via the Wasserstein--Fisher--Rao JKO Scheme

Key points

• Our approach is based on the Jordan--Kinderlehrer--Otto scheme for the Kullback--Leibler divergence in the Wasserstein--Fisher--Rao geometry (WFR JKO scheme).1
• First, we prove that, for any fixed step size, the exact WFR JKO iterates converge exponentially fast to the target as the number of iterations tends to infinity.1
• Second, we develop a neural implementation of the WFR JKO scheme that parametrizes its transport and reaction components using reweighted normalizing flows.1
• Numerical experiments on challenging multimodal targets demonstrate the promising performance of the proposed method.1

Sources (1)

• [1]Neural Sampling with Reweighted Normalizing Flows via the Wasserstein--Fisher--Rao JKO Scheme
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 03:44 PM

  We propose a neural algorithm for sampling from distributions specified by unnormalized Boltzmann densities.

  Our approach is based on the Jordan--Kinderlehrer--Otto scheme for the Kullback--Leibler divergence in the Wasserstein--Fisher--Rao geometry (WFR JKO scheme).

Extractive summary: sentences quoted from the sources.