Reproduce “Maximum Mean Discrepancy Gradient Flow” 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

Michael Arbel, Anna Korba, Adil Salim, Arthur Gretton

The authors analyze maximum mean discrepancy gradient flow, which moves particles toward a target using target samples rather than its density. The method evaluates the MMD witness gradient at Gaussian-blurred particle locations while updating the original particles. In a one-dimensional Gaussian example, noisy flow falls below 10^-4, while noise-free flow remains above 10^-3; this does not prove universal convergence.

Key insights

  • The Maximum Mean Discrepancy (MMD) functional is Λ-displacement convex, indicating its non-convex nature in the Wasserstein space, which can lead to convergence to local optima for its gradient flow.
  • Adaptive noise injection into the gradient updates provides a theoretical mechanism to ensure global convergence of the MMD flow, enabling it to escape spurious local minima.
  • The MMD gradient flow establishes a formal connection to the population limit dynamics of gradient descent in over-parameterized neural networks, offering a new perspective on deep learning optimization.

Summary by alphaXiv.

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