Reproduce “Kernel Belief Propagation” 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

Le Song, Arthur Gretton, Danny Bickson, Yucheng Low, Carlos Guestrin

KBP extends kernel belief propagation from tree models to loopy pairwise Markov random fields by learning conditional relationships from training pairs and representing messages in an RKHS, with tensor-product approximations for cheaper updates. It outperformed particle BP on protein angle prediction and achieved mean absolute error 0.145 in depth prediction, but convergence on loopy graphs remains unestablished.

Key insights

  • BP message updates can be re-expressed as conditional expectations, directly computable using RKHS embeddings without explicit conditional probability densities.
  • Representing messages and operators in RKHS provides a flexible framework for learning complex, non-Gaussian, and structured variable relationships nonparametrically from data.
  • Approximating feature matrices with a fixed, small number of basis functions allows for efficient, constant-time message updates, critical for scaling KBP to large datasets.

Summary by alphaXiv.

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