Reproduce “K2-ABC: Approximate Bayesian Computation with Kernel Embeddings” 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
Mijung Park, Wittawat Jitkrittum, Dino Sejdinovic
K2-ABC addresses likelihood-free Bayesian inference when models can simulate data but cannot evaluate likelihoods. It compares samples through kernel embeddings and maximum mean discrepancy, then applies soft ABC weighting, avoiding hand-selected summaries while retaining kernel and tolerance. In a constructed mixture example, its reported estimates were closer; in blowfly data, its chosen-summary distance was nearly 50% lower than SL-ABC.
Key insights
- Characteristic kernels provide an injective mapping from probability distributions to a Reproducing Kernel Hilbert Space, allowing MMD to act as an 'infinite-dimensional sufficient statistic' that captures all differences between distributions.
- By operating directly on empirical data distributions, K2-ABC inherently avoids the information loss and resulting bias associated with hand-crafted, finite-dimensional summary statistics.
- The fully non-parametric, data-driven nature of K2-ABC enhances its robustness and generalizability, making it applicable to a wider range of complex generative models without requiring specific domain knowledge for statistic selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.