Reproduce “Statistical Inference for Generative Models with Maximum Mean Discrepancy” 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
Francois-Xavier Briol, Alessandro Barp, Andrew B. Duncan, Mark Girolami
This research introduces a theoretical and algorithmic framework for statistical inference in generative models with intractable likelihoods, utilizing Maximum Mean Discrepancy (MMD). It establishes the consistency, asymptotic normality, and robustness of MMD estimators, while also proposing a natural gradient descent method for efficient optimization, validated across diverse complex models.
Key insights
- MMD estimators offer dimension-independent generalization bounds and exhibit strong consistency, providing a statistically sound alternative for inference in intractable generative models.
- The choice of kernel lengthscale in MMD estimators involves a trade-off between statistical efficiency and robustness, with intermediate lengthscales often yielding optimal robustness to data perturbations.
- Leveraging the MMD-induced information geometry through Natural Gradient Descent significantly accelerates optimization for complex models by accounting for the local structure of the parameter space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.