Computation

Routing Summary

This folder covers computational methods for Bayesian inference from BDA3 Part III and Statistical Rethinking Chapter 8. Contains 5 notes.

Sub-topics

Sub-topicNotesCovers
Variational Inference8ELBO and KL minimization, mean-field CAVI, stochastic / black-box VI, ADVI, reparameterization trick and VAEs, normalizing-flow posteriors, diagnosing VI with PSIS and VSBC — Blei 2017, Ranganath 2014, Kucukelbir 2017, Kingma & Welling 2013, Rezende & Mohamed 2015, Yao 2018
Neural Simulation-Based Inference8Neural posterior / likelihood / ratio estimation, sequential neural likelihood, conditional normalizing flows, amortized vs sequential inference, benchmarking and diagnostics (SBC, coverage, C2ST), neural SBI for economic ABMs — Cranmer 2020, Papamakarios 2016/2019, Hermans 2020, Lueckmann 2021, Dyer 2022

Concept Map

ConceptNoteTypeDepends OnKey Result
Numerical integration, rejection sampling, importance samplingIntroduction to Bayesian ComputationconceptProbability and Bayesian Inference, Multiparameter ModelsFoundation for all computational methods
Gibbs sampler, Metropolis-Hastings, R-hat, n_effMCMC BasicsconceptIntroduction to Bayesian Computation, Probability and Bayesian Inference, Hierarchical ModelsConvergence diagnostics with R-hat and n_eff
HMC, NUTS, Stan, reparameterizationEfficient MCMCconceptMCMC Basics, Introduction to Bayesian ComputationHMC scales to high dimensions via gradient info
Variational inference, Laplace, expectation propagationApproximation MethodsconceptAsymptotics and Frequentist Connections, Efficient MCMC, Probability and Bayesian InferenceFast approximate posteriors trading accuracy for speed
King Markov parable, HMC/NUTS intuition, map2stanHMC and Stan in PracticetutorialMCMC Basics, Efficient MCMC, Garden of Forking DataPractical HMC diagnostics and Stan workflow

Notes

  • Introduction to Bayesian Computation — CONTAINS: Numerical integration, rejection sampling, importance sampling, simulation basics
  • MCMC Basics — CONTAINS: Gibbs sampler, Metropolis-Hastings algorithm, convergence diagnostics (R-hat, n_eff), mixing
  • Efficient MCMC — CONTAINS: Hamiltonian Monte Carlo, NUTS algorithm, Stan interface, reparameterization tricks
  • Approximation Methods — CONTAINS: Variational inference (ADVI), Laplace approximation, expectation propagation, accuracy-speed tradeoff
  • HMC and Stan in Practice — CONTAINS: King Markov parable, HMC/NUTS intuition, map2stan interface, practical diagnostics

Sources

  • BDA3 — Bayesian Data Analysis, 3rd Edition (Gelman et al.), Part III (pp. 259-349)
  • StatRethink-Bayes — Statistical Rethinking (McElreath, 2015), Chapter 8