Dynamic inference in probabilistic graphical models
summary
Presents an algorithm that updates samples and estimates after small edits to a graphical model with cost roughly proportional to the edit size rather than redoing full MCMC, building a dynamic data structure that maintains an execution log coupling of two Gibbs chains so samples are transformed to exact Gibbs draws of the new model. Under the Dobrich-Schlossmann condition and dynamic Gibbs sampling the expected time is O(n' + |V'|) up to polylog factors, beating static MCMC’s Ω(n'·|V'|) for large models; supports hard constraints and weight edits.
AI-written summary · about
chapters (8)
- 0:00Intro
- 0:16Graphical model
- 3:28Sampling
- 4:23Graphical models
- 6:11Dynamic influence problem
- 8:21Dynamic influence algorithm
- 9:41Dynamic Gibbs sampling
- 14:03Gibbs sampling
related talks
59mWhen do dependencies in your data help?
25mStatically and Dynamically Delayed Sampling for Typed Probabilistic Programming Languages
26mInference algorithms and sampling techniques
27mDistribution Learning Meets Graph Structure Sampling
18mFast Inference for Probabilistic Graphical Models
01h 23mProbabilistic Graphical Models, HMMs using PGMPY
11mA Design for Massively Parallel Gibbs Sampling on the GPU via Static and Dynamic(…)
32mLearning Ill-Conditioned Gaussian Graphical Models