A Beginner's Guide to Variational Inference
summary
Tutorial shows concrete claims and timings: PyMC runs ADVI and the newer Pathfinder and compares them against MCMC NUTS on real models. Example: Gaussian process baseball model that takes ~20 minutes with NUTS fits in 49 seconds with VI; a 400k × 100 model drops from ~20 minutes to ~20 seconds with minibatching. Pathfinder uses BFGS paths, importance reweighting and multipath jitter to recover multimodality and often matches MCMC quality while costing far fewer gradient evaluations.
AI-written summary · about
chapters (9)
- 0:00Variational inference intro
- 16:54Variational inference loss
- 18:25Simple VI example
- 32:24Generalizing with ADVI
- 46:54PyMC variational intro
- 1:08:56Advanced VI methods
- 1:18:52Pathfinder mode choice
- 1:20:36PyMC VI examples
- 1:27:00Summary of VI trade-offs
related talks
37mVariational Inference in Python
35mVariational Inference and Python
39mProbablistic Programming Data Science with PyMC3
21mProbabilistic Programming with Programmable Variational Inference
44mAn introduction to PyMC3
01h 27mProbabilistic Python: An Introduction to Bayesian Modeling with PyMC
42mBayesian Non-parametric Models for Data Science using PyMC3
01h 30mMaking Gaussian Processes Useful