Actually using GPs in practice with PyMC
summary
PyMC's GP tools are shown as practical: a hierarchical model combines team, player and season effects with an age GP and yields a clear age–SLG peak at 26. The workflow uses Matern 5/2 kernels, informative priors on length scale (example PC prior with a 5‑year lower threshold and 2.5% tail), and approximations like HSGP for large n. PyTensor rewrites avoid costly matrix inverses and targets backends (JAX/Numba). Key checks are blocking sub‑resolution length scales and constraining overly large length scales.
AI-written summary · about
chapters (16)
- 0:00Gaussian processes in PyMC
- 1:03GPs: powerful but challenging
- 3:11Bayesian modeling with PyMC
- 5:14Bayesian model example: batting
- 9:30Hierarchical modeling for baseball
- 15:31Hierarchical models to GPs
- 17:21PyTensor: PyMC's engine
- 21:37GPs as infinite models
- 24:24GPs: challenges and approximations
- 25:27Choosing GP covariance
- 29:07Length scale prior setup
- 33:38Data under-informs GP hyperparams
- 35:25GPs for age effects in baseball
- 37:12Practical GP tips
- 40:22GPs vs parametric functions
- 41:49Q&A: Priors vs flexibility
related talks
01h 30mMaking Gaussian Processes Useful
01h 33mMastering Gaussian Processes with PyMC
01h 30mA Primer on Gaussian Processes for Regression Analysis
42mBayesian Non-parametric Models for Data Science using PyMC3
40mBdrygp: A Boundary-Integrated Gaussian Process Model For Computer Code Emulation
01h 23mProbabilistic Graphical Models, HMMs using PGMPY
32mpyGAM: balancing interpretability and predictive power using...
01h 27mProbabilistic Python: An Introduction to Bayesian Modeling with PyMC