Learning Bayesian Statistics BITESIZE | How Bayesian Additive Regression Trees Work in Practice
Oct 9, 2025
Gabriel Stechschulte, a Bayesian software developer known for his work with PyMCBART, dives into the re-implementation of Bayesian Additive Regression Trees (BART) in Rust. He discusses the technical hurdles and enhanced performance achieved through this project. Gabriel explains the value of BART in uncertainty quantification and how it contrasts with other tree-based methods. The conversation also covers practical aspects, like integrating BART with Python and balancing open-source contributions with a full-time job, all while exploring the innovative features of PyMCBART.
AI Snips
Chapters
Transcript
Episode notes
Joining The PyMCBART Rewrite
- Gabriel joined the PyMCBART speedup effort after Osvaldo asked for performance improvements and offered to help.
- He brought Rust experience and reimplemented parts to accelerate PyMCBART.
Reimplement First, Optimize Later
- Reimplement one-to-one from the original before optimizing to ensure correctness.
- Then iteratively profile and optimize specific functions for safe speedups.
BART Is A Bayesian Booster
- BART aligns with gradient boosting by learning and correcting residuals across many trees.
- It assembles trees via random perturbations evaluated by likelihood rather than greedy fitting.