NumPyro で学ぶ ベイズ統計モデリング 【応用編】

Deep Dive into Bayesian Statistical Modeling with NumPyro

This course is designed for individuals with a foundational understanding of Bayesian statistics, offering practical applications of Bayesian modeling. We move beyond basic concepts to explore diverse scenarios, including:

• Linear Regression: Starting with fundamental examples like linear and robust regression.
• Change Point Detection: Learn to identify shifts and transitions in data.
• Rating Modeling: Apply Bayesian methods to model and predict ratings.
• Differential Equation Parameter Estimation: Estimate parameters in models defined by differential equations.

Throughout the course, you'll work through these applications using Google Colaboratory, gaining hands-on experience. We'll leverage the NumPyro probabilistic programming package, known for its speed thanks to the JAX backend. This allows for efficient parameter estimation and model fitting.

Why NumPyro?

• Python-Friendly: NumPyro is a Python package, making it accessible to those with existing Python experience.
• Fast Computation: JAX provides significant speed advantages for complex models.
• Probabilistic Programming: Explore a powerful paradigm for building and analyzing statistical models.

NumPyro Version: The course utilizes NumPyro version 0.7.2 (as of September 2021). While stable, please be aware that NumPyro is still under active development. Exercise caution when using it in production environments. Limited Japanese resources are available; familiarity with English documentation is recommended.

Prerequisites:

This course builds upon basic Bayesian statistical concepts, including MCMC (Markov Chain Monte Carlo) and understanding of prior and posterior distributions. However, the initial sections provide a NumPyro introduction using linear regression, making it accessible even if you have some familiarity with tools like Stan or PyMC3.

Important Notes:

• This course is a continuation of "NumPyro for Bayesian Statistical Modeling [Basics]" but doesn't require mastery of the foundational course.
• Some sections utilize experimental NumPyro features; interfaces may change in future versions.
• Code examples may be shared on blogs and articles to promote the adoption of probabilistic programming.