Methods for Scalable Probabilistic Inference - IPAM at UCLA

Explore methods for scalable probabilistic inference in astrophysics through this 46-minute conference talk by Dan Foreman-Mackey of the Flatiron Institute. Delivered at IPAM's Workshop III on Source Inference and Parameter Estimation in Gravitational Wave Astronomy, the presentation delves into recent developments in probabilistic programming that enable rigorous Bayesian inference with large datasets and complex models. Learn about scalable methods for time series analysis using Gaussian Processes, and discover open-source tools and computational techniques for accelerating inference in astrophysics. The talk covers topics such as high-dimensional integrals, Monte Carlo methods, gradient-based sampling, Hamiltonian sampling, automatic differentiation, and practical applications in exoplanet research. Gain insights into the challenges and solutions for handling large-scale probabilistic modeling in modern astrophysical data analysis pipelines.

Introduction
Why Im here
Punchlines
Tools
Integrals
Highdimensional integral
Physical mod
Monte Carlo
Good sampler
Fast probability calculations
There are other answers
Gradients
Higherorder information
Potential energy
Multiple parameters
Integrating a dynamical system
Any questions right now
Any other questions
Derivatives
Hamiltonian Sampling
Automatic differentiation
What is automatic differentiation
Example from my work
Open source tools
Deep learning tools
JAX
Exoplanet
Solarite
Nonstationarity
Scaling linearly
Summary
Documentation
Interfaces