CCM Colloq: Adrian Seyboldt (PyMC-Labs), on Fisher HMC sampling

Wednesday Feb 19, 2025, 10:00 AM → 11:00 AM America/New_York

3rd floor classrm (Flatiron)

Presenter: Adrian Seyboldt (PyMC-Labs)

Title: If Only My Posterior Were Normal: Introducing Fisher HMC

Abstract: Markov Chain Monte Carlo (MCMC) methods struggle with complex posterior
distributions. This talk introduces Fisher HMC, a method that adaptively
transforms parameter spaces to improve sampling efficiency of Hamiltonian MCMC.
We use the Fisher divergence to reshape the posterior, so that we can explore it
more effectively. I'll discuss the theoretical foundations, and show how this
framework can be used to find better mass matrices, or how we can use
computationally intensive normalizing flows for more adaptive transformations. I
will show how this is implemented in nutpie, and can be used to improve sampling
of PyMC and Stan models.

Bio: Adrian Seyboldt is a core PyMC developer and statistics consultant with
PyMC-Labs. He works on developing probabilistic programming methods, with
expertise in Bayesian computational techniques and scientific software
development.