Learning Bayesian Statistics

#98 Fusing Statistical Physics, Machine Learning & Adaptive MCMC, with Marylou Gabrié

Jan 24, 2024
Guest
Marylou Gabrié
Marylou Gabrié, assistant professor at CMAP, Ecole Polytechnique in Paris, discusses the fusion of statistical physics and machine learning. Topics include machine learning for scientific computing, adaptive Monte Carlo with normalizing flows, sampling discrete parameters in generative models, and machine learning in scientific computing.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Multimodality and Metastability

  • Normalizing flows and deep neural networks help MCMC explore multimodal distributions efficiently.
  • This addresses metastability, a challenge in both physics and Bayesian inference.
ADVICE

FlowMC Package

  • Try the FlowMC package for multimodal posteriors.
  • It's a user-friendly Jax package that implements the discussed algorithms.
INSIGHT

Discrete Parameter Handling

  • Current implementation focuses on continuous spaces; discrete parameter sampling is theoretically possible.
  • Autoregressive models offer potential solutions for discrete distributions, similar to normalizing flows.
Get the Snipd Podcast app to discover more snips from this episode
Get the app