Probabilistic Inference Lab

Overview

The Probabilistic Inference Lab treats uncertainty as a first-class object in computation. We develop algorithms that represent, propagate, and reason about uncertainty rigorously — not as a post-hoc annotation, but as a core design constraint. Our methods range from foundational Bayesian computation (variational inference, Markov chain Monte Carlo, normalizing flows) to higher-level concerns: model comparison, experimental design, and the integration of prior scientific knowledge into learned models.

A central interest is scalability. Classical Bayesian methods are often computationally intractable at the scale required by modern scientific datasets. We study how to approximate posterior distributions faithfully while preserving the calibration properties that make probabilistic reasoning useful. Our work on amortized inference and meta-learning-for-inference has enabled Bayesian updating at previously impractical scales in collaborations with the Distributed Learning Systems Group and the Computational Neuroscience Group.

We also build probabilistic programming tools. Our open-source library Veyra Infer provides a composable primitive set for specifying probabilistic models and running inference in a unified framework. It is used by several other research groups on campus and by external collaborators. The group maintains close ties with the Meridian HPC facility for large-scale Monte Carlo computation.

Research themes

• Variational inference and normalizing flows for complex posteriors
• Scalable MCMC methods: HMC variants, parallel tempering, sequential Monte Carlo
• Amortized inference and neural posterior estimation
• Probabilistic programming language design and compilation
• Bayesian experimental design and active learning
• Model selection, Bayes factors, and marginal likelihood estimation
• Calibration of predictive distributions in large-scale models
• Simulation-based inference for scientific simulators