Date: May 22, 2026

Time: 12:00 pm – 1:00 pm

Location: Sessler Conference Room – 50A-5132 [Hybrid and In-Person]

Speaker: Jacob Robnik (UC Berkeley)

Title: Feasible MCMC in high-dimensional parameter spaces

Abstract: Statistical inference increasingly takes place in high-dimensional parameter spaces (thousands to billions of parameters) and is commonly the computational bottleneck of scientific workflow. To make this task scalable we have developed Microcanonical Langevin Monte Carlo (MCLMC), which is the first ever out-of-the-box MCMC algorithm without Metropolis adjustment. Avoiding the adjustment gives it superior scaling with the dimensionality of the problem, while still maintaining error guarantees. MCLMC outperforms current state-of-the-art samplers like the No-U-Turn Sampler (NUTS) on all of the problems we considered. For example, for cosmological field level inference from galaxy surveys with 2 million parameters, MCLMC inference takes 4 hours while NUTS needs more than 3 days for the same accuracy and computational resources. The gains are growing further as the dimensionality increases. This kind of improvement can lead to scientific breakthroughs, for example by enabling optimal information extraction or by enabling the study of solid state systems of unprecedented size. In the talk I will give intuition about how scalable MCMC must work in high dimensions and then introduce MCLMC. I will show practical applications to benchmarks, lattice field theories, cosmological field level inference and Bayesian neural networks.

Join Zoom Meeting
https://lbnl.zoom.us/j/95016696011?pwd=Tk1XOW1Xd3RYRnlsc2tEYmRWZlVVZz09

Meeting ID: 950 1669 6011

Passcode: 247722