Learning Bayesian Statistics #150 Fast Bayesian Deep Learning, with David Rügamer, Emanuel Sommer & Jakob Robnik
18 snips
Jan 28, 2026 David Rügamer, LMU professor working on uncertainty in deep models; Emanuel Sommer, PhD researcher building practical JAX sampling tools; Jakob Robnik, Berkeley physicist developing the Microcanonical Langevin sampler. They discuss scaling Bayesian neural networks, fast sampling tricks and software, microcanonical dynamics, bottlenecks in high dimensions, hybrid warm-start strategies, and tooling for practical uncertainty quantification.
AI Snips
Chapters
Transcript
Episode notes
Software Is As Important As Samplers
- JAX plus modular samplers enabled practical scaling of Bayesian neural network sampling on GPUs.
- Software choices and engineering (memory, callbacks) matter as much as algorithms.
Warm-Start Before Sampling
- Warm-start sampling from an optimized network to reduce initialization error.
- Use parallel short chains and hybrid optimization-sampling workflows to allocate compute efficiently.
Fixed-Velocity Dynamics Scale Better
- Microcanonical Langevin dynamics keep chain velocity fixed, improving stability in sharp likelihood regions.
- Eliminating Metropolis correction and controlling discretization bias lets step sizes remain constant as dimension grows.