NP-NUTS: A Nonparametric No-U-Turn Sampler
summary
A new algorithm, NP‑NUTS, automatically adapts HMC trajectory length for models whose parameter dimension varies by combining NUTS’s doubling/tree construction with NP‑HMC’s dynamic trace extension during the leapfrog integrator. It computes subtree weights by recursively combining extension‑likelihoods so the tree is built once without storing all nodes, and it uses a modified no‑u‑turn criterion that ignores irrelevant extensions to preserve detailed balance. The paper gives correctness proofs for several extension schemes and reports an implemented prototype aimed at scalable continuous + nonparametric models.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:28Nonparametric models motivation
- 3:16Hamiltonian Monte Carlo
- 6:51HMC and NUTS limits
- 8:18NUTS algorithm features
- 9:23HMC for nonparametric programs
- 11:56NP-NUTS hybrid approach
- 13:40NP-NUTS future directions
related talks
20mNonparametric Involutive Markov Chain Monte Carlo
34mFast, beautiful and easy Bayesian modelling: Can you have it all?
01h 29mA Beginner's Guide to Variational Inference
01h 06mNonparametric Bayesian Methods: Models, Algorithms, and Applications I
01h 27mProbabilistic Python: An Introduction to Bayesian Modeling with PyMC
54mTime Series forecasting with NumPyro
30mMatrix-free Construction of HSS Representations Using Adaptive Randomized Sampling
40mApproximation Algorithms for Optimization under Uncertainty