arXiv.itsnick.co

an independent field guide to 23 million papers from arXiv & OpenAlex — not affiliated with arXiv.org

Mathematics · Statistics and Probability

Stochastic gradient Hamiltonian Monte Carlo. ICML

Tianqi Chen, Emily B. Fox, Carlos Guestrin

See on the map

Abstract

Hamiltonian Monte Carlo (HMC) sampling methods provide a mechanism for defining dis-tant proposals with high acceptance probabilities in a Metropolis-Hastings framework, enabling more efficient exploration of the state space than standard random-walk proposals. The popularity of such methods has grown significantly in recent years. However, a limitation of HMC methods is the required gradient computation for simula-tion of the Hamiltonian dynamical system—such computation is infeasible in problems involving a large sample size or streaming data. Instead, we must rely on a noisy gradient estimate computed from a subset of the data. In this paper, we ex-plore the properties of such a stochastic gradient HMC approach. Surprisingly, the natural imple-mentation of the stochastic approximation can be arbitrarily bad. To address this problem we intro-duce a variant that uses second-order Langevin dynamics with a friction term that counteracts the effects of the noisy gradient, maintaining the de-sired target distribution as the invariant distribu-tion. Results on simulated data validate our the-ory. We also provide an application of our meth-ods to a classification task using neural networks and to online Bayesian matrix factorization. 1.

More like this