Neural network gradient hamiltonian monte carlo

Hamiltonian monte carlo (HMC) is an effective way to sample from posterior distirbutions. HMC explores the parameter space by following the gradient field. We can speed up HMC with neural network gradient approximation when data are abundant.

Sampler

HMC sampler

Surrogate

• Shallow neural network
• Gradient neural network
• Gaussian process
• Stochastic gradient

Models

• Multivariate Gaussian
• Logistic regression with Normal prior
• Logistic regression with Laplace prior