GENERALIZING HAMILTONIAN MONTE CARLO WITH NEURAL NETWORKS

HMC

• struggle to mix energy levels
• cannot easily traverse low-density zones
• slow mixing in some cases

introcude

• translation, rescale of gradient and rescale of mementum
• all implemented by multi layer perceptrons
• loss: max lag-one distance
  • on both target and prior dist.