A Generating the Hamiltonian bound

Algorithm 6 Generating the (non-differentiable) Hamiltonian AIS variational bound.Sample z It can be observed that tuning β and q (z) lead to the largest gains in performance. T uning more parameters leads to significantly better results. Table 4: ELBO on the test set (higher is better). For K = 1 both methods reduce to plain VI. For HMC, we use half of the budget for the warm-up phase and half to draw samples.