80a160ff31266be2f93012a2a3eca713-AuthorFeedback.pdf

For the Markov proposal, we take homogeneous random-walkq(x|y) = N(x|y,σI) and scale σ with dimen-22 sions as proposed in (Roberts, 2001, optimal scaling) to keep the acceptance rate about20%. For the indepen-23 dent proposal, we take homogeneous Gaussianq(x) = N(0,σI),σ = 1.2. Empirical results are in Figure 1.24 Further,we extend the algorithm tothe case ofMarkovproposal. DoyoutrainanencoderfortheVAE?"42 Yes, we train the VAE as in the original paper, then we use only the decoder as a generator by sampling the latent43 variablesfromtheprior.