f-DivergenceVariationalInference

For decades, the dominant paradigm for approximate Bayesian inferencep(z|x) = p(z,x)/p(x) has been Markov-Chain Monte-Carlo (MCMC) algorithms, which estimate the evidencep(x) = R p(z,x)dz via sampling. However, since sampling tends to be a slow and computationally intensive process, these sampling-based approximate inference methods fadewhendealing withthemodern probabilistic machine learning problems that usually involveverycomplexmodels, high-dimensional feature spaces andlargedatasets.