Variational Lower Bound Exact posterior inference is hard in most Bayesian models. Variational inference proposes a way of approximating a posterior over latent variables, $p(z \mid x)$, with a variational distribution, $q(z \mid x)$. The idea is that we can specify some family of distributions, $q_{\theta}(.)$ to approximate the true posterior, then learn the distributional parameters, $\theta$ that make $q_{\theta}$ as close as possible to the true posterior. How to measure closeness? One way ...