Variational Inference

Problem-style sheet on VI evidence lower bounds, ELBO–KL decompositions, mean-field updates, EM for a two-component Gaussian mixture, and ELBO-based views of EM.

Problem sets on ELBO vs KL to the true posterior, mean-field restriction, coordinate ascent updates, a small graphical model, EM vs VI, and diagonal Gaussian encoders in VAEs.

A full worked exercise: joint and complete-data log-likelihood, responsibilities, ELBO decomposition, M-step updates, and why EM is a special case of VI.