A tutorial on variational methods for machine learning

Abstract

In this work, we present a tutorial on deep generative models, specically Varia- tional Autoencoders (VAE) and deep Bayesian Neural Networks (BNN). In spite of being apparently distinct, both themes are intimately connected through the pro- babilistic view of machine leaning. Therefore, we describe approximate inference methods since exact approaches are only adequate in very limited conditions in which the computation time remains feasible. We discuss modern extensions to such methods, capable of adequately sca- ling to models with millions of parameters and equally large data sets. Furthermore, we illustrate their relevance in each chapter through applications using (deep) neural networks.