Given a supervised problem with X, y input pairs, one can do two things for obtaining the function f that maps X with y with Neural Networks (and in general in machine learning):
Deploy directly a supervised learning algorithm that maps X to y
Deploy a (variational) auto-encoder for learning useful features, and then using these for training the supervised learning algorithm
I would like to be pointed to some papers/blogs that explain which technique is better and when or where they conduct empirical benchmarking experiments.