In this chapter, we discussed some quite common neural models that are employed for solving unsupervised tasks. Autoencoders allow you to find the low-dimensional representation of a dataset without specific limits to its complexity. In particular, the use of deep convolutional networks helps to detect and learn both high-level and low-level geometrical features that can lead to a very accurate reconstruction when the internal code is much shorter than the original dimensionality too. We also discussed how to add sparsity to an autoencoder, and how to use these models to denoise samples. A slightly different variant of a standard autoencoder is a variational autoencoder, which is a generative model that can improve the ability to learn the data-generating process from which a dataset is supposed to be drawn.
Sanger's and Rubner-Tavan's networks are neural models...