One of my favorite related papers: https://arxiv.org/pdf/1703.00848.pdf
For example, the canonical way to train deep machine translation systems is to make use of large parallel corpora - large set of translation pairs. In the case where this data may not exist, for structurally similar languages, one approach is to learn unsupervised word embeddings independently (i.e. using word2vec) which only require monolingual data and to find a transformation that aligns the two spaces. In many cases, neural networks still struggle with this task. However, some work in the last few years has demonstrated that a few seed word-translations is all you need to apply simple algorithms that can learn a linear or simple nonlinear transformation that work pretty well!
[1] Here is a method that proposes locally linear transformation maps for alignment: http://nakashole.com/papers/2018-emnlp-norm.pdf
So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible.
You train two of these: one for the original face, and one for the target face. Then, you compress with the “source” autoencoder, and then uncompress with the “target” autoencoder. And, voila, ‘source‘ face becomes ‘target’ face.