Create new faces controlling unsupervise learned features.
Using a Deep Learning Network called Variational AutoEncoder I've used the latent space it creates for infering new faces. As the features are self-created, I don't have any controll from what they represent (as the data was unlabed anyway, btw here is the dataset link). The labels (in beautifull Spanish) are just to make the creation easier but they don't really have to represent exactly that.
For what? Imagine you have to describe a suspect as you are the only witness from a bank robbery. What an easier way to do it that implement real imagery algorithms. Also was a very fun project.
Here is an example:
(An entire video can be found on my Linkedn page somewhere here)
Would like to implement that using a GAN, my computation power is pretty poor right now unfortunatelly. Guess stuff may change in this summer. May post something if I get the oportunity.