There are ML architectures that can do that. The two halves of an autoencoder learn a “shared language” that allows them to communicate through a bottleneck.