I am interested in capturing higher-dimensional embeddings of a image dataset as a gaussian noise, such that a specific region of gaussian noise corresponds to embedding of a particular label. How do we get to do that? For instance, suppose I have two classes $A$, $B$ in say 100 dimensions. Now, I want to do train a Normalizing flow which would have $z_1, z_2, \ldots, z_{100}$ gaussian variables, such that $z_1 > 0$ implies the normalizing flow came from class $A$ and $z_1 < 0$ implies class $B$. How do we achieve that ?
Is conditional Normalizing flow help in achieving this?
