E.g. a high dimensional domain like the D: domain of images
Discriminative Model
Description: let's assume it has been trained to discriminate images of cats from non cats
Input: D
Output: Scalar Value representing the probability of belonging to cat category
This value can then be thresholded to get to a binary result
Comments
so from a statistical point of view the discriminative model (or discriminator) has learned the Cats PDF hence a function associating to a certain point in its domain (whatever it is) a scalar value that can be interpreted as a probability
from a set theory point of view, this function has split D into 2 subsets: the D_{cats} (images of cats) and ~D_{cats} (images of non cats)
Generative Model
Description: let's assume it has been trained to produce realistic cats images
Input: Random Value (e.g. a random scalar value)
Output: an element in D which, more precisely, is in its subset D_{cats} (realistic cat)
Comments
this is also a function but totally different from the previous PDF as its domain is completely different and its codomain is PDF domain
this codomain-domain connection is what makes GAN work
Overview
Basic Elements about Generative Models
TODO