Broadly speaking, machine learning models can be subdivided into discriminative models and generative models. Discriminative models learn a map from some input to some output. In discriminative models, learning the process that generates the input is not relevant; it will just learn a map from the to the expected output.
Generative models, on the other hand, in addition to learning a map from some input to some output, also learn the process that generates the input and the output.
In this context, we say that discriminative models estimate 
: the conditional probability distribution of 
conditioned on 
. Note that, in this case, the input x is fixed, known a priori, and the discriminative model estimates the probability of 
, 
, but does not have any information...
