This is about changing the way we build our training batch. We do this by changing the probability of selecting a certain class. For example if you have two classes, A and B, where we have much more instances of A than B, we could change our sampling system to select more B than A.