The smaller the VC dimension, the less training data is needed for learning. The goal of learning is to create an admissible set of functions with a small VC dimension, which contain good functions that can be picked up with a small amount of observations.