ERM Learns Finite Realizable Classes

Why

We want to make a statement like “good performance on a correctly labeled training dataset of independent and identically distributed inputs leads to good generalization error.”