The classical view of statistical learning (and generalization) is a parametric one. Empirical Risk Minimisation considers a population from which ones draws data iid to produce the training data.
We instead adopt a nonparametric perspective. We emphasize that nowhere do we make any iid assumptions.