Groups
Empirical Risk Minimization (ERM) chooses a model that minimizes the average loss on the training data.
PAC learning formalizes when a learner can probably (with probability at least 1−δ) and approximately (error at most ε) succeed using a polynomial number of samples.