Beyond Worst-Case Generalization In Modern Machine Learning