Improving Generalization via Meta-Learning on Hard Samples