Towards Understanding Generalization in Gradient-Based Meta-Learning