Exploring the Similarity of Representations in Model-Agnostic Meta-Learning