How t-SNE Outperforms PCA in Dimensionality Reduction