Fig. 8.

t-SNE representation of images from different datasets before and after transform. Each point is obtained by flattening the corresponding image and performing dimensionality reduction. For each dataset, only 200 images are represented.