Neural Networks with Keras Cookbook

Grouping similar images

In the previous sections, we represented each image in a much lower dimension with the intuition that images that are similar will have similar embeddings and images that are not similar will have dissimilar embeddings. However, we have not yet looked at the similarity measure or examined embeddings in detail.

In this section, we will try and plot embeddings in a 2D space. We can reduce the 32-dimensional vector to a two-dimensional space by using a technique called t-SNE. (More about t-SNE can be found here: http://www.jmlr.org/papers/v9/vandermaaten08a.html.)

This way, our feeling that similar images will have similar embeddings can be proved, as similar images should be clustered together in the two-dimensional ...