Python Machine Learning Cookbook - Second Edition

There's more...

A t-SNE algorithm is divided into two main phases. In the first phase, a probability distribution is constructed so that each pair of points in the original high-dimensional space associates a high probability value if the two points are similar, and low if they are dissimilar. Then, a second analogous probability distribution is defined in the small-sized space. The algorithm then minimizes the divergence of Kullback–Leibler of the two distributions by descending the gradient, reorganizing the points in the small-sized space.