Mastering Machine Learning Algorithms

t-SNE

This algorithm, proposed by Van der Mateen and Hinton and formally known as t-Distributed Stochastic Neighbor Embedding (t-SNE), is one of the most powerful manifold dimensionality reduction techniques. Contrary to the other methods, this algorithm starts with a fundamental assumption: the similarity between two N-dimensional points x_i and x_j can be represented as the conditional probability p(x_j|x_i) where each point is represented by a Gaussian distribution centered in x_i and with variance σ_i. The variances are selected starting from the desired perplexity, defined as:

Low-perplexity values indicate a low uncertainty, and are normally ...