A fast, universal algorithm to learn parametric nonlinear embeddings

Using the method of auxiliary coordinates, we derive a training algorithm that works by alternating steps that train an auxiliary embeddingwith steps that train the mapping.