Sublinear Variational Optimization of Gaussian Mixture Models with Millions to Billions of Parameters