Export Reviews, Discussions, Author Feedback and Meta-Reviews

This paper presents a scalable second-order stochastic variational inference method for models with normally distributed latent variables. In order to efficiently compute the gradient and Hessian, a representrization trick for general location-scale family is adopted with computation scales quadratically w.r.t the number of latent variables for both gradient and Hessian (in practice with diagonal Gaussian). Furthermore, Hessian-free optimization is used to account for the high dimensionality of the underlying embedded parameters. R-operator technique is used and it enables exact Hessian-vector product computation in Hessian-free optimization. L-BFGS can also be used in place of Hessian-free within the proposed framework.