A0662
Title: Parallel and sequential coordinate ascent in variational inference
Authors: Debdeep Pati - Texas A&M University (United States) [presenting]
Abstract: A surprising discordance is described between the sequential and parallel versions of coordinate ascent in variational inference. Focusing on the case of high dimensional linear regression, it is shown that the parallel version exhibits a lack of convergence under a general setting. This can be effectively remedied by using a sequential version of the algorithm under fairly relaxed assumptions. The techniques involve analyzing the spectral norm of the Jacobian of specific nonlinear dynamical systems.
