The recursive variational Gaussian approximation (R-VGA)

We consider the problem of computing a Gaussian approximation to posterior distribution parameter given N observations and prior. Owing need processing large sample sizes N, variety approximate tractable methods revolving around online learning have flourished over past decades. In present work, we propose use variational inference compute through single pass data. Our algorithm is recursive version called approximation. start from prior, for each observation, nearest in sense Kullback–Leibler divergence this observation. turn, considered as new prior when incorporating next This based on sequence optimal approximations leads novel implicit update scheme which resembles Newton shown boil down Kalman filter Bayesian linear regression. context logistic regression, may be solved, perform better than extended filter, while being less computationally demanding its sampling counterparts.