Bayesian Generalized Linear Models in a Terabyte World
Onno Zoeter · International symposium on image and signal processing and analysis/ISPA ... · 2007
This paper introduces extremely fast approximate inference schemes for Bayesian treatments of dynamic generalized linear models. The approximations are tailored variants of quadrature EP. The first forward pass of this fixed point iteration algorithm can be interpreted as a one-step unscented Kalman filter. For on-line applications this filter can handle tens of thousands of updates a second on a current day desktop machine.