Information driven parameter dynamics on-line Bayesian learning with sequential Monte Carlo
Kuniaki Yosui, Makio Wakahara, Yohei Nakada, Takashi Matsumoto · 2005
A new parameter dynamics that incorporates the information available for training instead of the standard "blind" parameter dynamics is proposed for on-line Bayesian learning. A significant improvement is realized over the schemes the authors have previously proposed. The particular advantage of the currently proposed approach is the speed at which it follows abrupt changes.