I have a quick question regarding the algorithms used for learning and inference in DBNs. After reading these threads (post 1, post 2) my understanding is, that inference in SMILE is always based on the unrolled DBN and the selected algorithm for BNs (default: junction-tree). Parameter learning with the EM is performed on the not unrolled DBN, since the transitional CPTs share the same parameters. What inference algorithm is used as a subroutine in the EM algorithm? Is it again the junction-tree algorithm?
1 post • Page 1 of 1