EM, Variational Message Passing

by allenlu2007

 

之前 EM algorithm 有幾何解釋,代數解釋。

幾何基本上 based on concave (or convex) function;  代數大多 based on Jensen’s inequality.

雖然可以說明,但缺少一個大的 picture.

 

 

另外的方法是用 variational message passing 來說明。

請參考 Tom and Bishop paper

For Bayesian DAG graph

NewImage 

 

 

 

 

 

 

 

 

 

EM algorithm:  Q(H) = ??    Lower bound, L(Q), 就變成??

NewImage

 

 

Variational message passing:  Q(H)  (i) factorization form; (ii) conjugate exponential model

NewImage

Advertisements