ON VARIATIONALMESSAGEPASSINGON FACTORGRAPHS

Inthis paper, itisshownhow(naive andstructured) variational algorithms maybederived fromafactor graph bymechanically applying generic message computation rules; inthis way,onecan bypass error-prone variational calculus. Inprior workbyBishop etal., Xingetal., andGeiger, directed andundirected graphicalmodels havebeenusedforthis purpose. Thefactor graph notation amounts tosimpler generic variational message computation rules; bymeansoffactor graphs, variational methods canstraightforwardly becompared toandcombined withvariousother message-passing inference algorithms, e.g., Kalmanfilters andsmoothers, iterated conditional modes, expectation maximization (EM),gradient methods, andparticle filters. Someof those combinations havebeenexplored intheliterature, others seemtobenew.Generic message computation rules forsuch combinations areformulated.