Gates
Tom Minka, John Winn Context-sensitive dependencies in graphical models m
1

M45

p
Gaussian Gate

m1
1

p1 c

m2
2

p2

c x

Gate

x Mixture of Gaussians

Gate

x depends on m,p iff c=1
Expectation propagation Variational message passing Gibbs sampling

Messages change at gate boundaries

Simpler inference