ebook img

Graphical)Models - University of Massachusetts Amherst PDF

33 Pages·2011·1.29 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Graphical)Models - University of Massachusetts Amherst

Graphical  Models Lecture  12:     Belief  Update  Message  Passing Andrew  McCallum [email protected] Thanks  to  Noah  Smith  and    Carlos  Guestrin  for  slide  materials. 1 Today’s  Plan • Quick  Review:    Sum  Product  Message  Passing  (also   known  as  “Shafer-­‐Shenoy”) • Today:    Sum  Product  Divide  Message  Passing  (also   known  as  “belief  update”  message  passing  “Lauritzen-­‐ Spiegelhalter”  and  “belief  propagaFon”) • MathemaFcally  equivalent,  but  different  intuiFons. • Moving  toward  approximate  inference. ϕ ϕ Quick  Review F A ϕ Flu All. • {H,  R,  S,  A,  F} FAS S.I. ϕ ϕ SR SH ϕ SR ψ R.N. H. SR SR τ S ϕ F ϕ A ν ϕ ϕ FAS SH FAS ψ SH HS SAF ν (C ) = φ j j τ S ψ φ Φ FAS ￿∈ j Message  Passing  (One  Root) • Input:    clique  tree  T,  factors  Φ,  root  C r • For  each  clique  C ,  calculate  ν   i i • While  C  is  sWll  waiWng  on  incoming  messages: r – Choose  a  C  that  has  received  all  of  its  incoming   i δ = ν δ i j i k i → → messages. C S k Neighbors j i￿\ i,j ∈ ￿ i\{ } – Calculate  and  send  the   message  from  C  to  C i upstream-­‐neighbor(i)β = ν δ r r k r → k Neighbors ∈ ￿ r = φ X C φ Φ ￿\ i ￿∈ = Z P(C ) r · Different  Roots;  Same  Messages δ SAF→SR β SR δ SR SR→SAF HS SAF root β β SAF HS δ HS→SAF δ SAF→HS Sum-­‐Product  Message  Passing • Each  clique  tree  vertex  C  passes  messages  to   i each  of  its  neighbors  once  it’s  ready  to  do  so. • At  the  end,  for  all  C : i β = ν δ i i k i → k Neighbors ￿ i ∈ – This  is  the  unnormalized  marginal  for  C . i Calibrated  Clique  Tree • Two  adjacent  cliques  C  and  C  are  calibrated   i j when: β = β i j C S C S i￿i,j j￿i,j \ \ = µ (S ) i,j i,j β β i j μ C i,j C i j S  = i,j C  ∩  C i j Calibrated  Clique  Tree  as  a   • Original  (unnormalized)  factor  model  and   calibrated  clique  tree  represent  the  same   (unnormalized)  measure: clique  beliefs β C C Vertices( ) ￿ φ = ∈ T µ φ Φ S ￿ ∈ S Edges( ) ￿ ∈ T sepset  beliefs unnormalized  Gibbs distribuWon  from  original factors  Φ Inventory  of  Factors • original  factors  ϕ • iniWal  potenWals  ν • messages  δ • intermediate  factors  ψ  (no  longer  explicit) • clique  beliefs  β • sepset  beliefs  μ Inventory  of  Factors • original  factors  ϕ • iniWal  potenWals  ν • messages  δ • intermediate  factors  ψ  (no  longer  explicit) • clique  beliefs  β • sepset  beliefs  μ New  algorithm   collapses   everything  into   beliefs!

Description:
Graphical)Models Lecture)12:) Belief)Update)Message)Passing Andrew)McCallum [email protected] Thanks)to)Noah)Smith)and))Carlos)Guestrin)for)slide)materials. 1
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.