1 / 48

Probabilistic Inference Lecture 7

Probabilistic Inference Lecture 7. M. Pawan Kumar pawan.kumar@ecp.fr. Slides available online http:// cvc.centrale-ponts.fr /personnel/ pawan /. Recap. Loopy Belief Propagation. Initialize all messages to 1. In some order of edges, update messages.

brenna
Télécharger la présentation

Probabilistic Inference Lecture 7

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Probabilistic InferenceLecture 7 M. Pawan Kumar pawan.kumar@ecp.fr Slides available online http://cvc.centrale-ponts.fr/personnel/pawan/

  2. Recap

  3. Loopy Belief Propagation Initialize all messages to 1 In some order of edges, update messages Mab;k= Σiψa(li)ψab(li,lk)Πn≠bMna;i Not Guaranteed !! Until Convergence Rate of changes in messages < threshold

  4. Loopy Belief Propagation B’a(i) = ψa(li)ΠnMna;i B’ab(i,j) = ψa(li)ψb(lj)ψab(li,lj)Πn≠bMna;iΠn≠aMnb;j Normalize to compute beliefs Ba(i), Bab(i,j) At convergence Σj Bab(i,j) = Ba(i)

  5. Outline • Free Energy • Mean-Field Approximation • Bethe Approximation • Kikuchi Approximation Yedidia, Freeman and Weiss, 2000

  6. Exponential Family P(v) = exp{-ΣaΣiθa;iIa;i(va) -Σa,bΣi,kθab;ikIab;ik(va,vb) - A(θ)} Πaψa(va) Π(a,b)ψab(va,vb) Probability P(v) = Z A(θ) : log Z ψa(li) : exp(-θa(i)) ψa(li,lk) : exp(-θab(i,k))

  7. Exponential Family P(v) = exp{-ΣaΣiθa;iIa;i(va) -Σa,bΣi,kθab;ikIab;ik(va,vb) - A(θ)} Πaψa(va) Π(a,b)ψab(va,vb) exp(-Q(v)) Probability P(v) = = Z Z A(θ) : log Z ψa(li) : exp(-θa(i)) ψa(li,lk) : exp(-θab(i,k)) Energy Q(v) = Σaθa(va) + Σa,bθab(va,vb)

  8. Exponential Family Πaψa(va) Π(a,b)ψab(va,vb) exp(-Q(v)) Probability P(v) = = Z Z Approximate probability distribution B(v) B(v) has a simpler form than P(v) Minimize KL divergence between B(v) and P(v)

  9. Kullback-Leibler Divergence B(v) D = ΣvB(v) log P(v)

  10. Kullback-Leibler Divergence D = ΣvB(v) log B(v) - ΣvB(v) log P(v)

  11. Kullback-Leibler Divergence D = ΣvB(v) log B(v) + ΣvB(v) Q(v) - (- log Z) Helmholz free energy Constant with respect to B

  12. Kullback-Leibler Divergence ΣvB(v) log B(v) + ΣvB(v) Q(v) Negative Entropy U(B)

  13. Kullback-Leibler Divergence ΣvB(v) log B(v) + ΣvB(v) Q(v) Average Energy S(B)

  14. Kullback-Leibler Divergence ΣvB(v) log B(v) + ΣvB(v) Q(v) Gibbs free energy

  15. Outline • Free Energy • Mean-Field Approximation • Bethe Approximation • Kikuchi Approximation

  16. Simpler Distribution One-node marginals Ba(i) Joint probability B(v) = Πa Ba(va)

  17. Average Energy ΣvB(v) Q(v)

  18. Average Energy * ΣvB(v)(Σaθa(va) + Σa,bθab(va,vb)) * = Simplify on board !!!

  19. Average Energy ΣaΣiBa(i)θa(i) + Σa,bΣi,kBa(i)Bb(k)θab(i,k)

  20. Negative Entropy * ΣvB(v) log (B(v))

  21. Negative Entropy ΣaΣiBa(i)log(Ba(i))

  22. Mean-Field Free Energy ΣaΣiBa(i)θa(i) + Σa,bΣi,kBa(i)Bb(k)θab(i,k) + ΣaΣiBa(i)log(Ba(i))

  23. * Optimization Problem minB ΣaΣiBa(i)θa(i) + Σa,bΣi,kBa(i)Bb(k)θab(i,k) + ΣaΣiBa(i)log(Ba(i)) s.t. ΣiBa(i) = 1

  24. KKT Condition log(Ba(i)) = -θa(i) -ΣbΣkBb(k)θab(i,k) + λa-1 Ba(i) = exp(-θa(i) -ΣbΣkBb(k)θab(i,k))/Za

  25. Optimization Initialize Ba (random, uniform, domain knowledge) Set all random variables to unprocessed Pick an unprocessed random variable Va Ba(i) = exp(-θa(i) -ΣbΣkBb(k)θab(i,k))/Za If Ba changes, set neighbors to unprocessed Guaranteed !! Until Convergence Tutorial: Jaakkola, 2000 (one of several)

  26. Outline • Free Energy • Mean-Field Approximation • Bethe Approximation • Kikuchi Approximation

  27. Simpler Distribution One-node marginals Ba(i) Two-node marginals Bab(i,k) Joint probability hard to write down But not for trees

  28. Simpler Distribution One-node marginals Ba(i) Two-node marginals Bab(i,k) Πa,b Bab(va,vb) B(v) = Πa Ba(va)n(a)-1 n(a) = number of neighbors of Va Pearl, 1988

  29. Average Energy ΣvB(v) Q(v)

  30. Average Energy * ΣvB(v)(Σaθa(va) + Σa,bθab(va,vb))

  31. Average Energy * ΣaΣiBa(i)θa(i) + Σa,bΣi,kBab(i,k)θab(i,k)

  32. Average Energy -Σa(n(a)-1)ΣiBa(i)θa(i) + Σa,bΣi,kBab(i,k)(θa(i)+θb(k)+θab(i,k)) n(a) = number of neighbors of Va

  33. Negative Entropy * ΣvB(v) log (B(v))

  34. Negative Entropy -Σa(n(a)-1)ΣiBa(i)log(Ba(i)) + Σa,bΣi,kBab(i,k)log(Bab(i,k)) Exact for tree Approximate for general MRF

  35. Bethe Free Energy -Σa(n(a)-1)ΣiBa(i)(θa(i)+log(Ba(i))) + Σa,bΣi,kBab(i,k)(θa(i)+θb(k)+θab(i,k)+log(Bab(i,k)) Exact for tree Approximate for general MRF

  36. * Optimization Problem minB -Σa(n(a)-1)ΣiBa(i)(θa(i)+log(Ba(i))) + Σa,bΣi,kBab(i,k)(θa(i)+θb(k)+θab(i,k)+log(Bab(i,k)) s.t. ΣkBab(i,k) = Ba(i) Σi,kBab(i,k) = 1 ΣiBa(i) = 1

  37. KKT Condition log(Bab(i,k)) = -(θa(i)+θb(k)+θab(i,k)) + λab(k) + λba(i) + μab - 1 λab(k) = log(Mab;k)

  38. Optimization BP tries to optimize Bethe free energy But it may not converge Convergent alternatives exist Yuille and Rangarajan, 2003

  39. Outline • Free Energy • Mean-Field Approximation • Bethe Approximation • Kikuchi Approximation

  40. Local Free Energy Cluster of variables c V1 V2 V4 V3 Gc = ΣvcBc(vc)(log(Bc(vc)) + Σd “subset of c” θd(vd)) G12 = Σv1,v2 B12(v1,v2)(log(B12(v1,v2)) + θ1(v1) + θ2(v2) + θ12(v1,v2))

  41. Local Free Energy Cluster of variables c V1 V2 V4 V3 Gc = ΣvcBc(vc)(log(Bc(vc)) + Σd “subset of c” θd(vd)) G1 = Σv1 B1(v1)(log(B1(v1)) + θ1(v1))

  42. Local Free Energy Cluster of variables c V1 V2 V4 V3 Gc = ΣvcBc(vc)(log(Bc(vc)) + Σd “subset of c” θd(vd)) G12 = Σv1,v2 B12(v1,v2)(log(B1234(v1,v2,v3,v4)) + θ1(v1) + θ2(v2) + θ3(v3) + θ4(v4) + θ12(v1,v2) + θ13(v1,v3) + θ24(v2,v4) + θ34(v3,v4))

  43. Sum of Local Free Energies V1 V2 V4 V3 Sum of free energies of all pairwise clusters G12 + G13 + G24 + G34 Overcounts G1, G2, G3, G4 once !!!

  44. Sum of Local Free Energies V1 V2 V4 V3 Sum of free energies of all pairwise clusters G12 + G13 + G24 + G34 - G1 - G2- G3 - G4

  45. Sum of Local Free Energies V1 V2 V4 V3 Sum of free energies of all pairwise clusters G12 + G13 + G24 + G34 - G1 - G2- G3 - G4 Bethe Approximation !!!

  46. Kikuchi Approximations V1 V2 V4 V3 Use bigger clusters G1234

  47. Kikuchi Approximations V1 V2 V3 V5 V6 V4 Use bigger clusters G1245+ G2356 - G25 Derive message passing using KKT conditions!

  48. Generalized Belief Propagation V1 V2 V3 V5 V6 V4 Use bigger clusters G1245+ G2356 - G25 Derive message passing using KKT conditions!

More Related