Approximating the Partition Function by Deleting and then Correcting for Model Edges Arthur Choi and Adnan Darwiche University of California, Los Angeles {aychoi,darwiche}@cs.ucla.edu
Edge Deletion: Idea Delete an edge Model M Model M'
Deleting an Equivalence Edge j
Deleting an Equivalence Edge j j
Deleting an Equivalence Edge j
Deleting an Equivalence Edge j
Deleting an Equivalence Edge j
Deleting an Equivalence Edge j
Edge Parameters: ED-BP i j
Edge Parameters: ED-BP i j Conditions on edge parameters imply an iterative algorithm: ED-BP
Edge Parameters: ED-BP i j Yields a weaker notion of equivalence:
A Spectrum of Approximations ED-BP networks: [CD06]
A Spectrum of Approximations ED-BP networks: [CD06] Exact Inference
A Spectrum of Approximations ED-BP networks: [CD06] Loopy BP marginals Exact Inference
A Spectrum of Approximations ED-BP networks: [CD06] Loopy BP marginals Exact Inference partition function?
A Partition Function i j
An Easy Case: Delete a Single Edge Prop.: If MI(Xi,Xj) = 0 in ED-BP network M', then: where i j
An Easy Case: Delete a Single Edge Prop.: If MI(Xi,Xj) = 0 in ED-BP network M', then: where i With multiple edges deleted (ZERO-EC): j
Bethe Free Energy and ZERO-EC Bethe free energy approximation: as a partition function approximation:
Bethe Free Energy is ZERO-EC Bethe free energy approximation: as a partition function approximation: Theorem: The Bethe approximation is ZERO-EC when M' is a tree :
An Easy Case: Delete a Single Edge Prop.: If MI(Xi,Xj) = 0 in ED-BP network M', then: where i With multiple edges deleted (ZERO-EC): j
An Easy Case: Delete a Single Edge Prop.: For any edge in an ED-BP network M', then where i With multiple edges deleted (GENERAL-EC): j
Overview tree exact exact marginals marginals LBP IJGP
joingraph free energies Overview tree exact exact marginals marginals LBP IJGP joingraph free energies zero-EC Bethe exact Z
Overview tree exact exact marginals marginals LBP IJGP joingraph free energies zero-EC Bethe exact Z improved approximations (higher order EP/GBP energies) general-EC
joingraph free energies Overview tree exact exact marginals marginals LBP IJGP joingraph free energies zero-EC Bethe exact Z recover edges general-EC
Edge Recovery: ZERO-EC i j Recover edges with largest MI(Xi;Xj)
Edge Recovery: GENERAL-EC j i t s Recover edges with largest MI(Xi,Xj; Xs,Xt)
Edge Recovery Bethe exact Z 6x6 grid EC-Z,rand 0.07 0.06 0.05 relative error 0.04 0.03 0.02 exact Z 0.01 edges recovered 25
Edge Recovery Bethe exact Z 6x6 grid EC-Z,rand EC-G,rand 0.07 0.06 0.05 relative error 0.04 0.03 0.02 exact Z 0.01 edges recovered 25
Edge Recovery Bethe exact Z 6x6 grid EC-Z,rand EC-G,rand 0.07 EC-Z,MI 0.06 0.05 relative error 0.04 0.03 0.02 exact Z 0.01 edges recovered 25
Edge Recovery Bethe exact Z 6x6 grid EC-Z,rand EC-G,rand 0.07 EC-Z,MI EC-G,MI 0.06 0.05 relative error 0.04 0.03 0.02 exact Z 0.01 edges recovered 25
Edge Recovery Bethe exact Z 6x6 grid EC-Z,rand EC-G,rand 0.07 EC-Z,MI EC-G,MI 0.06 EC-G,MI2 0.05 relative error 0.04 0.03 0.02 exact Z 0.01 edges recovered 25
Edge Recovery
joingraph free energies Overview tree exact exact marginals marginals LBP IJGP/GBP joingraph free energies zero-EC Bethe exact Z recover edges general-EC
joingraph free energies Overview tree exact exact marginals marginals LBP IJGP/GBP joingraph free energies zero-EC Bethe exact Z partial corrections: general-EC
Partial Correction pigs 1 ED-BP (Bethe) 0.8 0.6 relative error 0.4 general-EC 0.2 500 1000 1500 2000 2500 3000 time (ms)
Partial Correction pigs 1 ED-BP (Bethe) 0.8 0.6 relative error soft-sep see AAAI’08 0.4 general-EC 0.2 500 1000 1500 2000 2500 3000 time (ms)
Partial Correction
joingraph free energies Overview tree exact exact marginals marginals LBP IJGP/GBP joingraph free energies zero-EC Bethe exact Z general-EC
joingraph free energies Overview tree exact exact marginals marginals LBP IJGP/GBP marginal corrections (AAAI'08) joingraph free energies zero-EC Bethe exact Z general-EC
Thanks!