360 likes | 376 Views
Linear response formula and generalized belief propagation for probabilistic inference. Kazuyuki Tanaka Graduate School of Information Sciences Tohoku University, Sendai 980-8579, Japan kazu@statp.is.tohoku.ac.jp http://www.statp.is.tohoku.ac.jp/~kazu/. Reference
E N D
Linear response formula and generalized belief propagation for probabilistic inference Kazuyuki Tanaka Graduate School of Information Sciences Tohoku University, Sendai 980-8579, Japan kazu@statp.is.tohoku.ac.jp http://www.statp.is.tohoku.ac.jp/~kazu/ Reference K. Tanaka: Probabilistic Inference by means of Cluster Variation Method and Linear Response Theory, IEICE Trans. on Inf. & Syst., vol.E86-D, no.7, pp.1228-1242, 2003. SMAPIP-MSI2004
Contents Introduction Probabilistic Inference Cluster Variation Method and Generalized Belief Propagation Linear Response Concluding Remarks SMAPIP-MSI2004
Introduction Bayes Formula Probabilistic Inference Probabilistic Model Belief Marginal Probability Probabilistic Inference and Belief Propagation Probabilistic models on networks with some loops =>Good Approximation Probabilistic models on tree-like networks with no loops =>Exact Results Belief Propagation Generalization SMAPIP-MSI2004
Introduction Probabilistic model with no loop Belief Propagation (Lauritzen, Pearl) Probabilistic model with some loops Probabilistic model with no loop Transfer Matrix Statistical Mechanics and Belief Propagation Transfer Matrix = Belief Propagation Recursion Formula for Beliefs and Messages Bethe/Kikuchi Method Cluster Variation Method Probabilistic model with some loops SMAPIP-MSI2004
Review of generalized loopy belief propagation based on cluster variation method. • Calculation of correlations between any pair of nodes by combining the cluster variation method with the linear response theory Purpose SMAPIP-MSI2004
Contents Introduction Probabilistic Inference Cluster Variation Method and Generalized Belief Propagation Linear Response Concluding Remarks SMAPIP-MSI2004
Probabilistic Inference SMAPIP-MSI2004
Probabilistic Inference Probabilistic Inference and Probabilistic Model SMAPIP-MSI2004
Probabilistic Inference Probabilistic Inference and Probabilistic Model SMAPIP-MSI2004
Probabilistic Inference Probabilistic Model and Beliefs Statistics: Marginal Probability Statistical Mechanics: One-body distribution Probabilistic Inference: Belief SMAPIP-MSI2004
Contents Introduction Probabilistic Inference Cluster Variation Method and Generalized Belief Propagation Linear Response Concluding Remarks SMAPIP-MSI2004
Belief Propagation • Belief Propagation is equivalent to Bethe Approximation Y. Kabashima and D. Saad: Europhys. Lett. 1998. • Correctness of Loopy Belief Propagation Y. Weiss: Neural Computation 2000. Y. Weiss and W. T. Freeman: Neural Computation 2001. • Generalized Belief Propagation was proposed based on Cluster Variation Method J. S. Yedidia, W. T. Freeman and Y. Weiss: NIPS 2000. • Interpretation of Generalized Belief Propagation based on Information Geometry S. Ikeda, T. Tanaka and S. Amari: Neural Computation 2004. SMAPIP-MSI2004
Cluster Variation Method • Original Formulation R. Kikuchi: Phys. Rev. 1951. T. Morita: J. Phys. Soc. Jpn 1957 J. Math. Phys. 1972 J. Stat. Phys. 1990. • Extension to Random Spin System T. Morita: Physica A, 1979. T. Horiguchi: Physica A, 1981. • Convergence of a Sequence of Approximations in CVM A. G. Schlijper: Phys. Rev. B 1983 J. Stat. Phys. 1985. SMAPIP-MSI2004
Cluster Variation Method and Generalized Belief Propagation SMAPIP-MSI2004
Generalized Loopy Belief Propagation Extreme Condition of Kullback-Leibler Divergence Message Passing Algorithm SMAPIP-MSI2004
Generalized Loopy Belief Propagation Expression of Marginal Probability in terms of Messages SMAPIP-MSI2004
Generalized Loopy Belief Propagation SMAPIP-MSI2004
GBP Numerical Experiments Exact SMAPIP-MSI2004
Numerical Experiments SMAPIP-MSI2004
Contents Introduction Probabilistic Inference Cluster Variation Method and Generalized Belief Propagation Linear Response Concluding Remarks SMAPIP-MSI2004
Linear Response Theory (LRT) • MFA + LRT • Stat. Phys. [R. J. Elliott and W. Marshall: Rev. Mod. Phys. 1958] • [M. E. Fisher and R. J. Bufford: Phys. Rev. 1967] • NIPS [T. Tanaka: Phys. Rev. 1998] • [H. J. Kappen et al Neural Computation 1998] • Bethe Approx. (BP) + LRT • Stat. Phys. [T. Morita: J. Phys. Soc. Jpn. 1968] • NIPS [M. Welling and Y. W. The: Neural Computation 2004] • Kramers-Wannier-Kikuchi Approx. + LRT • Stat. Phys. [J. M. Sanchez: Physica A 1982] • CVM (GBP) + LRT • Stat. Phys. [T. Morita: Prog. Theor. Phys. 1991] • [K. Tanaka, T. Horiguchi and T. Morita 1991] • NIPS [K. Tanaka: IEICE Trans. Inf. & Syst. 2003] Co-variance between any pair of nodes SMAPIP-MSI2004
Linear Response SMAPIP-MSI2004
Final Result SMAPIP-MSI2004
Basic Cluster and Subcluster • Example SMAPIP-MSI2004
Probabilistic Model • Joint Probability Distribution Example SMAPIP-MSI2004
Cluster Variation Method • Kullback-Leibler Divergence and Kikuchi Free Energy SMAPIP-MSI2004
Present Probabilistic Model Marginal Distribution in CVM SMAPIP-MSI2004
Probabilistic Model with External Field Marginal Distribution in CVM SMAPIP-MSI2004
The approximate marginal probability is expanded in powers of Lagrange multipliers and only the first order terms are remained. Linear Response in CVM SMAPIP-MSI2004
This equation can be regarded as a system of linear equations for Lagrange multipliers. Linear Response in CVM SMAPIP-MSI2004
Correlation Function in CVM SMAPIP-MSI2004
Correlation Function in CVM Example SMAPIP-MSI2004
GBP+LR Numerical Experiments Exact SMAPIP-MSI2004
GBP Numerical Experiments GBP + Linear Response SMAPIP-MSI2004
Contents Introduction Probabilistic Inference Cluster Variation Method and Generalized Belief Propagation Linear Response Concluding Remarks SMAPIP-MSI2004
Cluster Variation Method + Linear Response Theory → General CVM Approximate Formula for Correlation Concluding Remarks Future Problems • Accuracy of GBP • Inconsistency between GBP and GBP+LRT • Statistical Learning of Conditional Probability. • Maximum Likelihood Framework • EM algorithm SMAPIP-MSI2004