Presentation is loading. Please wait.

Presentation is loading. Please wait.

Belief Propagation in a Continuous World Andrew Frank 11/02/2009 Joint work with Alex Ihler and Padhraic Smyth TexPoint fonts used in EMF. Read the TexPoint.

Similar presentations


Presentation on theme: "Belief Propagation in a Continuous World Andrew Frank 11/02/2009 Joint work with Alex Ihler and Padhraic Smyth TexPoint fonts used in EMF. Read the TexPoint."— Presentation transcript:

1 Belief Propagation in a Continuous World Andrew Frank 11/02/2009 Joint work with Alex Ihler and Padhraic Smyth TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: A

2 Graphical Models Nodes represent random variables. Edges represent dependencies. C B AC B A C B A

3 CE DB A Markov Random Fields E DB CA D ACE B B  E | C, DA  C | B

4 Factoring Probability Distributions Independence relations  factorization D C BA p(A,B,C,D) = f(A) f(B) f(C) f(D) f(A,B) f(B,C) f(B,D)

5 Toy Example: A Day in Court W A EV A, E, W є {“Innocent”, “Guilty”} V є {“Not guilty verdict”, “Guilty verdict”} I G G I I G

6 Inference Most probable explanation: Marginalization:

7 Iterative Message Updates x

8 Belief Propagation W A EV m AE (E) m WE (E) m EV (V)

9 Loopy BP C A BD C A BD Does this work? Does it make any sense?

10 A Variational Perspective Reformulate the problem: True distribution, P “Tractable” distributions Best tractable approximation, Q Find Q to minimize the divergence.

11 Desired traits: – Simple enough to enable easy computation – Complex enough to represent P Choose an Approximating Family e.g. Fully factored: Structured:

12 Choose a Divergence Measure Kullback-Liebler divergence: Alpha divergence: Common choices:

13 Behavior of α-Divergence Source: T. Minka. Divergence measures and message passing. Technical Report MSR-TR-2005-173, Microsoft. Research, 2005.

14 Resulting Algorithms Assuming a fully-factored form of Q, we get…* Mean field,α = 0 Belief propagation,α = 1 Tree-reweighted BP,α ≥ 1 * By minimizing “local divergence”: Q(X 1, X 2, …, X n ) = f(X 1 ) f(X 2 ) … f(X n )

15 Local vs. Global Minimization Source: T. Minka. Divergence measures and message passing. Technical Report MSR-TR-2005-173, Microsoft. Research, 2005.

16 Applications

17 Sensor Localization A B C

18 Protein Side Chain Placement RTDCYGN +

19 Common traits? ? Continuous state space:

20 Easy Solution: Discretize! 10 bins Domain size: d = 100 20 bins Domain size: d = 400 Each message: O(d 2 )

21 Particle BP We’d like to pass “continuous messages”… C A BD B m AB (B) 144.252.5……… Instead, pass discrete messages over sets of particles: { b (i) } ~ W B (B) m AB ({b (i) }) b (1) b (2) b (N)...

22 PBP: Computing the Messages Re-write as an expectation: Finite-sample approximation:

23 Choosing“Good” Proposals C A BD Proposal should “match” the integrand. Sample from the belief:

24 Iteratively Refine Particle Sets (2) f(x s, x t ) (1)Draw a set of particles, {x s (i) } ~ W s (x s ). (2)Discrete inference over the particle discretization. (3)Adjust W s (x s ) (1) (3) XsXs XtXt (1) (3)

25 Benefits of PBP No distributional assumptions. Easy accuracy/speed trade-off. Relies on an “embedded” discrete algorithm. Belief propagation, mean field, tree-reweighted BP…

26 Exploring PBP: A Simple Example xsxs ||x s – x t ||

27 Continuous Ising Model Marginals Approximate Exact Mean Field PBP α = 0 PBP α = 1 TRW PBP α = 1.5 * Run with 100 particles per node

28 A Localization Scenario

29 Exact Marginal

30 PBP Marginal

31 Tree-reweighted PBP Marginal

32 Estimating the Partition Function Mean field provides a lower bound. Tree-reweighted BP provides an upper bound. p(A,B,C,D) = f(A) f(B) f(C) f(D) f(A,B) f(B,C) f(B,D) Z = f(A) f(B) f(C) f(D) f(A,B) f(B,C) f(B,D)

33 Partition Function Bounds

34 Conclusions BP and related algorithms are useful! Particle BP let’s you handle continuous RVs. Extensions to BP can work with PBP, too. Thank You!


Download ppt "Belief Propagation in a Continuous World Andrew Frank 11/02/2009 Joint work with Alex Ihler and Padhraic Smyth TexPoint fonts used in EMF. Read the TexPoint."

Similar presentations


Ads by Google