1. Distributed Message Passing for Large Scale Graphical Models Alexander Schwing Tamir Hazan Marc Pollefeys Raquel Urtasun CVPR2011

2. Outline Introduction Related work Message passing algorithm Distributed convex belief propagation Experiment evaluation Conclusion

3. Introduction Vision problems → discrete labeling problems in an undirected graphical model (Ex : MRF) – Belief propagation (BP) – Graph cut Depending on the potentials and structure of the graph The main underlying limitations to real-world problems are memory and computation.

4. Introduction A new algorithm – distribute and parallelize the computation and memory requirements – conserving the convergence and optimality guarantees Computation can be done in parallel by partitioning the graph and imposing agreement between the beliefs in the boundaries. – Graph-based optimization program → local optimization problems (one per machine). – Messages between machines : Lagrange multipliers Stereo reconstruction from high-resolution image Handle large problems (more than 200 labels in images larger than 10 MPixel)

5. Related work Provable convergence while still being computationally tractable. – parallelizes convex belief propagation – conserves its convergence and optimality guarantees Strandmark and Kahl [24] – splitting the model across multiple machines GraphLab – assumes that all the data is stored in shared-memory

6. Related work Split the message passing task at hand into several local optimization problems that are solved in parallel. To ensure convergence we force the local tasks to communicate occasionally. At the local level we parallelize the message passing algorithm using a greedy vertex coloring

7. Message passing algorithm The joint distribution factors into a product of non-negative functions – defines a hypergraph whose nodes represent the n random variables and the subsets of variables x correspond to its hyperedges. Hypergraph – Bipartite graph : factor graph [11] one set of nodes corresponding to the original nodes of the hypergraph : variable nodes the other set consisting of its hyperedges : factor nodes N(i) : all factor nodes that are neighbors of variable node i

8. Message passing algorithm Maximum a posteriori (MAP) assignment Reformulate the MAP problem as integer linear program.

9. Message passing algorithm

10. Distributed convex belief propagation Partition the vertices of the graphical model to disjoint subgraphs – each computer solves independently a variational program with respect to its subgraph. The distributed solutions are then integrated through message- passing between the subgraphs – preserving the consistency of the graphical model. Properties : – If (5) is strictly concave then the algorithm converges for all ε >= 0, and converges to the global optimum when ε > 0.

11. Distributed convex belief propagation

13. Lagrange multipliers : – : the marginalization constraints within each computer – : the consistency constraints between the different computers

14. Distributed convex belief propagation

15. Experiment evaluation Stereo reconstruction – nine 2.4 GHz x64 Quad-Core computers with 24 GB memory each, connected via a standard local area network libDAI 0.2.7 [17] and GraphLAB [16]

16. Experiment evaluation

18. relative duality gap

19. Conclusion Large scale graphical models by dividing the computation and memory requirements into multiple machines. Convergence and optimality guarantees are preserved. Main benefit : the use of multiple computers.