Parallelizing a Convergent Approximate Inference Method

Probabilistic inference in graphical models is a prevalent task in statistics and artificial intelligence. The ability to perform this inference task efficiently is critical in large scale applications. The ever-evolving parallel computing technology suggests that dramatic speedup might be achieved by appropriately mapping the current inference algorithms to a parallel framework. Parallel exact inference methods still suffer from worst-case exponential complexity. Approximate inference methods have been parallelized and good speedup achieved. In this report, we focus on a variant of the Belief Propagation algorithm. This variant has better convergence properties and is provably convergent under certain condition. We show that this method is amenable to coarse-grained parallelization and propose techniques to parallelize it optimally without sacrificing convergence. Experiments on a shared memory system demonstrate that near-ideal speedup is achieved with reasonable scalability. Thi...

• Type: Journal
• Year: 2011
• Where: AI
• Authors: Ming Su, Elizabeth Thompson
• Keywords: AI 2011, Artificial Intelligence, Belief Propagation, Dramatic Speedup, Inference Algorithms