MAS: a Multiplicative Approximation Scheme for Probabilistic Inference Spotlight T51
Ydo Wexler & Christopher Meek - Microsoft Research · Fast and accurate approximate inference for factored probabilistic models for approximating likelihood, MAP and marginal probabilities · Partial computations are approximated with relative error guarantee that are translated into bounds on the result Computation graph
1 1   max  log Pe ~  1   max log P e

Error doesn't blow up! · Optimization of local errors · Fast optimal closed-form solution for L2 · Dynamic decision on local approximations · Significantly improves over generalized belief propagation and Mini-buckets