Learning Tractable Probabilistic Models from Inconsistent Local Estimates