NeuroBE: NN Approximations to Bucket Elimination

Sakshi Agarwal, Kalev Kask, Alexander Ihler, Rina Dechter

Abstract: A major limiting factor in graphical model inference is the complexity of computing the partition function. Exact message-passing algorithms such as Bucket Elimination (BE) require exponentially high levels of memory to evaluate the partition function, therefore approximations are often investigated. In this paper, we build upon one such recently introduced methodology called Deep Bucket Elimination (DBE) that uses classical Neural Networks (NNs) to approximate messages generated by BE in cases when buckets have large memory requirements. In the new scheme, called NeuroBE, we personalize the construction and learning of these NNs by utilizing prior information about message size and distribution. Our experiments demonstrate that this new scheme provides significant improvements over DBE, yielding more accurate estimates with less time. We also study the impact of the local errors in message approximations on the global accuracy of the estimand of the partition function.