Cloud Library for Directed Probabilistic Graphical Models.

摘要

The project aimed to build a massively parallel library for Bayesian networks by creating a data analytical capability with potential throughput commensurate with DoD data volumes. The goal was to implement data-parallel independent & identically distributed inference & learning in Bayesian networks & accomplish nearly-linear scaling. They re-examined & implemented data structures & algorithms needed for distributed-model inference. The inference aimed at being able to ask & answer privacy & adversarial learning questions where model distribution is due to private nature of the data. They looked for efficiently-parallelizable methods of inference & learning.