Parallel Algorithms and Generalized Frameworks for Learning Large-Scale Bayesian Networks

Student: Ankit Srivastava (Georgia Institute of Technology)
Supervisor: Srinivas Aluru (Georgia Institute of Technology)

Abstract: Bayesian networks (BNs) are an important subclass of graphical machine learning (ML) models that enable probabilistic reasoning about interactions between variables of interest. Their interpretability makes them an ideal model for making high-stakes decisions in fields where explainability is desirable. However, learning BNs with even few thousand variables using existing software libraries requires an infeasible amount of time. This has prevented BNs from becoming a viable alternative to other ML models. To address this, we have developed scalable high-performance libraries for learning large-scale BNs. In this poster, we present our work on parallelizing a variety of popular BN learning algorithms, including a method for constructing parameter-sharing specialization of BNs – module networks. Our experiments show that the optimized open-source implementations of our parallel algorithms reduce the time required for learning networks with tens of thousands of variables from multiple months to a few hours by efficiently utilizing thousands of cores.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing