Scaling up Bayesian variational inference using distributed computing clusters.

Andrés R. Masegosa, Ana M. Martínez, Helge Langseth, Thomas D. Nielsen, Antonio Salmerón, Darío Ramos-López, Anders L. Madsen

Abstract: Highlights • Bayesian inference over modern data sets is challenging because data does not fit into the main memory of a computer. • Stochastic methods have been so far the main approach to scale up Bayesian inference for this problem. • Apache Spark or Apache Flink provide a simple framework for efficiently processing distributed data. • This paper proposes alternatives approaches for scaling up Bayesian inference using modern distributed computing frameworks. • This approach is derived from new theoretical insights from variational message passing algorithms. Abstract In this paper we present an approach for scaling up Bayesian learning using variational methods by exploiting distributed computing clusters managed by modern big data processing tools like Apache Spark or Apache Flink, which efficiently support iterative map-reduce operations. Our approach is defined as a distributed projected natural gradient ascent algorithm, has excellent convergence properties, and covers a wide range of conjugate exponential family models. We evaluate the proposed algorithm on three real-world datasets from different domains (the Pubmed abstracts dataset, a GPS trajectory dataset, and a financial dataset) and using several models (LDA, factor analysis, mixture of Gaussians and linear regression models). Our approach compares favorably to stochastic variational inference and streaming variational Bayes, two of the main current proposals for scaling up variational methods. For the scalability analysis, we evaluate our approach over a network with more than one billion nodes and approx. 75 % latent variables using a computer cluster with 128 processing units (AWS). The proposed methods are released as part of an open-source toolbox for scalable probabilistic machine learning ( http://www.amidsttoolbox.com ) Masegosa et al. (2017) [29] . Previous article in issue Next article in issue