BIGDATA: F: DKA: Collaborative Research: Theory and Algorithms for Parallel Probabilistic Inference with Big Data, via Big Model, in Realistic Distributed Computing Environments

点击翻译按钮获取中文摘要

This project develops a new framework that enables machine learning (ML) systems to automatically comprehend and mine massive and complex data via parallel Bayesian inference on large computer clusters. The research has a profound impact on the practice and direction of Big Learning. The developed technologies have a catalytic effect on both ML research and applications: ML scientists are able to rapidly experiment on novel, cutting-edge ML models with minimal programming effort, unhindered by the limitations of single machines. Researchers from other fields, like biology and social sciences, are able to run contemporary advanced ML methods that transcend the capabilities of simple models, yielding new scientific insights on data whose size would otherwise be daunting. Data scientists at small start-ups are able to conduct ML analytics with complex models, putting their capabilities on par with huge companies possessing dedicated engineering and infrastructure teams. Students and beginners are able to witness distributed ML in action with just a few lines of code, driving ML education to new heights. Technically, this research focuses on scaling up and parallelizing Bayesian machine learning, which provides a powerful, elegant and theoretically justified framework for modeling a wide variety of datasets. The research team develops a suite of complementary distributed inference algorithms for hierarchical Bayesian models, which cover most commonly used Bayesian ML methods. The project focuses on combining speed and scalability with theoretical guarantees that allow us to assess the accuracy of the resulting methods, and allow practitioners to make trade-offs between speed and accuracy. Rather than focus on a few disconnected models, the project develops techniques applicable to a broad spectrum of hierarchical Bayesian models, resulting in a toolkit of building blocks that can be combined as needed for arbitrary probabilistic models - be they parametric or nonparametric, discriminative or generative. This is in contrast to much existing work on parallel inference, which tends to focus on parallelization in a specific model and cannot be easily extended. The project provides a solid algorithmic foundation for learning on Big Data with powerful models. The research contributes to democratizing advanced and large-scale ML methods for broad applications, by offering the user and developer community a library of general-purpose parallelizable algorithms for working on diverse problems using computer clusters and the cloud, bridging the gap between practical needs from data and basic research in ML.