Development of Large-Scale Tensor Libraries with Applications to Machine Learning

Abstract

Program Manager: Predrag Neskovic, Math, Computer & Information Sciences. Data deluge is occurring in every conceivable domain and handling “big data” is one of the primary research challenges of this decade. This proposal considers development of fast algorithms for a variety of challenging machine learning tasks. These algorithms are based on a unified framework of tensor decomposition. Tensors are higher order generalizations of matrices, and can encode rich information about underlying relationships in the collected data. We aim to develop fast and scalable tensor decomposition algorithms and apply them on a range of exciting applications, including learning latent variable models and training neural networks. IntellectualMerit: The tools techniques exploited in the proposal are a result of cross-pollination between a number of areas in machine learning and statistics such as probabilistic models, nonconvex optimization, tensor algebraicmethods, graphicalmodels, and so on. This inter-disciplinary fusion of approaches will lead to novel algorithms for a variety of machine learning tasks. The PI is uniquely placed to develop large-scale open source libraries that will have a huge impact, both in academia and industry. Broader Impact: This proposal has the potential to completely transform the way probabilistic models and deep learning methods are being deployed. It will make training computationally far less expensive, as well as be trained using fewer number of samples. This will accelerate the adoption of sophisticated machine learning techniques into newer domains and applications. Tensor methods for learning latent variable models will have applications in a number of domains, currently overwhelmed by the data deluge. This includes efficient methods for discovering, monitoring, and reasoning about social networks, reasoning about biological pathways, predicting drug efficacy, improving computer vision, speech recognition, text understanding, monitor student learning in online courses, and so on. Large-scale tensor decomposition tools developed via this project will be released publicly, allowing widespread dissemination to researchers and practitioners in a range of fields. Key Words: Tensor methods, open source libraries, large-scale machine learning, latent variable

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 12, 2016
Source ID
N000141512737

Entities

People

  • Anima Anandkumar

Organizations

  • Naval Information Warfare Center Pacific
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Linear Algebra
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks