Parallel and Distributed Systems for Large-Scale Machine Learning

• Activity Number: 157
• Type: Invited
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Carlos Guestrin*
• Companies: University of Washington
• Keywords: Machine Learning ; Distributed Systems ; Parallel Algorithms ; Out-Of-Core Computation ; Graphical Models

Abstract

One of the biggest challenges in statistical learning today is the scale of data we must handle. Scaling up methods requires us to develop new distributed algorithms; these algorithms must be informed by the constraints of the underlying computer systems we are using. In this talk, we will discuss three key concepts in developing distributed ML methods (data partitioning, asynchrony and data layout), the underlying systems requirements that motivate the focus on these concepts, and the impact they have on the design of distributed ML methods. We will motivate the methods from the perspective of real-world, large-scale applications.