EE Seminar: Distributed Machine Learning

~~(The talk will be given in English)

Dr. Ohad Shamir
Weizmann Institute
Monday, May 18th, 2015
15:00 - 16:00
Room 011, Kitot Bldg., Faculty of Engineering
Distributed Machine Learning

Abstract
Handling increasingly large datasets is one of the major problems faced by machine learning today. One approach is to distribute the learning task, and split the data among several machines which can run in parallel. Ideally, a distributed learning algorithm on k machines should provably (1) Run k times faster than an algorithm designed for a single machine; (2) Reach the same statistical learning performance with the same amount of training data; And (3) Use minimal communication between the machines, since it is usually much slow than internal processing. In other words, such an algorithm should combine computational efficiency, statistical efficiency, and communication efficiency. In this talk, I'll survey the challenges of designing such algorithms for convex learning problems, and describe some recent advances as well as fundamental limitations.

Includes joint work with Andrew Cotter, Ofer Dekel, Ran Gilad-Bachrach, Nathan Srebro, Karthik Sridharan, Lin Xiao and Tong Zhang.