Recent Developments in Distributed Machine Learning
Michael I. Jordan
University of California, Berkeley

The field of statistical machine learning grown rapidly in recent years, and is a currently thriving engineering discipline, 
having significant impact on numerous other academic disciplines and on industry.  This growth has revealed some weaknesses in 
the classical machine learning paradigm, however, most notably in the centralized nature of many machine learning algorithms.  
Many large-scale problems are of a distributed nature, and it would be desirable to avoid transmitting data to a central location 
for analysis.  In this talk I discuss some recent progress on this problem. The basic approach is to characterize the tradeoff 
between partial transmission of data and the loss of accuracy incurred, and to optimize along that tradeoff. This approach also 
yields insights into methods for solving very large centralized problems, where the tradeoff involves computation/accuracy 
instead of communication/accuracy.

[Joint work with Ling Huang, XuanLong Nguyen, Martin Wainwright and Donghui Yan]