ae0eb3eed39d2bcef4622b2499a05fe6-Reviews.html

Running machine learning algorithms on larger datasets is becomming more and more a necessity. Recently, a practically very relevant line of research has been to look at various programming paradigms for turning well known machine learning algorithms into distributed algorithms - meaning they can run on an infrastructure with no shared memory and slow communication between processing units. This paper introduces a well known pattern called "optimistic concurrency control" into the machine learning literature. As the authors point out, there has been some work on embarrasingly parallel algorithms, distributed algorithm using the locking paradigm and coordination-free approaches to distributed algorithms. Optimistic concurrency control is a technique which starts out by assuming that each individual processing unit can freely access shared state.