NOMAD : A Framework for Distributed Machine Learning

Опубликовано 8 июля 2016, 0:43

In this talk I will describe NOMAD, which is an asynchronous, distributed framework for large scale matrix factorization. A unique feature of our framework is the ability to keep the network and CPU busy by simultaneously performing communication and computation. Extensive experiments using public datasets confirm that NOMAD is able to outperform other algorithms both in a multi-core as well as distributed memory setting. If time permits, I will also discuss some preliminary results on extending our framework to solve a larger class of regularized risk minimization problems.