Training machine learning algorithms using actor framework in multi-core or cluster system

Abstract

Machine learning algorithms have shown great promises in many applications, the increase of data has fueled the production of great frameworks that could leverage the combination of data with tools. There have been focus on providing applications to ease the process of training different learning models like Apache Spark, Tensor flow that unifies streaming, batch, and interactive big data workloads. The Actor model is asynchronous message passing protocol for computations in distributed systems and is suitable for exploiting large-scale parallelism. The support that actor system brings with it like fault tolerance, actor spawning, multi-core usage makes it a good model for building Machine learning applications that wants to benefit in local and distributed cloud computing.

This thesis provides one such framework for iterative machine Learning algorithms that train the models by asynchronous message passing. We show results of a number of ML algorithms and analyze the benefits of using a scalable (horizontal and vertical) platform of actors to do large-scale data model training. The thesis implements two protocols to do scalable training and compares their performance.