Technical Perspective: Implicit Parallelism through Deep Language Embedding

Abstract

Modern “big data” analysis was motivated by the needs of the large Internet players, but it was enabled by two main technical developments: parallel data processing technologies that support reliable and scalable computation over unreliable shared-nothing clusters of computers, and continued advances in machine learning algorithms and techniques. Initial work on these two areas happened largely independently: MapReduce was developed for aggregate computations over large multitudes of records, with minimal control flow and no evident goal of supporting machine learning. Conversely, many of the advances in machine learning research targeted a single machine.