Foreword
Several technological drivers that arose in the last decade that I made big data possible,
irreversibly reshaping the world completely. Among these, machine learning plays a
singular role, since it provides the main functionalities required for data analysis, mining,
knowledge discovery, and many other features that provide actionable autonomous
intelligence in a way invisible but pervasive to most of the systems we use in our daily
lives. Although not new, the formalisms and methods within machine learning have
quickly evolved, driven by the growing demands of e-commerce, social networks, internet-
related services and products, and similar enterprises centered on online business.
Breakthroughs in machine learning have been fueled by other technological innovations
sprouted and matured within the Hadoop ecosystem, including horizontally scalable
computational resources and superior warehousing capabilities that have made the real-
time analysis of huge datasets feasible. At the same time, community supported initiatives
around the Python programming language have made the use and evolution of
sophisticated analysis libraries widspread, giving rise to a remarkable amount of
knowledge and experience, at the same time fast and easy to deploy and to put into
production.
Within machine learning, neural networks play a singular role nowadays. Perhaps the first
artificial intelligence paradigm to be proposed more than 70 years ago, neural networks
have experienced several cycles of being abandoned by the community only to be
rediscovered some years later. This was likely due to the lack of computational power to
perform really complex analysis adequately, together with the burdensome task of
assembling, training, and testing different topologies by trial and error. This has changed
dramatically in recent years, mostly due to the availability of cloud computing, GPUs, and
programming libraries that allow the set up of networks with simple scripts. Today, a
network with hundreds of millions of degrees of freedom can be assembled in minutes,
trained in hours, and put into production in a few days, (obviously, if you know the right
technologies to so). This is one of the reasons why most of the radical advancements in
computer vision, language understanding, and pattern recognition in general are being
driven specifically by different flavors of neural networks that have been proposed recently.
