Abstract:
|
Various narrow artificial intelligence architectures are on the rise due to the
development of Graphics Processing Units and, thus, computational capabilities. Massive
number multiplication capabilities of GPUs enabled researches to create more
complicated and advanced algorithms. Initially, a gaming hardware became a base for
modern time Industrial Revolution.
Machine learning, once a forgotten branch of computer science, attracts huge investments
and interest. In 2014, Google acquired an UK-based start-up Deep Mind for over £400M.
In 2016 Volkswagen invested $680M in autonomous vehicle and cyber security start-ups
(1). Same year Microsoft announced a newly created AI fund (2) and in May this year it
resulted in investment of $7.6M in Bonsai, an AI start-ups that hopes to help companies to
integrate machine learning in the infrastructure (3).
It seems that almost never-ending pockets of investors are motivated by a promise of
automation of difficult tasks, which, until now, have never been performed by humans.
This thesis explores various supervised machine learning algorithms, beginning with
the simplest k-Nearest Neighbours and Multi-layer Perceptron, to the state of the art
architecture created by the industry experts (Deep Residual Network from Microsoft
Research), and prominent academic figures (i.e. GG from Oxford).
Furthermore, the author of the thesis proposes two additional network structures,
named Deep Inception and Stacked Artificial Residual Architecture, inspired by previously
mentioned research. |