The Bridge - Issue 2, 2018 - 7

determines it has succeeded based on how tightly
the groups are packed, and how many outliers
are present. In this way, unsupervised learning is
typically very good at finding patterns in the data,
though with complex datasets, the pattern found
may be difficult to interpret. While this can make it
unsuitable for some of the problems in supervised
and reinforcement learning, it does have some
very profound applications. For example, it can
be used to divide and conquer other problems
such as the combinatorically-demanding Traveling
Salesman Problem. Unsupervised learning was used
as a heuristic to divide the problem and achieve a
dramatic speedup over the best previous solution.
[5] Whether the final processing is done by another
algorithm or a human, reducing the complexity
of data analysis is among many applications of
unsupervised learning.
Overall, each method to machine learning has
its own strengths and weaknesses. It is clear that
humans exhibit traits from each, being able to
learn from either a teacher (supervised), or trial
and error (reinforcement), or being able to classify
objects based on their similarities to find patterns
(unsupervised). While human level AI may be
far away, these methods can be applied to most
problems with very good results.

WHY WILL MACHINE LEARNING
OUTLIVE MOORE'S LAW?
While Moore's law has certainly helped machine
learning, it is not needed for the continued growth
of the field. Algorithmic advances are continually
improving the field, with new methods for learning
along with better parallel processing constituting
the most significant increase in performance.
Furthermore, increasingly specialized hardware that
focuses on the operations needed for machine
learning has led to tremendous growth.

Garry Kasporov, in 1997. Deepmind's AlphaGo Lee
defeated Lee Sedol, 18-time world champion of
Go, in 2016. Chess has a state space complexity
of 10^47, while Go has a state space complexity
of about 10^170. This means in those 19 years,
computational efficiency would have increased
by 123 orders of magnitude, or 297,600,000%
per year. While this figure is very approximate,
it highlights just how impressive the growth of
machine learning is.
To further highlight the growth of AlphaGo
particularly, the original version used 176 GPUs, and
required months of training. When it defeated Lee
Sedol, it had switched to 48 Tensor Processing Units
(TPUs), which are optimized for machine learning.
Just months later Deepmind launched AlphaGo Zero,
which started learning without any human game
data, and in 3 days was stronger than the version
that beat Lee Sedol (AlphaGo Lee), and ultimately
after 40 days was fully trained, far outmatching any
previous results. This newest version only runs on 4
TPUs, and even despite this is many times stronger
then AlphaGo Lee.
Just as algorithmic advances have accelerated
machine learning, newer, more optimized hardware
has made its impact as well. When self-driving car
research took off around 2010, GPU acceleration
had started to become mainstream. Since then,
programs that could utilize these resources became
more common, and has since become the defacto standard for machine learning. The more
specialized TPU that is aimed solely at deep learning
applications has finally stepped into consumer
grade products with Nvidia's release of the Titan V,
which claims up to 110 TFlops of compute power in
deep learning applications. This shows a large step
towards creating specialized hardware specifically
for machine learning, and the potential demand for
such systems.

To put this in perspective, IBM Deep Blue managed
to defeat the reigning world champion of chess,

HKN.ORG

7
http://www.HKN.ORG
Table of Contents for the Digital Edition of The Bridge - Issue 2, 2018
