Skymind Wiki: Part 1

An Introduction to Artificial Intelligence

Chris V. Nicholson
Identification of popular use cases

This book is a beginner’s guide to important topics in AI, INTRODUCTION 1

machine learning, and deep learning. The goal is to give readers
DEEPLEARNING, MACHINE LEARNING 2
an intuition for how powerful new algorithms work and how & AI USE CASES
they are used.
Feature Introspection 2

Advances in the field of machine learning (algorithms that Text 2

adjust themselves when exposed to data) are driving progress
Images 3
more widely in AI. But many terms and concepts will seem
incomprehensible to the intelligent outsider, the beginner, and ARTIFICIAL INTELLIGENCE (AI) VS. 3
MACHINE LEARNING VS. DEEP
even the former student of AI returning to a transformed
LEARNING
discipline after years away. We hope this book helps you better
understand AI, the software used to build it, and what is at stake Artificial Intelligence, Symbolic AI and GOFAI 3
in its development.
Machine Learning: Programs That Alter 4
Themselves
The line between mathematics and philosophy is blurry when
Deep Learning: More Accuracy, More Math & 5
we talk about artificial intelligence, because with AI, we ask the More Compute
mineral called silicon to perceive and to think – actions once What’s Next for AI? 5
thought exclusive to meat, and now possible with computation.
We hope that by reading this book, you will find new ways of CONCLUSION 6

thinking about life and intelligence, just as we have by writing it.

PAGE 1 (May 2019)

 DEEP LEARNING, Feature Introspection
MACHINE LEARNING & AI Traditional machine learning has the advantage of feature
USE CASES introspection – that is, it knows why it is classifying an
input in one way or another, which is important for
Deep learning excels at identifying patterns in unstructured analytics. But that very advantage is what excludes it from
data, which most people know as media such as images, working with unlabeled and unstructured data, as well as
sound, video and text. Below is a list of sample use cases attaining the record-breaking accuracy of the latest deep
we’ve run across, paired with the sectors to which they learning models. Feature engineering is one of the major
pertain. chokepoints of traditional machine learning, since so few
people are able to do it well and quickly enough to adapt to
changing data.

For cases where feature introspection is necessary (e.g.

the law requires that you justify a decision to, say, deny a
loan due to predicted credit risk), we recommend using a
deep net in an ensemble with machine-learning algorithms,
allowing each one to vote and relying on each for its
strength. Alternatively, one can perform various analyses
on the results of deep nets to form hypotheses about their
decision-making.

Text

NAMED-ENTITY RESOLUTION
One use of deep-learning networks is named-entity
recognition, which is a way to extract from unstructured,
unlabeled data certain types of information like people,
places, companies or things. That information can then be
stored in a structured schema to build, say, a list of
addresses or serve as a benchmark for an identity
validation engine.

SPEECH-TO-TEXT
With the proper data transforms, a deep network is
capable of understanding audio signals. This can be used
to identify snippets of sound in larger audio files, and
transcribe the spoken word as text.

PAGE 2 (May 2019)

 Image Artificial Intelligence,
Symbolic AI and GOFAI
OBJECT RECOGNITION
Object recognition is an algorithm’s ability to identify
arbitrary objects – such as balls, animals, or even faces –
within larger images. This is typically used in engineering
applications to identify shapes for modeling purposes. It’s
also used by social networks for photo tagging. Facebook’s
Deep Face is a good example of a deep-learning
application in this realm.

MACHINE VISION + NATURAL LANGUAGE PROCESSING

Advances in reality capture and reality computing are

making virtual and real worlds converge. One application of
John McCarthy, widely recognized as one of the godfathers
deep learning to this newly available data is to recognize
of AI, defined it as “the science and engineering of making
and label objects in 3D environments, and in real life.
intelligent machines.”
From there, it’s a short step to simulated semantics, in
which machines learn the nature and constraints of objects
Here are a few other definitions of artificial intelligence:
in the world, through their virtual representations, and then
bring that understanding to the language they generate A branch of computer science dealing with the
and ingest. We believe that is one of many futures in store simulation of intelligent behavior in computers.
for neural nets. The capability of a machine to imitate intelligent
human behavior.
A computer system able to perform tasks that
ARTIFICIAL normally require human intelligence, such as visual
INTELLIGENCE (AI) VS. perception, speech recognition, decision-making,
and translation between languages.
MACHINE LEARNING VS.
There are a lot of ways to simulate human intelligence, and
DEEP LEARNING some methods are more intelligent than others.

You can think of deep learning, machine learning and AI can be a pile of if-then statements, or a complex
artificial intelligence as a set of Russian dolls nested within statistical model mapping raw sensory data to symbolic
each other, beginning with the smallest and working out. categories. The if-then statements are simply rules
Deep learning is a subset of machine learning, and explicitly programmed by a human hand. Taken together,
machine learning is a subset of AI, which is an umbrella these if-then statements are sometimes called rules
term for any computer program that does something engines, expert systems, knowledge graphs or symbolic AI.
smart. In other words, all machine learning is AI, but not all Collectively, these are known as Good, Old-Fashioned AI
AI is machine learning, and so forth. (GOFAI).

The intelligence that rules engines mimic could be that of

an accountant with knowledge of the tax code, who takes
information you feed it, runs the information through a set
of static rules, and gives your the amount of taxes you owe
as a result. In the US, we call that TurboTax.

PAGE 3 (May 2019)

 Usually, when a computer program designed by AI
researchers actually succeeds at something – like winning
at chess – many people say it’s “not really intelligent”,
because the algorithm’s internals are well understood. The
critics think intelligence must be something intangible, and
exclusively human. A wag would say that true AI is
whatever computers can’t do yet.

Machine Learning: Programs

That Alter Themselves
Samuel taught a computer program to play checkers. His
goal was to teach it to play checkers better than himself,
Machine learning is a subset of AI. That is, all machine which is obviously not something he could program
learning counts as AI, but not all AI counts as machine explicitly. He succeeded, and in 1962 his program beat the
learning. For example, symbolic logic – rules engines, checkers champion of the state of Connecticut.
expert systems and knowledge graphs – could all be
described as AI, and none of them are machine learning. The “learning” part of machine learning means that ML
algorithms attempt to optimize along a certain dimension;
One aspect that separates machine learning from the i.e. they usually try to minimize error or maximize the
knowledge graphs and expert systems is its ability to likelihood of their predictions being true. This has three
modify itself when exposed to more data; i.e. machine names: an error function, a loss function, or an objective
learning is dynamic and does not require human function, because the algorithm has an objective. When
intervention to make certain changes. That makes it less someone says they are working with a machine-learning
brittle, and less reliant on human experts. algorithm, you can get to the gist of its value by asking:
What’s the objective function?
A computer program is said to learn
from experience E with respect to
How does one minimize error? Well, one way is to build a
some class of tasks T and
framework that multiplies inputs in order to make guesses
performance measure P if its
as to the inputs’ nature. Different outputs/guesses are the
performance at tasks in T, as
product of the inputs and the algorithm. Usually, the initial
measured by P, improves with
guesses are quite wrong, and if you are lucky enough to
experience E.
have ground-truth labels pertaining to the input, you can
–Tom Mitchell measure how wrong your guesses are by contrasting them
In 1959, Arthur Samuel, one of the pioneers of machine with the truth, and then use that error to modify your
learning, defined machine learning as a “field of study that algorithm. That’s what neural networks do. They keep on
gives computers the ability to learn without being explicitly measuring the error and modifying their parameters until
programmed.” That is, machine-learning programs have they can’t achieve any less error.
not been explicitly entered into a computer, like the if-then
They are, in short, an optimization algorithm. If you tune
statements above. Machine-learning programs, in a sense,
them right, they minimize their error by guessing and
adjust themselves in response to the data they’re exposed
guessing and guessing again.
to (like a child that is born knowing nothing adjusts its
understanding of the world in response to experience).

PAGE 4 (May 2019)

 Deep Learning: More Accuracy, What’s Next for AI?
More Math & More Compute
The advances made by researchers at DeepMind, Google
Deep learning is a subset of machine learning. Usually, Brain, OpenAI and various universities are accelerating. AI
when people use the term deep learning, they are referring is capable of solving harder and harder problems better
to deep artificial neural networks, and somewhat less than humans can.
frequently to deep reinforcement learning.
This means that AI is changing faster than its history can
Deep artificial neural networks are a set of algorithms that be written, so predictions about its future quickly become
have set new records in accuracy for many important obsolete as well. Are we chasing a breakthrough like
problems, such as image recognition, sound recognition, nuclear fission (possible), or are attempts to wring
recommender systems, natural language processing etc. intelligence from silicon more like trying to turn lead into
For example, deep learning is part of DeepMind’s well- gold?
known AlphaGo algorithm, which beat the former world
champion Lee Sedol at Go in early 2016, and the current There are four main schools of thought, or churches of
world champion Ke Jie in early 2017. A more complete belief if you will, that group together how people talk about
explanation of neural works is here. AI.

Deep is a technical term. It refers to the number of layers Those who believe that AI progress will continue apace
in a neural network. A shallow network has one so- tend to think a lot about strong AI, and whether or not it is
called hidden layer, and a deep network has more than one. good for humanity. Among those who forecast continued
Multiple hidden layers allow deep neural networks to learn progress, one camp emphasizes the benefits of more
features of the data in a so-called feature hierarchy, intelligent software, which may save humanity from its
because simple features (e.g. two pixels) recombine from current stupidities; the other camp worries about the
one layer to the next, to form more complex features (e.g. a existential risk of a superintelligence.
line). Nets with many layers pass input data (features)
through more mathematical operations than nets with few Given that the power of AI progresses hand in hand with
layers, and are therefore more computationally intensive to the power of computational hardware, advances in
train. Computational intensivity is one of the hallmarks of computational capacity, such as better chips or quantum
deep learning, and it is one reason why a new kind of chip computing, will set the stage for advances in AI. On a
call GPUs are in demand to train deep-learning models. purely algorithmic level, most of the astonishing results
produced by labs such as DeepMind come from combining
So you could apply the same definition to deep learning different approaches to AI, much as AlphaGo combines
that Arthur Samuel did to machine learning – a “field of deep learning and reinforcement learning. Combining deep
study that gives computers the ability to learn without learning with symbolic reasoning, analogical reasoning,
being explicitly programmed” – while adding that it tends Bayesian and evolutionary methods all show promise.
to result in higher accuracy, require more hardware or
training time, and perform exceptionally well on machine Those who do not believe that AI is making that much
perception tasks that involved unstructured data such as progress relative to human intelligence are forecasting
blobs of pixels or text. another AI winter, during which funding will dry up due to
generally disappointing results, as has happened in the
past. Many of those people have a pet algorithm or
approach that competes with deep learning.

PAGE 5 (May 2019)

 Finally, there are the pragmatists, plugging along at the
math, struggling with messy data, scarce AI talent and user
acceptance. They are the least religious of the groups
making prophesies about AI – they just know that it’s hard.

Footnotes

It’s interesting to note that even when certain technologies

are physically impossible, they can still be regulated. In
1404, during the reign of Henry IV, the English parliament
passed a law called the “Act against multiplication”,
outlawing the creation of gold and silver from other
materials, since that was seen as a threat to the throne’s
control over the currency. The law was later modified to
allow only certain people to create gold and silver through
alchemical processes, until it was finally repealed in the
17th century. Regulations outlawing strong AI, a technology
that may or may not be possible, and for which there exists
no strong theoretical foundation, would be similarly absurd.

Skymind offers AI infrastructure that enables corporate data scientists and IT

teams to rapidly prototype, deploy, maintain, and retrain machine learning
workflows that accelerate time to value. Skymind bridges the gap between data
science, DevOps and the big data stack.

PAGE 6 (May 2019)