An AI Talk to understand AI Talks
when i was
at university my professor wanted me to
write a white paper
about the application of an algorithm
that belongs to the area of artificial
intelligence and at the beginning i was
a little bit skeptical
i mean how to find a feasible use case
can i handle the complexity will i still
be able to hang out with my friends a
lot of
important questions to clarify before i
started but then i thought
challenge accepted let’s do it i was
motivated i mean i was
really motivated i went home i opened up
my laptop
and the first question i had to look at
was
what the hell is artificial intelligence
right and now imagine you
are a student you are lazy as hell
there’s only
one way how to approach this and this is
you open up a search engine
like google and you type in dear google
please show me what you say
i and the result i was presented looked
like this
so the first sentence of my white paper
was artificial intelligence
is blue what i also learned is that ai
has something to do with
digital brains and robots that look like
humans i mean
in the end of the day i did not learn
anything at all by looking at these
results right and exactly this is the
problem
because to most people in society ai
is just blue and let me illustrate this
problem with a comparison to online
banking
i think most of you guys use online
banking
and i think most of you have a rough
idea how a bank actually works
like you’ve got a bank account you
deposit money
the bank rents the money to other people
you get paid interest rates
on the other hand millions of people are
using digital services
of google amazon facebook whatever
without being aware of their value of
the personal data
and the power of artificial intelligence
and this is a problem
and the only way i see to escape this
problem
is learning and the good news is there
are many great ways how to learn
nowadays for example
you can watch all those really really
fantastic ted talks
about ai which you can find on youtube
however
in order to understand what the speakers
are talking about
you need to have a certain understanding
about this topic and you need to be able
to ask the right
questions and exactly this is what i
want to deliver you today
so in my ted talk you will receive two
learnings number one
you will get a basic understanding of
modern
artificial intelligence and number two
you will be equipped with a simple
toolkit
consisting of three guiding questions
that you can apply
to each and every ai talk you’re going
to listen to
and now let’s get started and the first
thing we have to clarify
is the terminology because i very often
observe
that people mix up the terms and this
leads to confusion
so the main concept we talk about today
is artificial intelligence and this is a
very broad concept so
the official definition of ai is the
science and engineering
of making intelligent machines however
where does intelligence start where does
it end
all this is really up to your
interpretation and this leads to the
fact that
different people can have a different
perspective on the same topic
of ai for example if you talk to a
computer scientist
or group psychologist those guys can
have a completely different
understanding of ai and this is
absolutely okay because it is a broad
topic
however what most people mean when they
talk about ai today
is machine learning and machine learning
is about the imitation
of the cognitive ability to learn
so what you do you apply mathematics
and statistics to historic data with the
goal to identify
pattern in this data in order to
understand more recent data
or to create predictions and
in other words this means machine
learning is the perfect tool
to support rational decision making in
an uncertain
environment because it’s completely
based on math
and data and when i talk about ai today
i refer to machine learning based
artificial intelligence
and deep learning again is one subset of
machine learning and it does
pretty much the same so again we look at
historic data
apply math and statistics to find
pattern in order to create predictions
however the difference is the type of
algorithm that is applied so
whereas in classical machine learning we
apply pretty straightforward statistical
mathematical math methods whereas in
deep learning you apply something called
deep
neural networks and this is a very
sophisticated type of algorithm that is
certainly inspired by the human brain
but it does also certainly not work
the same way than a human brain works
so after being familiar with the
definition you want to know
what can i do with artificial
intelligence and again here are
different ways how to look at this topic
but i prefer those four application
clusters
let me start with computer vision so
computer vision is about recognizing
and interpreting the content of images
and an image is nothing else than a
piece of unstructured data right because
we can
break down an image into single pixels
and express each pixel with the value
and when we’ve got the numbers
we can apply the math and computer
vision for example is one of the
core technologies of autonomous driving
because together with other sensors
it helps the car to perceive its
environment
for example to identify other cars
pedestrians traffic signs whatever
in the field of vision in order to
derive the most
suitable follow-up action in the current
situation
what you can also do with computer
vision is to distinguish
a chihuahua from a muffin and if you
think this is an easy task
just have a look at this picture
so the one thing is the thing you want
to bite in
the other one is the thing that might
want to bite you
this is a small but very important
difference right
and as you can see not every problem
that sounds easy
is actually easy for ai to solve in the
end
of the day and now let me continue with
natural language processing so human
natural language is very complex for
machines
to understand it has got a broad
terminology
it has got many rules it has got
exceptions its context based and i did
not even start to talk about accents and
dialects right and
natural language processing is about the
understanding
interpretation and even generation of
human natural language for example think
about all those
automatic translation services where you
can translate the text from one language
to another language or think about those
smart voice assistants
like alexa or siri and as long as
a huge metal fan can either keep me cool
during a hot day or keep me awake
during a quiet night language will
remain a complex thing
for machines to understand
so now let’s have a look at data mining
data mining is about leveraging the
potential
of structured and semi-structured data
that you have captured just think about
a time
series analysis where you have got costs
or sales figures from the past and you
want to analyze how they will continue
or think about an automatic proper
customer analysis
where you try to provide the right
customer with the right information in
the
right point in time and this is heavily
applied in social media and is called
micro targeting
what you can also do and we are familiar
with this
are recommendation engines i mean think
about amazon or netflix right
so people who look at this product were
usually also interested in these
products
and i’ll just imagine i want to give an
example you visit an e-commerce
store and you search let’s say a
baseball
bat you know to hit some balls together
with your friends
and now imagine you are recommended to
buy
these items based on popular customer
combinations again
i have got two learnings for you number
one some people might have
interesting hobbies number two ai does
not always generate the results
that you have expected
and now let me finish with behavioral ai
and behavioral ai
is about the ability of autonomous
systems
to learn a successful strategy by trial
and error
just think about a smart agent in a
computer game this agent has to learn
how to win the game
by exploring different strategies and
one very impressive example
was presented by google by a google
company a couple of years ago
so alphago zero is a computer program
that learned how to play the board game
go and go is a very famous sport game
in asia it’s played with a 19 by 19 grid
board and what’s so special about this
game is the complexity
because the amount of possible
constellations on this board
surpasses the amount of atoms in the
universe
this is a really huge number
and this algorithm learned how to play
the game
basically by getting provided with the
rules
and by playing against itself so
exploring the right
strategy in the end of the day it was
even able to beat the south korean grand
master
in an official match but behavioral ai
is also
applied in robotics for example to teach
robots
the desired behavior i mean just think
about
assembly lines where you want robots to
assemble certain production items
in the right way
so now you know the definition of ai and
you know what you can do with ai
the last missing piece is how to create
ai
and let me start with the training
because this is the biggest difference
to classical computing because
ai is not programmed but trained
and for ai training we need three
ingredients
data the algorithm and computing power
and let me first quickly start with the
algorithm because
most people think the algorithm is the
most important
asset in the game you know this musical
algorithm that controls our lives
however an algorithm is nothing else
than a simple
cooking recipe that tells us how to
process data
to solve a certain mathematical problem
so it’s the math it’s there since the
1950s i can
read it in papers and books i can even
download it from the internet so yes
the algorithm is important but it’s not
the most important asset
what we also need is computing power and
looking back we observe
a strong increase of inexpensive
computing resources and this was
certainly
an accelerator however the most
important
part in the game is the data because
what you have learned already is
machine learning is based on historic
data
if we do not have this data we cannot do
machine learning right
but in times of internet of things
social media online shopping and so on
we produce
tons and tons of data every day and the
more data you have
the more fun you will have with machine
learning and when all three elements
come together the data the algorithm
plus the computing
an iterative training process starts
for example i provide the algorithm with
thousands of examples how to hours look
like on images
and after the training time what we get
is a
trained machine learning model and this
model is basically a
mathematical representation of the real
world
so it contains for example all the rules
how to identify a chihuahua
on an image the good thing is this model
can be so complex especially in deep
learning
i would never ever be able to come up
with a model of this complexity
in a reasonable amount of time when i do
it manually
however the downside because of this
complexity
i might not be able to understand how
the model works and how it will behave
in different situations so after the
training
we come to the inference phase and in
the inference phase
we need new data points so new images
of chihuahuas the algorithm has never
seen before again we apply
computing power but here not as much as
during
training i mean your iphone is doing
machine learning every day
so that’s very often sufficient and we
apply now
the trained model from the training
phase to the new data points and what we
get is a prediction
and this prediction could be i identify
a chihuahua on this image and i’m
confident to
97 so this tells us that machine
learning is a
probabilistic approach rather than a
deterministic approach
and this implies we always have a
certain amount of uncertainty
and depending on the use case we have to
think about
how much uncertainty we can accept
so congratulations now you are familiar
with the basics of modern
artificial intelligence and now i think
you guys are excited about
the toolkit the three questions so let’s
have a look at the first one
and the first question you should ask
yourself when you watch an ai talk is
is the speaker talking about artificial
narrow intelligence
or something superior like artificial
general intelligence and let me quickly
tell you uh what’s the difference so
narrow ai is all the stuff we have
talked about today
it’s a very specialized type of ai that
performs well
in a certain niche whereas the other
types like
general ai or super ai those types of
i come close to human level of
intelligence or even surpass
human level of intelligence for example
just think about this
smart small robot in the star wars movie
called
r2d2 and just think about this guy
had to save a jedi’s life but the only
thing it can do
is to distinguish a chihuahua and a
muffin i mean
this does not really work out well right
so
those types of ai what they can do they
can learn
new stuff independently and they can
apply
knowledge from one domain to another
domain
based on the world knowledge they have
and this is really human-like
and the reason why you should ask this
question so
if a speaker talks about narrow ai the
talk might be much
more based on concrete scientific
statistical mathematical
solutions and methods whereas if she
talks about something superior like
generally i or super ai
the talk might be much more on a
visionary or philosophical
level the speaker
actually talks about narrow ai the next
question is is she talking about machine
learning or other ai methods
so if it’s machine learning the good
thing is you can apply
all the fundamentals you have learned
already you know the pros
and you know the cons whereas if it’s
another method you first have to find
out about this method
you have to understand how it works in
order to be able to understand the talk
so this is also very important
because machine learning is one subset
of ai a very famous one at the beginning
at the moment
but it’s not the only one and the last
question is
if she talks about machine learning is
the funnel from data
to decision observable so in a high
quality talk
the speaker will always make transparent
on which data the algorithm was trained
on
and which method was applied right and
only with this information you can
really scope the solution the speaker
talks about
and you can really think about how well
it can perform in which kind of
situation
whereas if ai is represented as black
box
you might always have problems to fully
understand the solution
so artificial intelligence
is important because it will impact
our private lives and it will change the
way
how we work so learning about ai
is key for everybody and if you decide
to do so
by watching all those great tedx talks
and if you apply
the fundamental knowledge of my talk
plus the three guiding questions
you will definitely understand much more
about ai
than this thank you