
today I’m going to talk about AI why
it’s important why it’s important for
everybody in the room and why diversity
matters in AI and so imagine a world
where you can just think of something
and it happens you’re kind of won and
won with computers and they
automatically do what you’re thinking so
maybe you’re hungry and you want to
order food and it automatically comes to
your door and it might even be prepared
by machines and flown with drones to
your house because it knows where you
are and gets the exact dish the way you
like it
prepared for you right when you need it
stuff like that is not too far off
and it needs to learn from you and learn
from data to be able to do that kind of
thing and so I was struck by this video
this got me into machine learning so
this supposed to be a video and being
playing but it’s completely generated by
artificial intelligence and I saw this
in right after my second year in this
engineering science program which is a
pretty cool program the first two years
you take every discipline of engineering
and the last two you pick one out of
eight different options and when I was
deciding one of Jeff Hinton’s PhD
students showed me this so Jeff is kind
of the pioneer of a field of artificial
intelligence called neural networks and
I’ll teach you how those work in this
talk but when I went to this PhD student
who happened to be my resident advisor
on the floor I was living in at
University of Toronto he showed me this
and said it was completely generated by
neural networks and I was blown away
because I knew how to program at that
time and there was no way I could write
functions and loops or any kind of
regular programming to create something
like this create a realistic video so I
had to learn more and I took the
computer option and took Jeff Hinton’s
machine learning course to learn more
about it and did my undergrad thesis
with Geoff Hinton and that was a great
learning experience I got a little taste
of machine learning but I didn’t know
enough to start a company and it’s been
my childhood dream to start a company in
the tech space but I knew I was hooked
on machine learning and so I took a PhD
at New York University and that’s what
brought me down to the United States and
I’m down there right now in New York
where I started clarified
and so if you think of artificial
intelligence it’s everywhere around us
already
it’s an apps you use every single day
whether it’s Siri when you talk into
your phone and it turns it into text
it’s automatically understanding that
with neural networks type of artificial
intelligence or I heard Spotify playing
at lunch those playlists that are
automatically created that’s another use
case for artificial intelligence which
is based on your past listening habits
and other people that are like you
similarly for buying products on Amazon
other retailers they understand in
aggregate from multiple users buying
stuff over time how you might be
interested in other products as you’re
shopping and then with image recognition
and video recognition we can now
understand who’s in the image or what’s
in the image the dog the tree the
mountain etc all automatically just from
the pixels and that’s what we do at
clarify and so let’s get it fast so
there’s a lot of really big problems for
humanity that artificial intelligence
can also help out with maybe it can
understand how you like your temperature
in your office or in a place like this
the optimize how everybody feels while
minimizing energy to affect climate
change or optimize the crop yields that
farmers are producing in order to
minimize world hunger or even drug
discovery finding new drugs that can
help cure global health issues and all
these other important issues like human
trafficking gun violence human rights
they can all be improved
with artificial intelligence so it’s a
broadly applicable new way of
programming computers to act as if they
were humans or like humans or even
smarter than humans well why No
why is it so important why are you
hearing so much about it in the news and
hopefully in school well it’s actually
starting to work these algorithms have
been in progress since the 70s and 80s
when people like chef Hinton were
actually in their school in their PhD is
coming up with these algorithms they’re
almost exactly the same algorithms from
back then many decades ago and they
actually work today in real world
applications and that happens because of
two main factors one is lots of data
everyone
has a cell phone in their pocket right
now and they can take pictures they can
create text messages they can talk into
it that’s all data that can feed into
these algorithms to teach them how
people interact and computation so now
these same algorithms that were created
back then with really slow computers now
have orders of magnitude more
computation and we even use these
graphics cards which are made for video
games to process these machine learning
algorithms because they’re really
parallel and a lot of the things you do
when you’re thinking are really parallel
so they’re perfectly suited for
artificial intelligence so data and
computation make these old algorithms
actually work and they work really
really well so how does it work it’s not
that much magic it’s actually pretty
structured mathematical operations which
I’ll get into now and it works just like
a human does so does anybody know what
this little guy is called I think I
heard somebody shove it yes okay so one
person said axolotl and that’s correct
so does everybody can everybody repeat
that axolotl awesome okay
what’s this called awesome and how about
these
all right so you learned from one
picture what this creature is and you
realize that even though there’s you
know black one’s different colored one’s
different poses different lighting
conditions different viewpoints from the
camera they’re all axolotls so it didn’t
take much to teach a human how to learn
a new concept and this is a visual
concept but you can hear new things you
can read new things you can taste the
things you learn very quickly as humans
and so to simulate that this is the type
of algorithm we use in computers and
this is a type of neural network that is
perfectly suited for images and video
understanding and there’s different
types different architectures of neural
networks that are good for audio and
good for text reading and good for
financial data and so forth but neural
network is just a computer algorithm
that’s all it is that tries to simulate
how the brain works nobody really knows
exactly how the brain works but this is
a good approximation and it shares a lot
of the same properties and so a neural
network is a bunch of layers of
processing and you get to configure them
in however you want to map some input in
this case the pixels of an image to some
output in this case the categories
people romance grass and house and so to
teach this algorithm how to recognize
these different things you have to feed
it data data that people have labeled
like axolotl manually and so that’s the
pixel and in this case person our people
as category and these different layers
of operations sound complicated but
they’re pretty quite simple so
convolution so this is a big word but it
all it means is template matching so
it’s gonna look at that little white
patch that’s the template and it’s gonna
slide it over every location in the
image and when it matches very closely
to the pixels at that location it gives
a strong activation in that feature map
so that that blue square that is the
output of the convolution layer creates
a feature map the same size as the image
so it goes every location the stronger
the activation the higher the value you
have in that feature map you can see
there’s a stack of those feature Maps
it’s not just one
because you want to recognize different
patterns different templates and most
importantly those templates are learned
from the data they start off at complete
random and they start updating in a
training algorithm that I’m going to
describe in a second the second layer
there is called pooling and this is
meant to help with some variability in
the network so that it doesn’t care if
the person’s face is in the middle of
the image or on the left hand side or
the right hand side is still a face so
to do that in the pooling layer it looks
at a small region and takes the
strongest activation out of it that’s
why it’s called max pooling in many
cases and there’s different types of
booing but max works great and that’s
also why these maps get smaller cuz
you’re taking one out of a set a small
region and then you do alternating
layers of convolution pooling and so
forth remember the convolution has
templates that are going to be learned
the pooling is just a fixed math
function and then eventually you get to
the end of the network these fully
connected layers are just a fancy word
for convolution template matching where
there’s only one pixel all of the
pooling has reduced the image size down
to one pixel and the number of output
maps is corresponding to how many
categories you want to recognize so here
there’s only four so that’s how the math
works going through you take the pixels
template match pool template match and
so forth you have many many more layers
than this and even different types but
these are the most important and so
since the templates start off at random
off the start it goes through and makes
a random guess it doesn’t know anything
about the world and then you compare it
to what was labeled by a person in this
case people and you would get your error
signal this is the mistakes you’re
making so just like the interlude was
talking about the mistakes we make and
learn from is exactly how a eye learns
as well so it compares to the ground
truth sees the mistakes it’s making and
then it goes back through the network to
update those templates and it updates
them a little bit for this image and
then it goes looks at another image and
another image and another and millions
of images later it can start to
recognize a bunch of different
categories so that’s how these neural
networks work and they are really good
at image and video understanding but
you can think of neural networks just
mapping inputs to outputs so you can
configure them in different ways for
different problems so what do they
actually learn it’s pretty amazing that
these algorithms start off at random but
the first layer those templates I was
telling you about learn this so there is
96 templates here not just 3 like the
previous diagram and these are all
learned from data we never told it that
edges would be important or colors would
be important to recognize in the pixels
it learns different edges vertical
different angles different frequencies
you see sharp ones and blurry ones and
even different colored edges and colored
blobs and when you look into a primates
brain and people have done this they’ve
probed into cat’s eyeballs and monkeys
eyeballs and you see exactly this in the
first layer of the visual system which
is pretty amazing because we just came
up with these algorithms over the years
to replicate how the brain works and in
practice it actually does learn to
replicate how the brain works and what’s
more exciting is in certain tasks ai is
actually working better as a visual
system than humans and primates and so
even though it’s learning in the same
ways we can make it work better and
faster and so now you know how a eye
works but let’s shift gears and talk
about why diversity is important in AI
and how all you can have an impact on
that so if you see an image like this
you might be thinking black and white
you might be thinking World War 2 or
sailor or love or kissing everybody has
a different mindset right now as they
see this a different perspective and
because of your diverse backgrounds
growing up in different places different
parents different nations perhaps and
you’ve moved here you have a different
way of thinking about the world and if
you want to have a really powerful AI
system that is fair to all different
users and use cases we really want a
diverse set of people training the AI
and we view it as this critical period
right now in artificial intelligence in
the last three years we got it working
really well and now there’s this period
where we have to make it work for all
different users all different use cases
and there was this really interesting
experiment by Blakemore and Cooper
where they did some horrible stuff to
cats but it was for science so it was
good
they they they made these drums that had
on the inside only horizontal lines in
some drums and only vertical lines in
other drums and they put a newborn cat
into the drum right when it was born so
that was the only thing that’s ever seen
in the world was in the inside of the
drums and they left the cat in there and
fed them and all that for weeks but even
when they’re feeding they made sure not
to expose it to any other of the
different types of lines in the world
that we all see every day and then they
took the cats out of the drums after
several weeks and the cats that were in
the drum with a vertical line could see
all vertical lines in the world so they
would see like the chair legs and build
a weed between them but they wouldn’t
see any horizontal lines at all they
couldn’t see the seat of the chair they
would never they would even think the
the floor or a table as they were being
placed on a table was a weird thing they
just hit it and then they could feel it
they couldn’t see it at all and then the
horizontal the cats that were in the
horizontal drums they could see all
horizontal things so they could jump up
onto tables and and chair tops and all
that but they would run into any of the
chair legs so the visual system is in
this critical period where it learns
from when you’re born for a certain
number of weeks and the the really scary
thing is that no matter how long the
cats were then left into the real world
their system never came back and never
learned the other types of lines and
they showed several weeks later that
they couldn’t see the horizontal lines
or vertical lines depending on what drum
they’re in so we view that as the
critical period right now for AI and we
don’t want that to affect how our
algorithms work and how you can
interface with AI in the future and so
it’s not good vs. bad it’s not
Terminator and Skynet and all this scary
stuff that’s all in the movies it is
important to think about but we view it
as a couple different criteria that we
look at smart versus dumb so we want AI
to be incredibly smart because that’s
what makes it useful and it could be
less smart than a human worse more than
human doesn’t really matter but does it
provide a human some
is it smart enough to do that that’s
kind of one of the criteria we look at
when pushing the limits of AI ethical
versus unethical is artificial
intelligence going to take people’s jobs
or is it going to augment their jobs
make them able to do 10 times more work
or do it 10 times faster etc and is it
gonna is AI gonna abide by the laws or
what happens when AI gets so smart that
it’s equivalent to a person or maybe
smarter does the AI have more rights
than a human or less rights and same
rights there’s lots of different ethical
implications of AI as well and then
unbiased versus biased so this is a
really important one because if you only
teach just like the cat example if you
only teach AI with a few small sets of
the world it’s not going to understand
the rest of the world and so right now
we have to teach it with all of our
backgrounds and all the things we know
so that it doesn’t fall into case and
this has happened very recently so this
made a lot of press Microsoft had really
cool technology chatbot that they let
free on social media and what happened
was a small community of people taught
it a bunch of offensive stuff so when
you communicated with the chat bot it
would learn how to talk the way you talk
and it would respond to you in that
language and it learned very offensive
stuff because only a subset of the
planet really taught it whereas if the
whole planet we don’t talk that way
always and so the whole planet teaching
it would be much better off and it would
not have that bias another perspective
there was recently a beauty contest that
was judged by AI and it essentially
picked only white girls as what it
considered Beauty which is clearly not
true and this is one of the cases where
diversity is crucial because duty is
really in the eye of the beholder and
the same applies to AI so we have to
teach it what everybody thinks is
beautiful and that’s because AI is going
to be everywhere you see it in products
today you see it in the news almost
every day and so how do you get involved
how can you have this impact because the
next generation is going to be driving
they I field forward and you can think
even
a generation ahead your children
whenever you have them probably will
never have to get a driver’s license
that’s how fast a eye is improving and
it’s really exciting to think about it
that way and so there’s lots of
different ways you can get into AI you
can work in AI at startups like clarify
or large companies Google Microsoft
Facebook they all have really big
research labs and lots of universities
are now teaching it even an undergrad so
there’s lots of ways to learn about AI
and contribute to it and as well
contribute to open source projects
there’s a huge flurry of open-source
tool kits for neural networks and
applications on top of those toolkits to
apply ai in the real world and then you
can think about the ethics of AI there’s
think tanks like open AI which Elon Musk
drove and even write your own blogs on
how you view is gonna improve life and
improve humanity and then bring AI into
your world whether it’s your school as a
science project start an AI club or you
all be going a university soon perhaps
started up there and bring it in to your
everyday because AI is really for every
human on the planet that’s why we need
every single one of you teaching it
thank you [Applause]