All You Need to Know About AI But Were Afraid to Ask
have you ever seen this
nice fellow before this robot now romeo
and pepper
i’ve been working with them for 10 years
and
romeo the big one is almost my fifth
child
i discovered robotics after my master’s
in ai at the end of the 80s and i never
let them since
first i’ve been working 20 years at cea
research institute where i developed
robots for nuclear application
and others for disabled handicapped
people
then 10 years at aldebarans of
bankrobotics with these fellows
and today i’m an expert in ai at
renault for intelligent car
compared to the ai i discovered when i
was young
the ai of today is incredibly more
powerful
you can find any information you want on
the web
uh ai can defeat you defeat any chess or
go master
uh was supposed to be the the most
intelligent guy
in the world and computer vision has
made a tremendous gap
ai computers can recognize your face as
we do but
it can also detect teeny tumor much
better than any expert radiologist
at aldebaran we were trying to make our
robot
ever clever more clever and more clever
for instance we taught now to climb a
ladder
and uh pepper has been trained to to
play you know cup and bold game
and um by training so we showed it a
good gesture once and then it train
again
and again and again and finally succeed
and when
you succeed to find a good gesture it
can redo it exactly
perfectly and win at every time i’m
always completely amazed
when i see robots doing uh
wonderful tricks and i’m not the only
one
when we conducted experiments with
pepper
with elderly people they were completely
amazed uh
to see the robots doing uh a target
by themselves and the favorite moment
was when the robot goes
to charging to the charging station the
fact that the robot decides by itself
well my battery are low i have to go
charging and does it was completely
amazing for them
another thing that is really amazing to
me are the magic tricks of my son
my oldest son simon is an amateur
magician and i’m
always delighted when the cartoon i
picked up in the deck reappeared
in my in my wallet or in my pocket or in
my ear
sometimes he explained me the trick and
it’s
always off very simple
very clever requiring some dexterity of
course
and i think ai is the same kind of magic
it looks magic
but it’s not when you know what is
behind the curtain
you are amazed but a little bit
disappointed
for instance if you
change the ladder of now it will redo
exactly the same gesture and of course
it will falls down and pathetically keep
on crawling on the ground
ridiculously and and the day paper
succeeded
in in this cup and ball game we brought
it in the in the ceo’s office to show
the performance
and of course it failed actually we
we did return a paper in our lab
and with a special slope on the ground
and when we were in the ceo’s office you
know on the thick carpet
the vertical angle of the robot was not
exactly the same so the perfect gesture
there was not the perfect gesture here
and it failed of course
all these little troubles can be fixed
but what i want you to to see that the
intelligence of your machine is very
narrow
as soon as you change a little bit uh
the situation
things can go completely wrong do you
know this story
of the tesla autonomous car that went
directly
through a truck knocked over in the
middle of the highway
the vision system of tesla is one of the
best in the world
it has been trained by watching millions
of
driving hours millions of hours of
driving video
each image of it was annotated
you know by showing on this picture you
have pedestrian
a car a cycle this is a tree
this is a building this is a road
you have to show millions of bicycles to
the system for it to recognize another
bicycle and this tedious work
made by human beings is necessary to
make machine
intelligent and thanks to this
supervised learning
you are your your computer your vision
system is able to recognize 99.9
of the object you will ever cross on the
road but not 100
during the training phase it never have
seen
such a track in the middle of the road
so it was unable to recognize
this object as an obstacle you probably
thought it was a
the sky or or a billboard above the road
but not
an obstacle that should be avoided and
then it went through
today we consider that autonomous
driving will avoid
plenty of accidents that human beings
have
but it will also generate crashes that
even the worst driver
could never have ai is perfect in its
domain but as soon
as it outside of it it’s fair
it’s it’s a drama and and and the the
the worst is that
ai does not have the consciousness of
being outside of its domain
so what is going to happen now
will researchers solve all these little
problems
or will ai reach a kind of limit
in 1997
when uh deep blue defeated the the chess
world master kasparov
people say yes you know but chess it’s
simple
it will never happen with go that is
really complicated
and 20 years later it happened alphago
defeated lisa doll thanks to the huge
progress of
ai and the moore’s law you know that
makes the
the computing power twice more powerful
every two years
now we can implement solution that would
even
dare to hope in in the 80s because of
the required
computing power but today you can run
neural deep neural network on a laptop
your smartphone can defeat you at any
game you play
so it’s really difficult to guess what
will be the limitation of ai
the victory of alphago made me think
that
saying that something it is impossible
to ai
it’s impossible today researchers are
working on
unsupervised learning it means that you
don’t have to explain for each input
what is supposed to be the expected
output
in the domain of out of domain detection
we are progressing too it means that the
ai
will be able to say wow but this object
i’ve never seen before so i don’t take a
decision
yes that’s that’s the first step to our
consciousness it’s nice to know that i
don’t know
but the common sense is missing
you as a driver when you see an object
that you don’t know the name of in front
of you you know what to do you know if
you should break or turn around or or
or or or keep on but ai does not have
this implicit understanding of unknown
situations
when you think about the way you learn
to to to to make
computers learning you know how it’s
simple to learn but oh it’s complained
complex to explain
and watching children learning is a good
way to try to understand
a baby does not need to to see millions
of teddy bear to recognize another teddy
bear
but in the other way around it’s
incredible
to see how difficult it is for them to
remember that 7 times 8
is 56 when a very cheap calculator can
do it
hey daddy if a very cheap calculator can
do it why should i remember that
hey i’m talking to these people ask your
mother
excuse me um yeah so it means that um
a a community of researchers
is working on what we call developmental
robotics
they want to implement the way that
children learn and without or there is
very few supervision
only by experiment by experience the
computer will discover the world
the effect of the is motioned with the
rest of his body
and this is this is a very very complex
task it’s it requires a lot of
mathematics to do that
but it will take very very long time for
this kind of intelligence
to become a threat for humanity
because at the end of the day that is
the question that the reason why you’re
listening to me
will i overcome humanity the idea of
singularity you know singularity will be
the moment where
computers will be more clever than human
beings
as we’ve seen we’re already there
computers can defeat you at
any games a computer can store
and use more information that all humans
can do all together
alan turing you know that is one of the
father of the ai
he said ai will really exist today
you will not be able to detect if you’re
talking to a machine or
to a human being when you discuss with
uh with some chat bots today you can
have long discussion
sometimes more interesting than a
discussion that you can have with a
football fan at the stadium
some people have tried to evaluate the
computing power
of a human brain they say it’s about 20
petaflops
as you know one petaflop is one million
of billion operations per second
others say it’s a thousand petaflops
anyway
today the most powerful computers are
hundred petaflops
you’re not so we’re not so far
and what do we have do we have computers
able to create by themselves
even more intelligent computers
no so okay futurologists say that
singularity will happen in 15 20 years
from now
but what will happen in the meantime
always more computing borrower
sure self-learning machine
self-replicating machine why not
but does it mean that this will lead to
the destruction of humanity
in a way it’s a very anthropomorphic way
to see the world
think that if you are more clever than
people you will overcome them
why couldn’t we couldn’t we imagine that
we can live
peacefully side by side with very
intelligent machines
like the electricity like the internet
ai
will be more and more important in our
lives
if we try to compete in its excellence
domain
we are lost it’s like trying to run
faster than a car on the highway
it’s silly but being a pedestrian
you can climb mountains you can dance or
you can play football
of course you cannot play anymore and
you cannot run on the highway anymore
and and do you regret it and that’s my
point
why shall we regret that ai can do
for us ai is invitable at chess
okay but did we stop playing chess no
but we can have ai as a sparring partner
to prepare
the next game against our friends
it will be difficult to turn ai off
so now we have to imagine the life that
go with it
a study in 2018 showed that
60 of the jobs created in the u.s
between 1980 and 2015
had names that did not exist before you
know
because of uh internet and computers of
course
but with ai it will be the same
it’s it’s it’s funny to see that every
week you can have
new application on your smartphone i
think there are about
two million applications available on
the apple store
none of them are not all of them are
made of ai but plenty are
and and it’s incredible to see how
imaginative people are to create these
new services for us
and ai is a wonderful support for them
for their creativity
by the way it’s interesting to see that
so many people
are able to develop ai-based
applications because
ai algorithms rely on very high level
mathematics
of course you know the principle of
neural networks so you have a
data coming from the previous layers
coming here and
and each value is multiplicated by the
white a weight all together they are sum
up and then transfer
to the next neural layers
that is pretty simple what is a little
bit tricky is to compute
the multiplicative weight for each
neuron and to have the expected
output and what is a little bit
complicated
is when you try to deal with images for
instance then you need
convolutive networks be more complicated
and if you want to to deal with video
you need
convolutive and recursive neural
networks
this is really complicated and ai
researchers are working hard
to make the ai more and more efficient
thousands of
very high level scientific papers are
published every
year what i want you to understand with
this
is that even if ai is based on learning
it does not develop by itself brilliant
human beings are
improving the performance of ai
and the learning part of it is just how
to apply this generic
architecture to specific use case like
for instance
translating greek instead of hungarian
or recognizing plants
instead of birds
this make possible for almost anyone to
develop
ai based applications the idea that
researchers find as soon as a researcher
finds this this new architecture they
publish it
and you can reuse it today amongst the
best researchers in ai
come from google and and facebook they
are not really public institutional
philanthropic associations
but they publish almost everything and
then good engineer can reuse that
for its own application
it’s um actually it’s a massive change
in the way an economy can grow
finally i would like to compare the
ai to a big wave you can
consider it as a terrifying tsunami
but because ai comes from california you
also can consider it
like surfer do it’s a way to have fun
and to go fast
you have different kind of of of server
you have the researchers you have the
developers and you have the users
and as users what is really important is
you
understand what can and cannot do ai
ai cannot be human er
ai is still hard time to deal with
complex human emotions
ai cannot improvise in this complex
situation and that’s where we have a
place
to deal with people to take care of
people in complex situation
in this situation if you behave like a
robot
sooner or later you will be replaced by
a robot
but if you behave like a human you will
always
have your place thank you