
gamer and modulation my recent subject
is really interesting and partly but
when what we need before also a urban
danger I was asked to talk about our
survey today I ask you what I were them
I’d like to go but the only one in Dar
but also what I think of are we giving
great decision will be more and more
affordable series in the future so I am
very happy if i can give you a clear as
so I put one TI and why so much
attention very people grass fire eater a
peta kappa poor dog but if you have to
write up how so it’s other of their
little ponies are my a movie a million
so you did that all technical exchange
who knows so you have to take so there
are many other criteria Saturday able to
be able to classify this part picture
into adult categories but probably it is
difficult to classify all of these
pictures cry 3 why we can classify them
so easily but why we cannot know how to
do that Chinese because what we are
doing is what our brain is doing our
Granger Browne posible pero and very
free food system and it is said that a
hundred billions of euros are working in
parallel in a brain on the other hand
when we write down something or speak
something we use a world and suppose so
it should be sequential so due to that
each difference between a parallel
situation and so we cannot ascribe what
we are doing in a brain so the
researchers have developed a for the
system into a similar paths we call this
an actual module for general recognition
planning control and memory and so on
and they try to understand the function
of each modules and
also try to an together human rock until
Robert by giving the knowledge into the
each modules and his recommission mother
his father divided like this I say we
put image and through the pre-processing
image segmentation and fit extraction
and classification and then finally it
can say this is Cathy the researcher try
to give Malone bon voyage and that means
sometimes they try to prepare well a
useful features Randy’s and so the
recognition rate has increased but
actually it has been such rated and he
will reach a human evil solder deep
running has appeared it is a central
technique in time presently I’ve and
either just Mario you will it work in an
hour in another world ah the very simple
model of a break so it implies that it
can do the boss error processing and
learning and he employed a special
structure for sieved and scary barriers
and a GPU does sound for the graphical
processing unit has a so far alvaro
computing and it enables right
computation of the large-scale neural
network and turn running so useful
features emerged from the neuronal to
was wracked highly though after running
these pictures for example could be
classified into one of the 1000 calories
so this is a motor scooters and this one
is can be classified into a robot not
each other and so on so what it is not
enough to we are as a human life
intelligence where you take a look at
this picture what do you recognize you
aren’t going to only cross by this
picture to the cat categories but you
think our magazines for example the
Cascades Otis and they are gathered into
being fault others heaters and they
don’t seem to notice fees they did not
give me away easily so we can recognize
and seek so many things from this one
picture in a year award we face in
niggas situations and so it is very
difficult to decide what should be the
output of this revision manual what is
the origin of this problem to another
the brain does researchers divided it
into modules to system introduced and
they try to understand and develop to
each module but to do them its input and
output should be decided at first and to
destroy them they have to they have to
understand brain so this is a kind or a
chicken and egg problem so as you can
see that the important thing is without
division
into modules the parallel process for
the sensor to notice k rod at the same
time and the internal retreat addition
we don’t need to prepare in advance but
they emerged through running what will
see that is necessary to hear the humor
are contusions is a rallying without a
teacher this rod Quran to push a member
even though it didn’t get in any
direction from the outside but it can
learn by itself so the labels running
has appeared so it is a learning from
the reward and punishment through trials
and errors it enables a water was
running without a teacher so pottery we
have reached end-to-end rivers Ronnie
using a neural network the file system
from the sensor the motors is consisted
of one euro network and using the motion
and perception dude the wrinkles running
disturb right and to running this is
Rapunzel’s emerged without any
directions so i have been preceded in
this framework for a lot of three years
so far and then unfortunately and you
are scared about that so now the google
deepmind maybe you know v2 advocated
advantage of this approach attached
impressive demonstration this framework
bina is getting a lot of attention now
so I am surely had to demonstration
videos the first one is what we did in
2008 named the kissing I would ask
whether I whoa a robot it has a camera
attached to its nose and it was around
and wake it kissed the other angle so it
can get a rework the camera image is the
input of the neural network and I was
according to the output of the neural
network so after Pearl thousand trials
groaning it finally approached I will
approach and kissed the arrival even
though the successful ratio did not
reach other person actually a rather
ninety percent but before running the
island you know how to recognize the
arrival and how to act to approach the
arrival and also he didn’t know it
should approach to kiss the arrival so
that it could run by itself so this is a
demonstration by google deepmind so it
runs to play the game pick out so as you
can see that first it can he couldn’t
receive the ball and in this case that
this image is also the input of the
neural network and the power goes
according to the output of the neural
network so now that it can it could see
the ball and the game score is used as a
reward and after 600 trial
running so they can it can it could get
very nice strategy rightness no one
very nice and it was a surprising news
the last year Alfred oh that was
developed also good dirty mind it one
provision human co-op real goal is a
really competent game and it wanted to
ensure that the official integers
neighbor win the gold failure so
something the first part is that if you
expect a purse of intelligence so don’t
interrupt a couple it was running and
your network so Reba to O’Donnell’s
running so indigence emerged a quarter
emerged as a fruit of houghton throwing
in the second bird let me talk what I
think about AI and model science so free
from the fear salvation is very very
important for humans and it brings us so
much happiness in such situation however
if a serious man well a lot instead of a
person so how much happiness could you
get so the benefit of science has
increased as a progress of science but
combined with the passage that increase
it is not a sorry it is difficult to
expect a large increase
in happiness so to get my happiness our
power grows is expected that means a
self-replicating technologies is
demanded nuclear technology
biotechnology running are all
self-replicating technologies they have
a great potential because it grows
exponentially right this club but it is
very difficult to predict the outcome
and also is very difficult to control
them especially running has a potential
to enable robots asleep by themself so
the risk increase as a progress science
and there is a critical point so I think
that it’s a time to we have the very
little risk of science and technologies
I’m afraid that the critical point has
passed already who should make a
decision even researchers cannot predict
the outcome and they don’t want to deny
their research and it’s in southern poly
poly will come sadri and it’s a human
hubris to say that it can be controlled
or can we use multiple benefits and also
the captors privatize the short term
profits over the absurdity of our human
future our future anyway I really hate
I’m on air base all the robots sync by
themselves we slaughter I and they are
wondering a lot week but the present
arrives are so awful yet and they will
not be able to sync by themselves in 10
years I think but it comes soldering and
once it occurs it may not be able to be
stopped easily but now we can ok so
another de Bruyne said on their homepage
that sort of it engines use it make the
world a better place so how do you want
making our future so thank you for your attention and from