let me take you first to 1956 so this is
a workshop at the Dartmouth summer
school in Vermont and it’s considered
date of birth of artificial intelligence
and let me read the statement they made
why they had this meeting to proceed on
the basis of a conjecture that every
aspect of learning or any other feature
of intelligence can be in principle be
so precisely described that the machine
can be made to simulate it think about
it they want to crack intelligence they
want to have a computer do what humans
do
that’s 1956 that’s a bold goal because
computer in 1956 is like this hmm so
when you have this and you think about
making a computer simulated intelligence
you really have a vision think about it
if you swipe on your iPhone the
computing power to do that simple action
might take this machine months so a
session on Facebook may take 15 years so
it’s a long shot to think of artificial
intelligence in that environment and in
fact it took something like 50 to 60
years before that promise could be
materialized so we are more or less in
2012 2010 2012 that the first wave of
big important results came out the first
wave of important algorithms that sort
of simulate the brain and do something
like the brain do does better image
recognition is the first big example
image recognition implies that you have
an image a picture of something and a
machine tells you what’s in there right
so this has been worked on for many many
years and you can see the albums were
okay for a long time they could
recognize three out of four images which
is not very good for doing things this
is just plain software and then a group
of researchers in Toronto say forget
about that way of doing things we don’t
take an algorithm a software which you
have put all the instruction to say how
to recognize an image but we teach the
algorithm we take an algorithm which is
actually designed to learn not to do
things but to learn what happens is that
from 75% it goes up to 80% and then year
after year it becomes super human so it
means that does a human task better than
a human and now you can do that in image
recognition in text recognition
in speech in motion and so on and so on
and so on
in the past everything we did with
software was instructions you tell us
after what to do and everything you do
with the software has to be told to the
software by somebody so when you click a
key on a keyboard somebody has told the
machine to do everything that happens
until something comes up on the screen
except for bugs right when you have bugs
that’s the only time when the software
invents something even if it’s something
that you don’t really want so the new
set of algorithms is but has been are
based on a different premise the premise
that you copied the way the brain works
and the brain works in this way many
simple pieces of computing which are
core neurons which are compiled together
in a very complex layer and this is how
it happens
within this image recognition algorithm
so you put an image at the entrance that
image is segments in different places
and after you’ve shown many many images
that algorithm can make the association
between the input and the output you
show that image and this is Sarah is so
another image and this is another person
Facebook does that all the time so
whenever you post something in facebook
the first thing that happens if there is
a picture there it goes through a series
of neural networks which actually split
out the content because of that you have
a change in the parting you go from
software or algorithms in the old way in
which you actually do instructions you
tell the algorithm everything that you
want it to do that’s why you call
instructions in software you go from the
earth to what we call supervised
learning so you show algorithms many
many examples of images of speech of
applications for university and so on
and that algorithm recognizes the
relationship between something that says
this and the output it makes the
connection between input and output and
then you show it something it hasn’t
seen and it can make it itself it’s a
very very simple way of doing
intelligence it’s actually very very
simple this is not how our brain works
in practice we do many other things but
this is what we have now so it’s not
really artificial intelligence what we
call restricted artificial intelligence
is extremely powerful nonetheless where
we are going to go it’s unsupervised
learning
so you need very few examples to learn
think about the first time you saw a
Segway right you just needed one frame
one second of that object to figure out
it was in the same category of scape
works bicycles and so on you didn’t need
anything else just one example computers
cannot do that it might take years or
decades for them to learn it or maybe
never but in any case so this is the
part where we are so this moment we know
how to do supervised learning we know
how to teach machines from many examples
to do it better than humans and that’s
powerful it’s powerful because if a
person can do it task a mental task in
less than a second probably it can be
automated probably there’s an algorithm
that does it better than you but not in
ten years not in five years with what we
have now technology this is now if you
take that app step further
what is the impact why is it important
why it’s important because about 60% of
jobs occupations and so on will be
changed by that without any development
anything new just what we have now so
all my job your future jobs and so on
will be affected by this in a
substantial fundamental way so in five
years from now it would be impossible to
imagine the doctor that does diagnoses
without consulting an algorithm that’s
just impossible to imagine it’s just
impossible to conceive a lawyer that
starts the case without an agent doing
the background work for the lawyer
itself so it’s important so the question
is can we trust these algorithms do they
tell you the truth and when they make
decisions for you do they have a moral
standard that they use for making the
decision so I’m going to touch on some
of these elements in the presentation
now the first problem you have is this
many of the powerful tools that we have
developed within what we call artificial
intelligence are actually black boxes so
you know that they work well you know
what you put in you know what you put
out you don’t really know why they work
that well it is surprising but we
actually don’t know the real function
the real inner working of a neural
network we know the wiring but why does
it usually
we actually don’t know so you have to
start thinking about how do you actually
poke inside this algorithm to make sure
that they do what you want and they do
it in a correct way
so let’s start with a simple case in
this case let’s say that you want to
interpret this image and you want to
have an algorithmic figure out where
there’s more people that’s a simple case
right there will be no algorithm in the
world and make a different assumption
this is Obama inauguration and this is
Trump in aberration where there’s more
people pretty clears on the right
every machine will be able to do this is
simple it’s explicit information
actually every human being would do the
same except perhaps 10 or 12 in
Washington so it’s fairly clear
situation here but then what happens
when you have this so here you have a
machine that takes an algorithm that
takes many many medical dossiers
compiles them and it’s capable of
detecting sepsis which is you know a
serious condition 90% of the time better
or in actually faster than a doctor so
you have an algorithm that actually
performs better than a doctor in the
sense that it detects things quicker
right and then you doctor assume you’re
a doctor you have to find out what do I
do with that do I trust it do I act upon
it because I got 16 hours advantage but
I have to find out if this algorithm is
correct so what is it that made this
algorithm go from the evidence it had to
the conclusion it makes so you have to
have that clear because otherwise you
can’t use it it gets even more
complicated in situations like this so
this is an algorithm that predicts the
road accidents by the way this is a
simulation we cannot put out in the real
panels on the roads because we don’t
know how people would react to this but
these are algorithms anyway that
predicted in the next 10 kilometers in
the next 30 minute there’s going to be
an accident or not right and then if
they do that how should you need to be
before you actually put out that
information because once the information
is out and maybe this is not the best
way to do it that this is going to
change the way people behave so you
might have a change of traffic which we
proved remove the incident but you know
it might actually make it work so you
don’t really know and worse in even
worse in any case even deeper is that
you have machines that are really so
strong and powerful like this one that
can raise a person that cannot do that
but as much as they can raise
also crush it so really have to trust
them completely so here’s the issue that
we try to solve and I’ll give you a
couple of examples how we inside the
research community practitioners those
work in AI try to do it
the main problem you have is you have
two types of algorithms does that work
extremely well but you don’t really know
why those that you know why they work
but they don’t do it that well so
explain ability and performance are in
conflict there’s a trade-off you want
good performance you don’t know how to
explain it you want to explain it we
have to give up performance we want to
get rid of that conflict we don’t want
that trade-off we want to make sure that
actually explainable and they work well
so some of the things that people are
doing is of course to test what are the
limits of these algorithms the first
thing you want to do since we are
talking about training and so you’re not
coding an hour you’re training an
algorithm are you introducing biases are
you introducing biases that make the
algorithm work in a way which is not
supposed to do this is called compass
it’s a tool used by courts in the United
States some courts to decide if convict
can be released on parole or not so you
decide in freedom or not for a person
and it does it based on history and a
questionnaire and puts out the score and
what comes out is that this algorithm
this has been tested by number of people
who try to push the argument to the
limit is that white 2d3 event I need to
read it because you can’t whites who d3
offend we’re classified as low-risk
twice as much as blacks auditory offend
so that tolerance was really biased in
favor of white people right why is that
is it somebody did it on purpose
probably not there’s nobody wants to do
that on purpose just the history the
cases the judges that decided in the
past that had that bias coded in in the
information used to train the algorithm
and that’s now is part of the code to
actually advise judges so this is one of
the main things we cannot afford to do
this has to be sifted out we can do this
because it’s it’s a matter of fact that
worse is not a fair place it’s in just
and so on but there’s one thing we
cannot afford to take that and just
encode it and make it part of algorithms
so
people are doing something experiments
like this stretching the algorithms to
the limit to find out where they fail
where they break or in any case where
they are biased another case of that
exercise is this so a lot of image
recognition perhaps everything that has
to do with image recognition at this
point in time is automatic so you just
have software tool and Algrim that does
it for you
so if you have a face you recognize that
that face is my face is the same with or
without it doesn’t matter it’s still me
right and if I change color of glasses
it’s still me but is it the same for a
machine actually turns out is not so
some people are using tools and
techniques to see where the algorithms
break so you want to stretch it to the
limit and find out places and training
sets that fool them this is what these
guys have found it is sufficient in this
case to wear glasses that have a very
specific color and pattern to fool the
image recognition algorithm so what
happens is that this guy on the top
right that’s the guy but is recognized
as the lady the middle lady is
recognized as that guy and so on so the
algorithm is completely confused and and
the reason why it does is because that
piece of information that patterns
actually sifts through the neural
network at that some layers of the
neural network alters something
fundamental there remember that we don’t
really know why this works so the only
thing that we can do is to test it to
the limit so we know where it breaks it
will never use something that you don’t
know where it breaks so this is fairly
important piece of training but it’s all
about training so there is some limit to
what you can do by testing you have to
do something upfront
this is a typical dilemma of autonomous
car this is a vehicle that cannot stop
anymore and it has three terrible
choices in the next few seconds either
you hit the person right or you hit the
tree on the zebra stripes or you go
against that wall and you risk the life
of the passengers of the car which one
you cannot stop the car this is going to
happen in action unfortunate has
happened to people before it’s it is one
of the situations in which the outcome
of your choice is Nega
there’s nothing positive that Alka right
so how do you choose here well if you do
training the way we do now with
supervised training you probably do this
so this is the way you train an
algorithm to beat a go player this is go
it’s a board games much more complex
than chess and it was told and nobody no
algorithm could actually win a human
player but it has been demonstrated last
years now it’s a done deal how do you
train this you take many many games play
in the past you have the algorithm to
recognize the patterns of this game and
at the end this algorithm does it as the
good players right and then actually the
trick is we make the algorithm play
against itself so it’s going to train
even better after a million games so you
can train a machine a car this way
because then you have to create a
million debts before figuring out that
that’s not really what you want from the
car so you have to find another way to
do it so the way the tech tree is
developing now is this you actually
raise algorithms the way the same ways
you raise children exactly the same way
with one catch you cannot wait 20 years
so you have to find ways to accelerate
that training so there are a couple of
ways that actually work well one that I
like a lot is Georgia Tech in the u.s.
is that with the create stories about
the behavior of people in that situation
thousands of stories it just narratives
and you have algorithms learned from
these narratives what are the code the
moral code or the choices that people
have made in those situations and then
that way you actually have the machine
learn what humans have done in a
terrible situation like that it might
not be the best choice but at least is
as good as humans have done it and from
there you can construct so there are a
few experiments going on at this moment
one that I like a lot is an algorithm
like this which has been trained to
recognize what is the choice is the
codes that you use when you go out on a
date and you go to the movies so
thousands of people have constructed
narratives about that event and guess
what the algorithm detects a number of
things but what also detects is that a
key element in that schema is this so
apparently for that algorithm kissing is
an extremely important element of going
out in a day to the movies based on the
stories
of many people this is a way of learning
you know what if an algorithm turns out
and suggested kissing I think we’re in a
good place
thank you
[Applause]
a workshop at the Dartmouth summer
school in Vermont and it’s considered
date of birth of artificial intelligence
and let me read the statement they made
why they had this meeting to proceed on
the basis of a conjecture that every
aspect of learning or any other feature
of intelligence can be in principle be
so precisely described that the machine
can be made to simulate it think about
it they want to crack intelligence they
want to have a computer do what humans
do
that’s 1956 that’s a bold goal because
computer in 1956 is like this hmm so
when you have this and you think about
making a computer simulated intelligence
you really have a vision think about it
if you swipe on your iPhone the
computing power to do that simple action
might take this machine months so a
session on Facebook may take 15 years so
it’s a long shot to think of artificial
intelligence in that environment and in
fact it took something like 50 to 60
years before that promise could be
materialized so we are more or less in
2012 2010 2012 that the first wave of
big important results came out the first
wave of important algorithms that sort
of simulate the brain and do something
like the brain do does better image
recognition is the first big example
image recognition implies that you have
an image a picture of something and a
machine tells you what’s in there right
so this has been worked on for many many
years and you can see the albums were
okay for a long time they could
recognize three out of four images which
is not very good for doing things this
is just plain software and then a group
of researchers in Toronto say forget
about that way of doing things we don’t
take an algorithm a software which you
have put all the instruction to say how
to recognize an image but we teach the
algorithm we take an algorithm which is
actually designed to learn not to do
things but to learn what happens is that
from 75% it goes up to 80% and then year
after year it becomes super human so it
means that does a human task better than
a human and now you can do that in image
recognition in text recognition
in speech in motion and so on and so on
and so on
in the past everything we did with
software was instructions you tell us
after what to do and everything you do
with the software has to be told to the
software by somebody so when you click a
key on a keyboard somebody has told the
machine to do everything that happens
until something comes up on the screen
except for bugs right when you have bugs
that’s the only time when the software
invents something even if it’s something
that you don’t really want so the new
set of algorithms is but has been are
based on a different premise the premise
that you copied the way the brain works
and the brain works in this way many
simple pieces of computing which are
core neurons which are compiled together
in a very complex layer and this is how
it happens
within this image recognition algorithm
so you put an image at the entrance that
image is segments in different places
and after you’ve shown many many images
that algorithm can make the association
between the input and the output you
show that image and this is Sarah is so
another image and this is another person
Facebook does that all the time so
whenever you post something in facebook
the first thing that happens if there is
a picture there it goes through a series
of neural networks which actually split
out the content because of that you have
a change in the parting you go from
software or algorithms in the old way in
which you actually do instructions you
tell the algorithm everything that you
want it to do that’s why you call
instructions in software you go from the
earth to what we call supervised
learning so you show algorithms many
many examples of images of speech of
applications for university and so on
and that algorithm recognizes the
relationship between something that says
this and the output it makes the
connection between input and output and
then you show it something it hasn’t
seen and it can make it itself it’s a
very very simple way of doing
intelligence it’s actually very very
simple this is not how our brain works
in practice we do many other things but
this is what we have now so it’s not
really artificial intelligence what we
call restricted artificial intelligence
is extremely powerful nonetheless where
we are going to go it’s unsupervised
learning
so you need very few examples to learn
think about the first time you saw a
Segway right you just needed one frame
one second of that object to figure out
it was in the same category of scape
works bicycles and so on you didn’t need
anything else just one example computers
cannot do that it might take years or
decades for them to learn it or maybe
never but in any case so this is the
part where we are so this moment we know
how to do supervised learning we know
how to teach machines from many examples
to do it better than humans and that’s
powerful it’s powerful because if a
person can do it task a mental task in
less than a second probably it can be
automated probably there’s an algorithm
that does it better than you but not in
ten years not in five years with what we
have now technology this is now if you
take that app step further
what is the impact why is it important
why it’s important because about 60% of
jobs occupations and so on will be
changed by that without any development
anything new just what we have now so
all my job your future jobs and so on
will be affected by this in a
substantial fundamental way so in five
years from now it would be impossible to
imagine the doctor that does diagnoses
without consulting an algorithm that’s
just impossible to imagine it’s just
impossible to conceive a lawyer that
starts the case without an agent doing
the background work for the lawyer
itself so it’s important so the question
is can we trust these algorithms do they
tell you the truth and when they make
decisions for you do they have a moral
standard that they use for making the
decision so I’m going to touch on some
of these elements in the presentation
now the first problem you have is this
many of the powerful tools that we have
developed within what we call artificial
intelligence are actually black boxes so
you know that they work well you know
what you put in you know what you put
out you don’t really know why they work
that well it is surprising but we
actually don’t know the real function
the real inner working of a neural
network we know the wiring but why does
it usually
we actually don’t know so you have to
start thinking about how do you actually
poke inside this algorithm to make sure
that they do what you want and they do
it in a correct way
so let’s start with a simple case in
this case let’s say that you want to
interpret this image and you want to
have an algorithmic figure out where
there’s more people that’s a simple case
right there will be no algorithm in the
world and make a different assumption
this is Obama inauguration and this is
Trump in aberration where there’s more
people pretty clears on the right
every machine will be able to do this is
simple it’s explicit information
actually every human being would do the
same except perhaps 10 or 12 in
Washington so it’s fairly clear
situation here but then what happens
when you have this so here you have a
machine that takes an algorithm that
takes many many medical dossiers
compiles them and it’s capable of
detecting sepsis which is you know a
serious condition 90% of the time better
or in actually faster than a doctor so
you have an algorithm that actually
performs better than a doctor in the
sense that it detects things quicker
right and then you doctor assume you’re
a doctor you have to find out what do I
do with that do I trust it do I act upon
it because I got 16 hours advantage but
I have to find out if this algorithm is
correct so what is it that made this
algorithm go from the evidence it had to
the conclusion it makes so you have to
have that clear because otherwise you
can’t use it it gets even more
complicated in situations like this so
this is an algorithm that predicts the
road accidents by the way this is a
simulation we cannot put out in the real
panels on the roads because we don’t
know how people would react to this but
these are algorithms anyway that
predicted in the next 10 kilometers in
the next 30 minute there’s going to be
an accident or not right and then if
they do that how should you need to be
before you actually put out that
information because once the information
is out and maybe this is not the best
way to do it that this is going to
change the way people behave so you
might have a change of traffic which we
proved remove the incident but you know
it might actually make it work so you
don’t really know and worse in even
worse in any case even deeper is that
you have machines that are really so
strong and powerful like this one that
can raise a person that cannot do that
but as much as they can raise
also crush it so really have to trust
them completely so here’s the issue that
we try to solve and I’ll give you a
couple of examples how we inside the
research community practitioners those
work in AI try to do it
the main problem you have is you have
two types of algorithms does that work
extremely well but you don’t really know
why those that you know why they work
but they don’t do it that well so
explain ability and performance are in
conflict there’s a trade-off you want
good performance you don’t know how to
explain it you want to explain it we
have to give up performance we want to
get rid of that conflict we don’t want
that trade-off we want to make sure that
actually explainable and they work well
so some of the things that people are
doing is of course to test what are the
limits of these algorithms the first
thing you want to do since we are
talking about training and so you’re not
coding an hour you’re training an
algorithm are you introducing biases are
you introducing biases that make the
algorithm work in a way which is not
supposed to do this is called compass
it’s a tool used by courts in the United
States some courts to decide if convict
can be released on parole or not so you
decide in freedom or not for a person
and it does it based on history and a
questionnaire and puts out the score and
what comes out is that this algorithm
this has been tested by number of people
who try to push the argument to the
limit is that white 2d3 event I need to
read it because you can’t whites who d3
offend we’re classified as low-risk
twice as much as blacks auditory offend
so that tolerance was really biased in
favor of white people right why is that
is it somebody did it on purpose
probably not there’s nobody wants to do
that on purpose just the history the
cases the judges that decided in the
past that had that bias coded in in the
information used to train the algorithm
and that’s now is part of the code to
actually advise judges so this is one of
the main things we cannot afford to do
this has to be sifted out we can do this
because it’s it’s a matter of fact that
worse is not a fair place it’s in just
and so on but there’s one thing we
cannot afford to take that and just
encode it and make it part of algorithms
so
people are doing something experiments
like this stretching the algorithms to
the limit to find out where they fail
where they break or in any case where
they are biased another case of that
exercise is this so a lot of image
recognition perhaps everything that has
to do with image recognition at this
point in time is automatic so you just
have software tool and Algrim that does
it for you
so if you have a face you recognize that
that face is my face is the same with or
without it doesn’t matter it’s still me
right and if I change color of glasses
it’s still me but is it the same for a
machine actually turns out is not so
some people are using tools and
techniques to see where the algorithms
break so you want to stretch it to the
limit and find out places and training
sets that fool them this is what these
guys have found it is sufficient in this
case to wear glasses that have a very
specific color and pattern to fool the
image recognition algorithm so what
happens is that this guy on the top
right that’s the guy but is recognized
as the lady the middle lady is
recognized as that guy and so on so the
algorithm is completely confused and and
the reason why it does is because that
piece of information that patterns
actually sifts through the neural
network at that some layers of the
neural network alters something
fundamental there remember that we don’t
really know why this works so the only
thing that we can do is to test it to
the limit so we know where it breaks it
will never use something that you don’t
know where it breaks so this is fairly
important piece of training but it’s all
about training so there is some limit to
what you can do by testing you have to
do something upfront
this is a typical dilemma of autonomous
car this is a vehicle that cannot stop
anymore and it has three terrible
choices in the next few seconds either
you hit the person right or you hit the
tree on the zebra stripes or you go
against that wall and you risk the life
of the passengers of the car which one
you cannot stop the car this is going to
happen in action unfortunate has
happened to people before it’s it is one
of the situations in which the outcome
of your choice is Nega
there’s nothing positive that Alka right
so how do you choose here well if you do
training the way we do now with
supervised training you probably do this
so this is the way you train an
algorithm to beat a go player this is go
it’s a board games much more complex
than chess and it was told and nobody no
algorithm could actually win a human
player but it has been demonstrated last
years now it’s a done deal how do you
train this you take many many games play
in the past you have the algorithm to
recognize the patterns of this game and
at the end this algorithm does it as the
good players right and then actually the
trick is we make the algorithm play
against itself so it’s going to train
even better after a million games so you
can train a machine a car this way
because then you have to create a
million debts before figuring out that
that’s not really what you want from the
car so you have to find another way to
do it so the way the tech tree is
developing now is this you actually
raise algorithms the way the same ways
you raise children exactly the same way
with one catch you cannot wait 20 years
so you have to find ways to accelerate
that training so there are a couple of
ways that actually work well one that I
like a lot is Georgia Tech in the u.s.
is that with the create stories about
the behavior of people in that situation
thousands of stories it just narratives
and you have algorithms learned from
these narratives what are the code the
moral code or the choices that people
have made in those situations and then
that way you actually have the machine
learn what humans have done in a
terrible situation like that it might
not be the best choice but at least is
as good as humans have done it and from
there you can construct so there are a
few experiments going on at this moment
one that I like a lot is an algorithm
like this which has been trained to
recognize what is the choice is the
codes that you use when you go out on a
date and you go to the movies so
thousands of people have constructed
narratives about that event and guess
what the algorithm detects a number of
things but what also detects is that a
key element in that schema is this so
apparently for that algorithm kissing is
an extremely important element of going
out in a day to the movies based on the
stories
of many people this is a way of learning
you know what if an algorithm turns out
and suggested kissing I think we’re in a
good place
thank you
[Applause]
Please follow and like us: