
in October 2015 alphago a deep learning
network that was trained to play the
board game go beat the then reigning
European go champion fan y5 to zero this
was the first time that a computer
program had ever beaten a professional
go player five months later an improved
version of alphago would go on to play
the then top player in the world Lise
Aldo four to one now to put this in the
kind of context of why this was such an
impressive feat for a computer program
to accomplish the estimated number of
possible legal go positions on the board
is somewhere around two times ten to the
hundred and seventy possibilities to
give an idea of what that actually means
that’s the actual amount of zeroes of
how many possible positions ago board
can have now this isn’t so much about if
the program can explicitly represent all
of these possibilities because for
computational reasons that’s near
impossible what’s important is that the
program is one that intuitively plays
the game in much the same way that a
professional go player does now alphago
was developed using a combination of two
different methods one is supervised
learning where programmers take expert
datasets and this came in this case
games play by master go players
strategies developed over the years and
encode knowledge representations into
the system and this is used in
conjunction with reinforcement learning
which is the method by which the program
is trained on optimal next steps from
those already pre-programmed games into
a series of rewards and punishments now
fast-forward further this past December
those working at deep mind developing
alphago created a new program alphago
zero and this uses something slightly
different it takes a blank slate
approach so instead of explicitly
training the next steps in the game or
the best possible combinations and moves
the trainer’s take the basic rules of
the game and teach the program how to
learn itself so how does it go
developing its own strategies and now in
the conclusion of their paper on this
program the author’s state humankind’s
have accumulated go knowledge from
millions of games played over thousands
of years collectively distilling into
patterns proverbs and books and the
space of a few days starting tabla rasa
a blank slate
alphago zero was able to rediscover much
of the go knowledge as well as novel
strategies that provide new insights
into the oldest of games now what does
it mean for a computer program to
discover go knowledge or to develop
novel strategies and how is it that the
programmers themselves didn’t know how
this occurred and this really brings us
to the topic of what I want to talk with
you guys about today and that’s really
the role of machine intelligence and
these types of deep neural networks and
kind of the future of knowledge and what
does it mean for us to know something
and now I don’t really want to talk
about any kind of like singular idea
there’s not one thing I want to talk to
you about I want to motivate a way of
thinking about this thinking about the
role of these types of artificial
intelligences in our society and in how
we produce knowledge now ultimately this
is an issue of automation and humans
have been trying to automate our
activities since the beginning of human
civilization when we toiled in the
fields producing crops we try to do that
at a greater scale so we delegated a
task to be suburban quite cruelly but we
did it and that was an automation of the
intelligent activity of plowing fields
fast forward and in the Industrial
Revolution we automated the task of
manufacturing goods and we did so at a
dizzying pace freeing up time for humans
to perform other activities and then
with the advent of computers we
automated the process of arithmetic in
logical operations instead of humans
having to laborious ly calculate by hand
these difficult mathematical equations
we can automate that with computers
through a symphony of electrical buzzing
and now with artificial intelligence
we’ve automated problem solving tasks at
a level previously unimaginable
but I want to talk about the automation
of not only intelligence as in
problem-solving and a task at hand but
rather the automation of knowledge and
what does that look like now I want to
stay agnostic about whether or not
machine knowledge in this way is
inevitable or whether the stirrings are
already occurring and I rather want to
pose the kind of hypothetical question
I’m a philosopher so we’re interested in
these types of hypothetical questions to
really get a grip before what I think
we’re already seeing is going to occur
and that’s really what is the role of
this sort of artificially produced
knowledge in our society in general now
there’s a couple of questions that I
want you to keep in the back of your
heads while I discuss these types of
deep neural networks and that’s really a
question about yourself so how do you
know what you know what does it mean to
know something and this might seem like
kind of a typically confusing and
annoying philosophical question because
it is it is a philosophical question but
it’s something I think we need to be
asking ourselves in these types of
situations and these situations where
it’s not humans creating knowledge like
in the case of alphago zero it created
new NGO knowledge or rediscovered NGO
knowledge and novel strategies so what
does it mean for that program to know
and more importantly how do we
incorporate this type of knowledge into
our own bodies of knowledge what are the
what are the kind of criteria that we
need to select the kind of knowledge it
should be brought into our society and
that that shouldn’t now with the rods
alphago 0 and he’s alphago programs or
say any other deep neural network that
uses a similar style of architecture you
might be asking yourself if the
programmers and developers created this
system then isn’t it really them that
produced the knowledge and I think at
this moment it’s good to take a step
back and realize if the developers could
track kind of each step of the system
took to produce that knowledge or to
come to some sort of decision and the
answer is probably yes it is the
programmers or the developers in that
scenario that produced the knowledge but
that’s not the case of these types of
systems like alphago zero as we saw with
deep mind and the conclusion of their
pain
/ they stay said it came up with novel
strategies and rediscovered this
knowledge but they themselves couldn’t
actually track how that happened
so I really want to ask this question
what is the role of that type of
knowledge in our society now there’s a
name for these types of systems these
types of systems that are very complex
and may have a number of kind of hidden
layers and how they come into a decision
and those are blackbox systems now
they’re called blackbox systems
precisely because you can’t crack them
open and kind of discover how each step
was taken to come to some outputs and
that’s important because these systems
weren’t built with the idea of
transparency in mind they were built for
efficiency so we shouldn’t be
particularly surprised when we look at
these types of really complex neural
networks and we can’t figure out exactly
how it got from stage one to the end to
the end result stage two and I think
this sort of disconnects between the
outcomes of these blackbox systems and
our own kind of expectations of what we
would assume these systems would produce
kind of creates this intelligence jet
lag or we’re always playing catch-up the
conclusions of these really
sophisticated deep neural networks and
what we would have expected the outcomes
to be and I think this intelligence jet
lag between the really task-specific and
optimized efficient problem-solving that
these types of systems do and the more
kind of general and nuanced reasons that
humans generally look for and our
explanations presents a really
interesting challenge for us and how we
see the role of knowledge in our society
and specifically artificially produced
knowledge in our society and I think
what’s mostly important about this too
is in situations with real life
consequences
now with the alphago 0 if the program
was to make some sort of mistake in
moving a piece I don’t think any of us
would consider that to have sort of dire
ethical impacts or general devastating
impacts on society but what happens when
these types of deep neural networks get
applied to situations
with real-life consequences where it’s
not merely a quantitative judgment on
where to place a piece on a board but
rather a qualitative one and we’re
already sort of at the point where this
is happening so these types of black box
systems already use a number of
real-life scenarios with real-life
consequences say for instance in
determining whether or not someone will
default on their loan the likelihood
that they would or in cases of criminal
recidivism and determining sentencing
for particular cases the likelihood that
someone will be commit a crime now this
brings us to another kind of issue it’s
being discussed right now and that’s the
issue of algorithmic bias so the issue
of datasets themselves being biased
because humans collect data and we have
biases that we can’t control for and
these outcomes say in the case of the
criminal recidivism applying these deep
neural networks they found that some of
these systems have racial bias now this
is a situation where there’s real-life
consequences a situation where it’s not
merely a quantitative judgment but a
qualitative one where people’s lives are
impacted and a really meaningful and
important way and this also highlights
another issue and that’s sort of the
disconnect between our own motivations
in using these programs save for
efficiency and sentient in trial
sentencing or our motivations to be more
equitable and just but the motivations
of these systems not matching those of
our own we’re in the case of criminal
recidivism and the likelihood at someone
will be me commit a crime they’ve been
found to have racial bias and that’s an
unjust or inequitable outcome of these
programs and we need to start thinking
about how those motivations and our own
match or how to make them match now
there’s kind of a problem which I think
highlights this really well that sort of
insert itself into the popular sphere
and that’s what Sharlee problem I’m not
sure if any of you have heard of this
before I think it was on the good place
at some points
well the trolley problem is this imagine
that there’s a train sort of barreling
out of control and on that train tracks
there’s five people tied now you’re
standing next to a lever
that if you polled would switch the
tracks the train runs on avoiding those
five people now here’s the problem on
that train tracks that you can switch to
there’s one person tied so your choice
is do I let the Train continue on its
way and kill five innocent people tied
to the track or do i wilfully pull the
lever and condemn one person to a
gruesome fate now this is a tough
ethical decision and in that situation
hopefully you would act you would either
choose not to pull the track or pull the
lever for whatever reason or you would
choose to save five people and condemn
one person to death now more importantly
for this discussion for what we’re
talking about today how would you
actually come to that decision how would
you know which decision to make and I
think reflecting on our own and of
decision process and how we get to this
gives us clues where we should expect
out of these very complex deep neural
networks for one say you pull the lever
you might access all the different steps
that got you to that point why you would
pull the lever or not pull the lever you
might also offer some sort of
justification some reason why you think
it’s better so willfully condemned one
person to death while letting fire are
saving five others or vice versa
you might also offer up some sort of
motivations and this is what’s kind of
interesting so we might say in a
situation I was motivated to save as
many lives as possible so I decided to
pull that lever even though that meant
murdering a human on that track but for
all you know that one person on the
track or for all we know was your mortal
enemy and your motivations weren’t
particularly just because you wanted
that person to die in the end but the
point is is that you’re able to access
and justify and give reasons for your
motivations in this situation and I
think you know with the way that this
technology is going with the way that
neural networks are being applied to our
lives in general these are sort of
situations that will come up say in the
case of driverless cars where it might
be offered with two unpalatable choices
like in the trolley problem and it has
to make a decision
and we should expect it to make a
decision and make it make the right one
we should also I think demand more than
this we should demand that it has some
ability to not only access its own steps
and States justify its decision and
offer some indication of the motivations
of how it came to that decision just as
we would if you are in that situation
and I think really this leads us to the
point of demanding some sort of criteria
for knowledge and these types of
artificial systems and this is really
the way of thinking that I want to
impress upon you is that we need to
start thinking about how we assess
knowledge and these very complex deep
neural networks now I want to give sort
of a caveat
remember this was a hypothetical
question in the beginning to kind of
motivate thinking about what it means to
artificially produce knowledge in this
way
artificial knowledge or machine
knowledge in this sense is probably a
long ways off
alphago zero gives us some indications
that the stirrings are already occurring
and I think this is the perfect time to
start asking these questions to sort of
get ahead of the curve on what it means
for machines to produce knowledge in
this way but it’s a difficult task
there’s nothing easy about accomplishing
this so I think it’s going to take kind
of all hands on deck and now for people
in general say people who aren’t
developing these types of systems who
aren’t looking at the philosophical
implications it’s still important that
we ask ourselves the question of what we
should expect out of artificial
intelligence and machine knowledge in
this way because it has an impact on our
lives as we already saw with how it’s
being applied to cases of defaulting on
loans or criminal recidivism or in these
ethical situations like a hypothetical
trolley problem these have real-life
consequences and impacts on our society
and that affects all of us so I hope
that after today I’ve motivated you
enough to start thinking about these
worries of machine intelligence and the
future of knowledge thank you
you [Applause]
network that was trained to play the
board game go beat the then reigning
European go champion fan y5 to zero this
was the first time that a computer
program had ever beaten a professional
go player five months later an improved
version of alphago would go on to play
the then top player in the world Lise
Aldo four to one now to put this in the
kind of context of why this was such an
impressive feat for a computer program
to accomplish the estimated number of
possible legal go positions on the board
is somewhere around two times ten to the
hundred and seventy possibilities to
give an idea of what that actually means
that’s the actual amount of zeroes of
how many possible positions ago board
can have now this isn’t so much about if
the program can explicitly represent all
of these possibilities because for
computational reasons that’s near
impossible what’s important is that the
program is one that intuitively plays
the game in much the same way that a
professional go player does now alphago
was developed using a combination of two
different methods one is supervised
learning where programmers take expert
datasets and this came in this case
games play by master go players
strategies developed over the years and
encode knowledge representations into
the system and this is used in
conjunction with reinforcement learning
which is the method by which the program
is trained on optimal next steps from
those already pre-programmed games into
a series of rewards and punishments now
fast-forward further this past December
those working at deep mind developing
alphago created a new program alphago
zero and this uses something slightly
different it takes a blank slate
approach so instead of explicitly
training the next steps in the game or
the best possible combinations and moves
the trainer’s take the basic rules of
the game and teach the program how to
learn itself so how does it go
developing its own strategies and now in
the conclusion of their paper on this
program the author’s state humankind’s
have accumulated go knowledge from
millions of games played over thousands
of years collectively distilling into
patterns proverbs and books and the
space of a few days starting tabla rasa
a blank slate
alphago zero was able to rediscover much
of the go knowledge as well as novel
strategies that provide new insights
into the oldest of games now what does
it mean for a computer program to
discover go knowledge or to develop
novel strategies and how is it that the
programmers themselves didn’t know how
this occurred and this really brings us
to the topic of what I want to talk with
you guys about today and that’s really
the role of machine intelligence and
these types of deep neural networks and
kind of the future of knowledge and what
does it mean for us to know something
and now I don’t really want to talk
about any kind of like singular idea
there’s not one thing I want to talk to
you about I want to motivate a way of
thinking about this thinking about the
role of these types of artificial
intelligences in our society and in how
we produce knowledge now ultimately this
is an issue of automation and humans
have been trying to automate our
activities since the beginning of human
civilization when we toiled in the
fields producing crops we try to do that
at a greater scale so we delegated a
task to be suburban quite cruelly but we
did it and that was an automation of the
intelligent activity of plowing fields
fast forward and in the Industrial
Revolution we automated the task of
manufacturing goods and we did so at a
dizzying pace freeing up time for humans
to perform other activities and then
with the advent of computers we
automated the process of arithmetic in
logical operations instead of humans
having to laborious ly calculate by hand
these difficult mathematical equations
we can automate that with computers
through a symphony of electrical buzzing
and now with artificial intelligence
we’ve automated problem solving tasks at
a level previously unimaginable
but I want to talk about the automation
of not only intelligence as in
problem-solving and a task at hand but
rather the automation of knowledge and
what does that look like now I want to
stay agnostic about whether or not
machine knowledge in this way is
inevitable or whether the stirrings are
already occurring and I rather want to
pose the kind of hypothetical question
I’m a philosopher so we’re interested in
these types of hypothetical questions to
really get a grip before what I think
we’re already seeing is going to occur
and that’s really what is the role of
this sort of artificially produced
knowledge in our society in general now
there’s a couple of questions that I
want you to keep in the back of your
heads while I discuss these types of
deep neural networks and that’s really a
question about yourself so how do you
know what you know what does it mean to
know something and this might seem like
kind of a typically confusing and
annoying philosophical question because
it is it is a philosophical question but
it’s something I think we need to be
asking ourselves in these types of
situations and these situations where
it’s not humans creating knowledge like
in the case of alphago zero it created
new NGO knowledge or rediscovered NGO
knowledge and novel strategies so what
does it mean for that program to know
and more importantly how do we
incorporate this type of knowledge into
our own bodies of knowledge what are the
what are the kind of criteria that we
need to select the kind of knowledge it
should be brought into our society and
that that shouldn’t now with the rods
alphago 0 and he’s alphago programs or
say any other deep neural network that
uses a similar style of architecture you
might be asking yourself if the
programmers and developers created this
system then isn’t it really them that
produced the knowledge and I think at
this moment it’s good to take a step
back and realize if the developers could
track kind of each step of the system
took to produce that knowledge or to
come to some sort of decision and the
answer is probably yes it is the
programmers or the developers in that
scenario that produced the knowledge but
that’s not the case of these types of
systems like alphago zero as we saw with
deep mind and the conclusion of their
pain
/ they stay said it came up with novel
strategies and rediscovered this
knowledge but they themselves couldn’t
actually track how that happened
so I really want to ask this question
what is the role of that type of
knowledge in our society now there’s a
name for these types of systems these
types of systems that are very complex
and may have a number of kind of hidden
layers and how they come into a decision
and those are blackbox systems now
they’re called blackbox systems
precisely because you can’t crack them
open and kind of discover how each step
was taken to come to some outputs and
that’s important because these systems
weren’t built with the idea of
transparency in mind they were built for
efficiency so we shouldn’t be
particularly surprised when we look at
these types of really complex neural
networks and we can’t figure out exactly
how it got from stage one to the end to
the end result stage two and I think
this sort of disconnects between the
outcomes of these blackbox systems and
our own kind of expectations of what we
would assume these systems would produce
kind of creates this intelligence jet
lag or we’re always playing catch-up the
conclusions of these really
sophisticated deep neural networks and
what we would have expected the outcomes
to be and I think this intelligence jet
lag between the really task-specific and
optimized efficient problem-solving that
these types of systems do and the more
kind of general and nuanced reasons that
humans generally look for and our
explanations presents a really
interesting challenge for us and how we
see the role of knowledge in our society
and specifically artificially produced
knowledge in our society and I think
what’s mostly important about this too
is in situations with real life
consequences
now with the alphago 0 if the program
was to make some sort of mistake in
moving a piece I don’t think any of us
would consider that to have sort of dire
ethical impacts or general devastating
impacts on society but what happens when
these types of deep neural networks get
applied to situations
with real-life consequences where it’s
not merely a quantitative judgment on
where to place a piece on a board but
rather a qualitative one and we’re
already sort of at the point where this
is happening so these types of black box
systems already use a number of
real-life scenarios with real-life
consequences say for instance in
determining whether or not someone will
default on their loan the likelihood
that they would or in cases of criminal
recidivism and determining sentencing
for particular cases the likelihood that
someone will be commit a crime now this
brings us to another kind of issue it’s
being discussed right now and that’s the
issue of algorithmic bias so the issue
of datasets themselves being biased
because humans collect data and we have
biases that we can’t control for and
these outcomes say in the case of the
criminal recidivism applying these deep
neural networks they found that some of
these systems have racial bias now this
is a situation where there’s real-life
consequences a situation where it’s not
merely a quantitative judgment but a
qualitative one where people’s lives are
impacted and a really meaningful and
important way and this also highlights
another issue and that’s sort of the
disconnect between our own motivations
in using these programs save for
efficiency and sentient in trial
sentencing or our motivations to be more
equitable and just but the motivations
of these systems not matching those of
our own we’re in the case of criminal
recidivism and the likelihood at someone
will be me commit a crime they’ve been
found to have racial bias and that’s an
unjust or inequitable outcome of these
programs and we need to start thinking
about how those motivations and our own
match or how to make them match now
there’s kind of a problem which I think
highlights this really well that sort of
insert itself into the popular sphere
and that’s what Sharlee problem I’m not
sure if any of you have heard of this
before I think it was on the good place
at some points
well the trolley problem is this imagine
that there’s a train sort of barreling
out of control and on that train tracks
there’s five people tied now you’re
standing next to a lever
that if you polled would switch the
tracks the train runs on avoiding those
five people now here’s the problem on
that train tracks that you can switch to
there’s one person tied so your choice
is do I let the Train continue on its
way and kill five innocent people tied
to the track or do i wilfully pull the
lever and condemn one person to a
gruesome fate now this is a tough
ethical decision and in that situation
hopefully you would act you would either
choose not to pull the track or pull the
lever for whatever reason or you would
choose to save five people and condemn
one person to death now more importantly
for this discussion for what we’re
talking about today how would you
actually come to that decision how would
you know which decision to make and I
think reflecting on our own and of
decision process and how we get to this
gives us clues where we should expect
out of these very complex deep neural
networks for one say you pull the lever
you might access all the different steps
that got you to that point why you would
pull the lever or not pull the lever you
might also offer some sort of
justification some reason why you think
it’s better so willfully condemned one
person to death while letting fire are
saving five others or vice versa
you might also offer up some sort of
motivations and this is what’s kind of
interesting so we might say in a
situation I was motivated to save as
many lives as possible so I decided to
pull that lever even though that meant
murdering a human on that track but for
all you know that one person on the
track or for all we know was your mortal
enemy and your motivations weren’t
particularly just because you wanted
that person to die in the end but the
point is is that you’re able to access
and justify and give reasons for your
motivations in this situation and I
think you know with the way that this
technology is going with the way that
neural networks are being applied to our
lives in general these are sort of
situations that will come up say in the
case of driverless cars where it might
be offered with two unpalatable choices
like in the trolley problem and it has
to make a decision
and we should expect it to make a
decision and make it make the right one
we should also I think demand more than
this we should demand that it has some
ability to not only access its own steps
and States justify its decision and
offer some indication of the motivations
of how it came to that decision just as
we would if you are in that situation
and I think really this leads us to the
point of demanding some sort of criteria
for knowledge and these types of
artificial systems and this is really
the way of thinking that I want to
impress upon you is that we need to
start thinking about how we assess
knowledge and these very complex deep
neural networks now I want to give sort
of a caveat
remember this was a hypothetical
question in the beginning to kind of
motivate thinking about what it means to
artificially produce knowledge in this
way
artificial knowledge or machine
knowledge in this sense is probably a
long ways off
alphago zero gives us some indications
that the stirrings are already occurring
and I think this is the perfect time to
start asking these questions to sort of
get ahead of the curve on what it means
for machines to produce knowledge in
this way but it’s a difficult task
there’s nothing easy about accomplishing
this so I think it’s going to take kind
of all hands on deck and now for people
in general say people who aren’t
developing these types of systems who
aren’t looking at the philosophical
implications it’s still important that
we ask ourselves the question of what we
should expect out of artificial
intelligence and machine knowledge in
this way because it has an impact on our
lives as we already saw with how it’s
being applied to cases of defaulting on
loans or criminal recidivism or in these
ethical situations like a hypothetical
trolley problem these have real-life
consequences and impacts on our society
and that affects all of us so I hope
that after today I’ve motivated you
enough to start thinking about these
worries of machine intelligence and the
future of knowledge thank you
you [Applause]
Please follow and like us: