
artificial intelligence will change our world
world
world machines that can think like humans yet
machines that can think like humans yet
machines that can think like humans yet crunch through massive amounts of data
crunch through massive amounts of data
crunch through massive amounts of data in a very inhuman way may be exactly the
in a very inhuman way may be exactly the
in a very inhuman way may be exactly the tool we need to solve our biggest
tool we need to solve our biggest
tool we need to solve our biggest problems AI may also unlock the
problems AI may also unlock the
problems AI may also unlock the remaining mysteries of the human mind
remaining mysteries of the human mind
remaining mysteries of the human mind and this is what intrigues me the most
and this is what intrigues me the most
and this is what intrigues me the most I’m not a neuroscientist but as a
I’m not a neuroscientist but as a
I’m not a neuroscientist but as a computer scientist I write programs that
computer scientist I write programs that
computer scientist I write programs that can learn learn from their experience
can learn learn from their experience
can learn learn from their experience and change and adapt some of my first
and change and adapt some of my first
and change and adapt some of my first experiments were with a robot that
experiments were with a robot that
experiments were with a robot that needed to avoid obstacles outdoors and I
needed to avoid obstacles outdoors and I
needed to avoid obstacles outdoors and I still remember the amazement that I felt
still remember the amazement that I felt
still remember the amazement that I felt when the robot stopped running into
when the robot stopped running into
when the robot stopped running into trees and instead drove around the trees
trees and instead drove around the trees
trees and instead drove around the trees and down the winding forest trail
and down the winding forest trail
and down the winding forest trail without me changing a single line of
without me changing a single line of
without me changing a single line of code but because the robots program had
code but because the robots program had
code but because the robots program had learned from its experience and changed
learned from its experience and changed
learned from its experience and changed its own behavior
its own behavior
its own behavior a decade later I’m still writing
a decade later I’m still writing
a decade later I’m still writing programs that can learn from experience
programs that can learn from experience
programs that can learn from experience and change and adapt and I’m still
and change and adapt and I’m still
and change and adapt and I’m still amazed to be clear we have not yet
amazed to be clear we have not yet
amazed to be clear we have not yet created true artificial intelligence and
created true artificial intelligence and
created true artificial intelligence and it may be decades more before we do but
it may be decades more before we do but
it may be decades more before we do but we are making progress so let me tell
we are making progress so let me tell
we are making progress so let me tell you about machines that can learn to
you about machines that can learn to
you about machines that can learn to play video games solve mazes and run
play video games solve mazes and run
play video games solve mazes and run through an obstacle course the programs
through an obstacle course the programs
through an obstacle course the programs that I write use a learning algorithm
that I write use a learning algorithm
that I write use a learning algorithm called reinforcement learning we start
called reinforcement learning we start
called reinforcement learning we start with what we call an agent and this is
with what we call an agent and this is
with what we call an agent and this is the part of the algorithm that learns
the part of the algorithm that learns
the part of the algorithm that learns and takes actions and that’s why we call
and takes actions and that’s why we call
and takes actions and that’s why we call it an agent because it has agency an
it an agent because it has agency an
it an agent because it has agency an autonomy the agent interacts with an
autonomy the agent interacts with an
autonomy the agent interacts with an environment it sees observations it
environment it sees observations it
environment it sees observations it takes actions and it gets reward both
takes actions and it gets reward both
takes actions and it gets reward both both positive and negative
both positive and negative
both positive and negative this idea of reinforcement learning is
this idea of reinforcement learning is
this idea of reinforcement learning is simple and powerful it’s based on the
simple and powerful it’s based on the
simple and powerful it’s based on the the same way that a dog learns how to
the same way that a dog learns how to
the same way that a dog learns how to sit or we encourage a child to practice
sit or we encourage a child to practice
sit or we encourage a child to practice piano through strategically given M&Ms
let’s go a little bit deeper the agent
let’s go a little bit deeper the agent that I’m talking about is a mathematical
that I’m talking about is a mathematical
that I’m talking about is a mathematical model and at deep mind we’ve been very
model and at deep mind we’ve been very
model and at deep mind we’ve been very successful in using deep neural networks
successful in using deep neural networks
successful in using deep neural networks with many layers of neurons and millions
with many layers of neurons and millions
with many layers of neurons and millions of connections between those neurons as
of connections between those neurons as
of connections between those neurons as the agent
the agent
the agent thus merging the fields of deep learning
thus merging the fields of deep learning
thus merging the fields of deep learning and reinforcement learning as for the
and reinforcement learning as for the
and reinforcement learning as for the environment we mainly use games at deep
environment we mainly use games at deep
environment we mainly use games at deep mind why is that well games are very
mind why is that well games are very
mind why is that well games are very very useful because we can readily
very useful because we can readily
very useful because we can readily evaluate how well the agent does by
evaluate how well the agent does by
evaluate how well the agent does by pitting it against a human player for
pitting it against a human player for
pitting it against a human player for example or comparing its score to other
example or comparing its score to other
example or comparing its score to other agents but more importantly we can think
agents but more importantly we can think
agents but more importantly we can think of games as being like a microcosm of
of games as being like a microcosm of
of games as being like a microcosm of human ability because they are so
human ability because they are so
human ability because they are so diverse and so ubiquitous across human
diverse and so ubiquitous across human
diverse and so ubiquitous across human culture so they’re incredibly valuable
culture so they’re incredibly valuable
culture so they’re incredibly valuable if what we want is to both develop and
if what we want is to both develop and
if what we want is to both develop and demonstrate artificial intelligence at
demonstrate artificial intelligence at
demonstrate artificial intelligence at deep mind we’ve trained agents to play
deep mind we’ve trained agents to play
deep mind we’ve trained agents to play everything from simple video games to
everything from simple video games to
everything from simple video games to incredibly complex games such as go and
incredibly complex games such as go and
incredibly complex games such as go and chess one particular game I can show to
chess one particular game I can show to
chess one particular game I can show to you is this game it’s an Atari game
you is this game it’s an Atari game
you is this game it’s an Atari game called break out those of you that are a
called break out those of you that are a
called break out those of you that are a child of the 70s or 80s may remember it
child of the 70s or 80s may remember it
child of the 70s or 80s may remember it very well
very well
very well may be new to others of you and the way
may be new to others of you and the way
may be new to others of you and the way an agent learns to play this game is by
an agent learns to play this game is by
an agent learns to play this game is by seeing the pixels on the screen then
seeing the pixels on the screen then
seeing the pixels on the screen then making mathematical calculations which
making mathematical calculations which
making mathematical calculations which propagate through those millions of
propagate through those millions of
propagate through those millions of connections layer by layer in the
connections layer by layer in the
connections layer by layer in the network producing an action which is a
network producing an action which is a
network producing an action which is a command on the joystick now if that
command on the joystick now if that
command on the joystick now if that action is good and results in the agent
action is good and results in the agent
action is good and results in the agent getting more points more reward then the
getting more points more reward then the
getting more points more reward then the learning algorithm reinforces that
learning algorithm reinforces that
learning algorithm reinforces that action by adjusting ever-so-slightly
action by adjusting ever-so-slightly
action by adjusting ever-so-slightly all of the millions of connections in
all of the millions of connections in
all of the millions of connections in that neural network we can see how the
that neural network we can see how the
that neural network we can see how the behavior of the agent changes over time
behavior of the agent changes over time
behavior of the agent changes over time with its experience of the game so at
with its experience of the game so at
with its experience of the game so at first even after a hundred games of
first even after a hundred games of
first even after a hundred games of experience it’s still pretty rubbish it
experience it’s still pretty rubbish it
experience it’s still pretty rubbish it misses the ball all too often but if it
misses the ball all too often but if it
misses the ball all too often but if it keeps on playing after a few hours the
keeps on playing after a few hours the
keeps on playing after a few hours the 300 games then it gets better and better
300 games then it gets better and better
300 games then it gets better and better at this point it’s it about human
at this point it’s it about human
at this point it’s it about human ability can return the ball keep keep
ability can return the ball keep keep
ability can return the ball keep keep alive for a long time if we let it keep
alive for a long time if we let it keep
alive for a long time if we let it keep on learning then after a few more hours
on learning then after a few more hours
on learning then after a few more hours and a lot of play and a lot of games of
and a lot of play and a lot of games of
and a lot of play and a lot of games of experience then it gets really good it
experience then it gets really good it
experience then it gets really good it gets better than a human because it
gets better than a human because it
gets better than a human because it learns how to how to do this trick
learns how to how to do this trick
learns how to how to do this trick called tunneling that is systematically
called tunneling that is systematically
called tunneling that is systematically send the ball to the sides of the wall
send the ball to the sides of the wall
send the ball to the sides of the wall so that it bounces around on top and the
so that it bounces around on top and the
so that it bounces around on top and the agent has less work and more reward the
agent has less work and more reward the
agent has less work and more reward the learning algorithm doesn’t just work on
learning algorithm doesn’t just work on
learning algorithm doesn’t just work on breakout it works on most of the 57
breakout it works on most of the 57
breakout it works on most of the 57 games that we’ve tried it on and
games that we’ve tried it on and
games that we’ve tried it on and achieves superhuman level of play at
achieves superhuman level of play at
achieves superhuman level of play at most of them but not all some are still
most of them but not all some are still
most of them but not all some are still a challenge so if we look at the bottom
a challenge so if we look at the bottom
a challenge so if we look at the bottom of this list we see a game called
of this list we see a game called
of this list we see a game called Montezuma’s Revenge where the agent gets
Montezuma’s Revenge where the agent gets
Montezuma’s Revenge where the agent gets a score of zero which is zero percent
a score of zero which is zero percent
a score of zero which is zero percent relative to a human player why’s
relative to a human player why’s
relative to a human player why’s Montezuma’s Revenge hard well when I
Montezuma’s Revenge hard well when I
Montezuma’s Revenge hard well when I look at this game I immediately infer
look at this game I immediately infer
look at this game I immediately infer that my actions are going to control the
that my actions are going to control the
that my actions are going to control the little dude in the middle with the hat
little dude in the middle with the hat
little dude in the middle with the hat that ropes are for swinging and ladders
that ropes are for swinging and ladders
that ropes are for swinging and ladders for climbing that the skull at the
for climbing that the skull at the
for climbing that the skull at the bottom is probably a probably bad and
bottom is probably a probably bad and
bottom is probably a probably bad and the the key especially since its golden
the the key especially since its golden
the the key especially since its golden is probably good in other words I apply
is probably good in other words I apply
is probably good in other words I apply concepts high-level concepts that I’ve
concepts high-level concepts that I’ve
concepts high-level concepts that I’ve learned in a lifetime to this game and
learned in a lifetime to this game and
learned in a lifetime to this game and immediately understand it an agent
immediately understand it an agent
immediately understand it an agent however that learns only from the pixels
however that learns only from the pixels
however that learns only from the pixels on the screen is never going to get to
on the screen is never going to get to
on the screen is never going to get to this high level of concept my colleague
this high level of concept my colleague
this high level of concept my colleague sasha avez nevets had the insight that
sasha avez nevets had the insight that
sasha avez nevets had the insight that this game could be solved if we could
this game could be solved if we could
this game could be solved if we could break it into simpler problems for
break it into simpler problems for
break it into simpler problems for instance if we could say jump across the
instance if we could say jump across the
instance if we could say jump across the gap get to the ladder get past the skull
gap get to the ladder get past the skull
gap get to the ladder get past the skull and pick up the key then maybe this
and pick up the key then maybe this
and pick up the key then maybe this would be a solvable problem and if you
would be a solvable problem and if you
would be a solvable problem and if you think about it that’s what a teacher
think about it that’s what a teacher
think about it that’s what a teacher does a good teacher breaks down a hard
does a good teacher breaks down a hard
does a good teacher breaks down a hard problem into simpler problems that a
problem into simpler problems that a
problem into simpler problems that a student can learn so we developed an
student can learn so we developed an
student can learn so we developed an architecture that involves training to
architecture that involves training to
architecture that involves training to neural networks to agents at the same
neural networks to agents at the same
neural networks to agents at the same I’m from the same game one of them we
I’m from the same game one of them we
I’m from the same game one of them we call it the teacher produces those
call it the teacher produces those
call it the teacher produces those subproblems learns what good subproblems
subproblems learns what good subproblems
subproblems learns what good subproblems are and sends them to the student the
are and sends them to the student the
are and sends them to the student the student actually takes the actions
student actually takes the actions
student actually takes the actions controls the joystick tries to maximize
controls the joystick tries to maximize
controls the joystick tries to maximize reward in the game but also tries to do
reward in the game but also tries to do
reward in the game but also tries to do what the teacher tells it to and even
what the teacher tells it to and even
what the teacher tells it to and even though these two agents are only trained
though these two agents are only trained
though these two agents are only trained from those same pixels on the screen and
from those same pixels on the screen and
from those same pixels on the screen and with no other information the structure
with no other information the structure
with no other information the structure allows strategy and communication to
allows strategy and communication to
allows strategy and communication to emerge over time and we end up being
emerge over time and we end up being
emerge over time and we end up being able to learn how to play the game if we
able to learn how to play the game if we
able to learn how to play the game if we look at how the agent plays now it’s
look at how the agent plays now it’s
look at how the agent plays now it’s still not as good as a human my 10 year
still not as good as a human my 10 year
still not as good as a human my 10 year old son would not be impressed I think
old son would not be impressed I think
old son would not be impressed I think but it is managing to move around the
but it is managing to move around the
but it is managing to move around the screen it’s managing to get through it
screen it’s managing to get through it
screen it’s managing to get through it even manages to pick up the sword
even manages to pick up the sword
even manages to pick up the sword important piece before dying so we are
important piece before dying so we are
important piece before dying so we are making progress
let me tell you about another challenge
let me tell you about another challenge for our learning algorithms for these
for our learning algorithms for these
for our learning algorithms for these agents and that’s the problem of
agents and that’s the problem of
agents and that’s the problem of learning multiple games one after the
learning multiple games one after the
learning multiple games one after the other we human that humans have a
other we human that humans have a
other we human that humans have a seemingly limitless capacity to learn
seemingly limitless capacity to learn
seemingly limitless capacity to learn games one after the other we learn to
games one after the other we learn to
games one after the other we learn to play piano we learn to play poker we
play piano we learn to play poker we
play piano we learn to play poker we learn to play ping pong all without
learn to play ping pong all without
learn to play ping pong all without forgetting how to learn how to ride the
forgetting how to learn how to ride the
forgetting how to learn how to ride the bike that we learn before any of those
bike that we learn before any of those
bike that we learn before any of those but agents that we train have a very
but agents that we train have a very
but agents that we train have a very hard time learning just two games one
hard time learning just two games one
hard time learning just two games one after the other this is because when we
after the other this is because when we
after the other this is because when we learn on the first game then we saturate
learn on the first game then we saturate
learn on the first game then we saturate the neural connections the millions of
the neural connections the millions of
the neural connections the millions of neural connections they’ll get saturated
neural connections they’ll get saturated
neural connections they’ll get saturated with knowledge of how to play that first
with knowledge of how to play that first
with knowledge of how to play that first game and when we move on to play the
game and when we move on to play the
game and when we move on to play the second game all of those neural
second game all of those neural
second game all of those neural connections are overwritten obliterated
connections are overwritten obliterated
connections are overwritten obliterated with the second game thus producing the
with the second game thus producing the
with the second game thus producing the phenomenon known as catastrophic
phenomenon known as catastrophic
phenomenon known as catastrophic forgetting which is well known for
forgetting which is well known for
forgetting which is well known for neural networks we can see this
neural networks we can see this
neural networks we can see this happening in the simple game of pong so
happening in the simple game of pong so
happening in the simple game of pong so this is a trained pong agent it’s
this is a trained pong agent it’s
this is a trained pong agent it’s superhuman better than any of us because
superhuman better than any of us because
superhuman better than any of us because it can methodically get to a score of
it can methodically get to a score of
it can methodically get to a score of twenty to zero the agent is controlling
twenty to zero the agent is controlling
twenty to zero the agent is controlling the green paddle on the right and it
the green paddle on the right and it
the green paddle on the right and it will go through methodically win with a
will go through methodically win with a
will go through methodically win with a score of twenty to zero on every game
score of twenty to zero on every game
score of twenty to zero on every game that it plays against the computer
that it plays against the computer
that it plays against the computer opponent here what happens when we make
opponent here what happens when we make
opponent here what happens when we make a very small change to the input all
a very small change to the input all
a very small change to the input all we’re going to do is change the paddle
we’re going to do is change the paddle
we’re going to do is change the paddle and the score of the opponent from
and the score of the opponent from
and the score of the opponent from orange to black and now we see that the
orange to black and now we see that the
orange to black and now we see that the agent controlling the green paddle still
agent controlling the green paddle still
agent controlling the green paddle still still moves around but it’s missing the
still moves around but it’s missing the
still moves around but it’s missing the point and missing the ball as well and
point and missing the ball as well and
point and missing the ball as well and it loses with a score of zero to twenty
it loses with a score of zero to twenty
it loses with a score of zero to twenty everytime catastrophic failure so in
everytime catastrophic failure so in
everytime catastrophic failure so in order to solve this we took some
order to solve this we took some
order to solve this we took some inspiration from neuroscience in the
inspiration from neuroscience in the
inspiration from neuroscience in the mouse brain and presumably in our own as
mouse brain and presumably in our own as
mouse brain and presumably in our own as well there’s a process called synaptic
well there’s a process called synaptic
well there’s a process called synaptic consolidation that protects only neural
consolidation that protects only neural
consolidation that protects only neural connections that are formed when a
connections that are formed when a
connections that are formed when a particular new skill is learned and we
particular new skill is learned and we
particular new skill is learned and we developed an algorithm where after
developed an algorithm where after
developed an algorithm where after training on a game we
training on a game we
training on a game we automatically identify and protect only
automatically identify and protect only
automatically identify and protect only the neural connections that are most
the neural connections that are most
the neural connections that are most important for that skill on that game
important for that skill on that game
important for that skill on that game this means that when we move on to
this means that when we move on to
this means that when we move on to playing the second game that there are
playing the second game that there are
playing the second game that there are spare neurons that can be used to learn
spare neurons that can be used to learn
spare neurons that can be used to learn the new game and we’ve preserved the
the new game and we’ve preserved the
the new game and we’ve preserved the knowledge of how to play the old game
knowledge of how to play the old game
knowledge of how to play the old game thus preventing forgetting this is
thus preventing forgetting this is
thus preventing forgetting this is called elastic weight consolidation and
called elastic weight consolidation and
called elastic weight consolidation and when we apply it to our agents then we
when we apply it to our agents then we
when we apply it to our agents then we find that they can learn ten games one
find that they can learn ten games one
find that they can learn ten games one after the other without forgetting and
after the other without forgetting and
after the other without forgetting and all at superhuman level all right so
all at superhuman level all right so
all at superhuman level all right so let’s move on from video games and let’s
let’s move on from video games and let’s
let’s move on from video games and let’s take a look at a more complex
take a look at a more complex
take a look at a more complex environment this maze so consider this
environment this maze so consider this
environment this maze so consider this maze it’s not too hard to solve
maze it’s not too hard to solve
maze it’s not too hard to solve from this perspective clearly but if we
from this perspective clearly but if we
from this perspective clearly but if we are in the maze looking at it from this
are in the maze looking at it from this
are in the maze looking at it from this perspective it becomes very hard for
perspective it becomes very hard for
perspective it becomes very hard for both agents and for humans alike this is
both agents and for humans alike this is
both agents and for humans alike this is because as we move through this maze
because as we move through this maze
because as we move through this maze then we need to both remember where we
then we need to both remember where we
then we need to both remember where we have been and also where we are going to
have been and also where we are going to
have been and also where we are going to go to and that requires memory and the
go to and that requires memory and the
go to and that requires memory and the neural networks that have shown you so
neural networks that have shown you so
neural networks that have shown you so far don’t have memory so what we did in
far don’t have memory so what we did in
far don’t have memory so what we did in order to solve this game was we
order to solve this game was we
order to solve this game was we developed a neural network that has one
developed a neural network that has one
developed a neural network that has one layer in it that is composed of neural
layer in it that is composed of neural
layer in it that is composed of neural memory units and those persist over time
memory units and those persist over time
memory units and those persist over time so that as the agent explores the maze
so that as the agent explores the maze
so that as the agent explores the maze and tries to find that orange and red
and tries to find that orange and red
and tries to find that orange and red object that you see there at the bottom
object that you see there at the bottom
object that you see there at the bottom it can retain memory of where it saw it
it can retain memory of where it saw it
it can retain memory of where it saw it where that location was in the maze and
where that location was in the maze and
where that location was in the maze and then as it continues to move through the
then as it continues to move through the
then as it continues to move through the maze then that memory layer persists
maze then that memory layer persists
maze then that memory layer persists over time and enables the agent to find
over time and enables the agent to find
over time and enables the agent to find that goal again more efficiently and
that goal again more efficiently and
that goal again more efficiently and just to be clear this neural memory is
just to be clear this neural memory is
just to be clear this neural memory is not given for free the agent has to
not given for free the agent has to
not given for free the agent has to learn to use it just like it has to
learn to use it just like it has to
learn to use it just like it has to learn to use all of the rest of their
learn to use all of the rest of their
learn to use all of the rest of their neural network but it does work so we
neural network but it does work so we
neural network but it does work so we can see how the agent after training
can see how the agent after training
can see how the agent after training manages to find its way through this
manages to find its way through this
manages to find its way through this maze zooms around the maze picking up
maze zooms around the maze picking up
maze zooms around the maze picking up apples along the way those are worth one
apples along the way those are worth one
apples along the way those are worth one point in this game that that I made and
point in this game that that I made and
point in this game that that I made and it searches the whole maze to find where
it searches the whole maze to find where
it searches the whole maze to find where the goal is
the goal is
the goal is note that without memory it wouldn’t
note that without memory it wouldn’t
note that without memory it wouldn’t even be able to search the maze
even be able to search the maze
even be able to search the maze effectively it would just run in circles
effectively it would just run in circles
effectively it would just run in circles because it wouldn’t be able to remember
because it wouldn’t be able to remember
because it wouldn’t be able to remember where it had searched once it finds the
where it had searched once it finds the
where it had searched once it finds the goal object it has to find it again and
goal object it has to find it again and
goal object it has to find it again and again and again in this maze that it had
again and again in this maze that it had
again and again in this maze that it had not seen before and we can look at how
not seen before and we can look at how
not seen before and we can look at how long it takes to get back to the goal
long it takes to get back to the goal
long it takes to get back to the goal each successive time and see that it
each successive time and see that it
each successive time and see that it gets lower each time meaning that the
gets lower each time meaning that the
gets lower each time meaning that the agent is using its memory to improve its
agent is using its memory to improve its
agent is using its memory to improve its performance and get back to the goal
performance and get back to the goal
performance and get back to the goal again and again I’ll tell you about one
again and again I’ll tell you about one
again and again I’ll tell you about one last challenge for our learning agents
last challenge for our learning agents
last challenge for our learning agents today and that’s the challenge of
today and that’s the challenge of
today and that’s the challenge of controlling a humanoid figure like our
controlling a humanoid figure like our
controlling a humanoid figure like our own this is challenging for us think
own this is challenging for us think
own this is challenging for us think about how long it takes a baby to learn
about how long it takes a baby to learn
about how long it takes a baby to learn to take those first few steps alone and
to take those first few steps alone and
to take those first few steps alone and it’s an interesting problem my colleague
it’s an interesting problem my colleague
it’s an interesting problem my colleague Nicolas he’s trained an agent to control
Nicolas he’s trained an agent to control
Nicolas he’s trained an agent to control this simulated humanoid figure instead
this simulated humanoid figure instead
this simulated humanoid figure instead of the neural network producing a single
of the neural network producing a single
of the neural network producing a single joystick command the network produces a
joystick command the network produces a
joystick command the network produces a set of 21 different actions continuous
set of 21 different actions continuous
set of 21 different actions continuous values one for each of the different
values one for each of the different
values one for each of the different joints in this simulated humanoid figure
joints in this simulated humanoid figure
joints in this simulated humanoid figure and then the agent trains learns how to
and then the agent trains learns how to
and then the agent trains learns how to control the humanoid so that it can run
control the humanoid so that it can run
control the humanoid so that it can run through an obstacle course which can
through an obstacle course which can
through an obstacle course which can have random walls placed or can have
have random walls placed or can have
have random walls placed or can have gaps of increasing size let’s take a
gaps of increasing size let’s take a
gaps of increasing size let’s take a look at the video of how it does
all right so clearly this is not the
all right so clearly this is not the most sophisticated running agent that
most sophisticated running agent that
most sophisticated running agent that you’ve ever seen and you may think that
you’ve ever seen and you may think that
you’ve ever seen and you may think that it is awkward or inhuman like or simply
it is awkward or inhuman like or simply
it is awkward or inhuman like or simply funny the way it controls the body and
funny the way it controls the body and
funny the way it controls the body and particularly those arms but keep in mind
particularly those arms but keep in mind
particularly those arms but keep in mind that this agent has never been asked to
that this agent has never been asked to
that this agent has never been asked to do anything with its arms like juggle or
do anything with its arms like juggle or
do anything with its arms like juggle or carry a full pint glass nor has it ever
carry a full pint glass nor has it ever
carry a full pint glass nor has it ever seen any human examples to emulate nor
seen any human examples to emulate nor
seen any human examples to emulate nor has it undergone the merciless scrutiny
has it undergone the merciless scrutiny
has it undergone the merciless scrutiny of its peers it’s simply trying to use
of its peers it’s simply trying to use
of its peers it’s simply trying to use its arms to do the task better so it’s
its arms to do the task better so it’s
its arms to do the task better so it’s using them for momentum and for balance
using them for momentum and for balance
using them for momentum and for balance so it’s clearly going to be a little
so it’s clearly going to be a little
so it’s clearly going to be a little while before we can use this technology
while before we can use this technology
while before we can use this technology in order to control complex robots on
in order to control complex robots on
in order to control complex robots on real-world tasks but it is a very
real-world tasks but it is a very
real-world tasks but it is a very interesting research path that we’re
interesting research path that we’re
interesting research path that we’re going down with a lot of potential and
going down with a lot of potential and
going down with a lot of potential and where you may ask is this path actually
where you may ask is this path actually
where you may ask is this path actually going to lead us and why are we trying
going to lead us and why are we trying
going to lead us and why are we trying to solve artificial intelligence at deep
to solve artificial intelligence at deep
to solve artificial intelligence at deep mind I can give you three reasons
mind I can give you three reasons
mind I can give you three reasons personally so the first reason as a
personally so the first reason as a
personally so the first reason as a computer scientist I just am fascinated
computer scientist I just am fascinated
computer scientist I just am fascinated by the challenge of how we write
by the challenge of how we write
by the challenge of how we write programs that can actually learn and
programs that can actually learn and
programs that can actually learn and change themselves and gain intelligence
change themselves and gain intelligence
change themselves and gain intelligence through interactions with a complex
through interactions with a complex
through interactions with a complex world the second reason is as a human I
world the second reason is as a human I
world the second reason is as a human I am deeply curious about how we work
am deeply curious about how we work
am deeply curious about how we work an AI may be exactly the platform we
an AI may be exactly the platform we
an AI may be exactly the platform we need to probe those mysteries of how
need to probe those mysteries of how
need to probe those mysteries of how does a baby learn to stand up and walk
does a baby learn to stand up and walk
does a baby learn to stand up and walk how does a 10 year old master a
how does a 10 year old master a
how does a 10 year old master a videogame how does a 40 year old
videogame how does a 40 year old
videogame how does a 40 year old navigate through a urban maze day after
navigate through a urban maze day after
navigate through a urban maze day after day without getting lost how does an 80
day without getting lost how does an 80
day without getting lost how does an 80 year old remember a lifetime of
year old remember a lifetime of
year old remember a lifetime of experiences and lastly as a human on
experiences and lastly as a human on
experiences and lastly as a human on this planet like all of us I am
this planet like all of us I am
this planet like all of us I am incredibly optimistic that artificial
incredibly optimistic that artificial
incredibly optimistic that artificial intelligence may be the powerful tool
intelligence may be the powerful tool
intelligence may be the powerful tool that we need
that we need
that we need to solve our problems to cure disease to
to solve our problems to cure disease to
to solve our problems to cure disease to ease inequality or to save a warming
ease inequality or to save a warming
ease inequality or to save a warming planet thank you
planet thank you
planet thank you [Applause]
[Applause]
[Applause] [Music]
”
Be First to Comment