been moving extremely quickly in the
last few years do you know how AI will
affect your life in the future I have
some ideas
I’m Jeff Dean I’m a Google senior fellow
at Google that means a couple of things
first means I’m kind of old second it
means that I get to spend my time
working on problems that I think are the
most important for the company I’m
working on artificial intelligence I
lead the Google brain team our
artificial intelligence research team
based in Mountain View California our
team does long-term research on how to
make machines intelligent and we then
work with product teams at Google in
order to make machine products that are
intelligent we want to use intelligence
to augment the abilities of people to
enable us to accomplish more to
eliminate tedious repetitive tasks and
to and allow us to spend more time on
our creative endeavors ai is going to be
more impactful than the invention of the
personal computer and the spread of
mobile phones into your pocket so the
idea of artificial intelligence is not
new it’s been around since the very
earliest days of computing it’s a grand
project to build machines that are
intelligent and there’s many ways of
pursuing this its captivated computer
scientists ever since the most promising
approach though is the area of machine
learning rather than trying to embody
machines with everything they need to
know up front rather we want to enable
them to learn to learn how to learn so
that they can learn from their
observations of the world and to make
inferences based on those observations
the field of deep learning is a
particular kind of machine learning that
I’ll be talking about and it’s been
shown in the last four or five years to
be remarkably effective for a wide
variety of problems although it’s
actually much older before we start
though how do we learn why we learn from
examples of things and from repeated
practice and repeated practice and
Impuls are really key to machine
learning as well so in machine learning
we’re going to expose a system to
examples of the behavior we want it to
have and those examples are going to
teach it it’s going to learn from those
examples how to do something so in this
very simple diagram I have a model where
we’re going to try to teach a computer
to tell whether a photograph contains a
cat or a dog and we’re going to have
examples where we know the right answer
so we have some that we say that’s a cat
that’s a dog that’s a dog that’s a cat
and we’re going to essentially show the
computer an example and say what do you
think they’re going to say well I think
that’s a dog and there you go if you get
the right answer you’re done right just
like doing your homework problems if you
get it right you don’t have to do more
but if you get it wrong we’re going to
make little adjustments to how we
believe about what we believe about the
world and what we believe about this
example to make it more likely that the
next time we see that example or one
like it will get the right answer not
the right wrong one deep learning in
particular has a particular way of doing
this that is really important so it
builds these layers of abstraction
automatically as part of the learning
process where the lowest level things or
things like you know does this part of
the image contain a little splotch of
brown and then as you go up through the
layers blub why are the kinds of things
that it learns get more complicated
things like is there an ear at this part
of the picture or something looks like a
couple of eyes or maybe some whiskers
and those features emerge automatically
as part of the learning process which is
really critical aspect of deep learning
we don’t have to tell it how to tell a
cat from a dog it just learns that
there’s these things called whiskers and
they seem to appear in lots of photos
they seem to appear more often and ones
about cats and that’s really helpful and
those features develop automatically in
the learning process so neural nets
cannot can learn much more than just how
do you tell a cat from a dog they can
learn to distinguish thousands of
different categories of objects you know
tens of thousands this is an ostrich or
a fire truck or a fire boat things like
that they can learn from an audio stream
to give you a transcript of the words
that were said in that audio stream how
cold is it outside they can take in an
English sentence hello
how are you and spit out the
corresponding French sentence Bonjour
they can take in the pixels of an image
and give you more than just a category
about it they can actually write a
sentence a caption if you will about
that picture a blue and yellow train
travelling down the train tracks that
perch hose a pretty high level of
understanding of what’s going on in that
scene one of the great things about deep
learning is that all of those things
that I showed you can be expressed using
a relatively simple set of algorithms
and can be expressed in a common
software framework so we can build a
software framework that enables us to
express all these different learning
problems and then use it over and over
for our research and for our products
and so the system we built this called
tensorflow and we use it internally for
everything that we do for in these this
area and last year we decided we would
open-source it because we wanted people
to have the ability to take this
software download it for free and use it
for their learning problems it’s been
really great to see different things
that people have used it for so here’s
one example there was a Japanese
cucumber farmer and it turns out in when
you harvest your cucumbers you have to
sort them into all kinds of different
categories for sale small ones medium
ones large ones
prickly ones NOT prickly ones straight
ones curved ones this is pretty
complicated I’m pretty time consuming at
harvest time so the farmer was able to
take a camera and using a computer
vision model that he trained with
tensorflow actually have the vision
model determine what category of
cucumber I was looking at and then
rigged it up to some conveyor belts and
some little switches that would push the
cucumber into the right box and so this
eliminated many days of labor that the
farmer and his wife would have to do at
harvest time that’s just one tiny
example of something you can do now
that would be harder before so as I said
neural nets are not new they’ve actually
been around since the 1980s and 1990s
and they showed really promising results
on kind of small toy problems then but
they really couldn’t show great great
results on realistic large problems at
that time and the reason is we just
lacked enough computational power the
process of making these adjustments to
the model on every example and
processing every example many times in
order to build this model of the world
is very computationally intensive and so
we just needed faster computers
fortunately we have faster computers now
computers have been dramatically
improving in performance every year for
the last 30 to 40 years and we’ve
reached this point where neural nets are
actually practical on real problems so
the computer you have in your phone is
now a hundred to a thousand times as
powerful as the computers you had on
your desktop twenty or thirty years ago
and that makes all the difference so we
now have enough compute computation and
just to take an example the field with
computer vision every year there’s a
contest where teams compete to see who
can give the right categories out of a
thousand different categories when given
an image and in 2011 before people were
using neural nets
the winning team got an error rate of
26% which doesn’t sound too good when
you think that humans are at 5% on this
task but fast-forward five just five
years and we’re now at three percent
errors using deep learning and much more
computational power
we’re actually better than humans on
this task that’s a really powerful and
transformative thing think about this
computers can now see and they didn’t
used to be able to if you think back to
the time and evolutionary biology when
computer when animals of all dies that
was likely a time of great change and
incredible amazing things started to
happen and computer we’re now at that
point in computing for example this is
really really useful if you’re trying to
build a robot if you can’t see is really
hard to do stuff
so here we have an example of robots
using deep learning to teach themselves
hand-eye coordination so essentially
there’s a video camera for each robot
gets to look over its shoulder I guess I
don’t know it only has one arm but over
its shoulder and again that model is
going to take the pixel inputs from the
camera and go directly to six torque
motor commands for the different joints
of the robot and they’re essentially
just going to through trial and error
practice picking things up and so you
see they’re just trying things and they
can tell if they succeeded or failed by
whether their gripper closed all the way
or whether they were actually successful
at picking this up and they’re learning
like what kinds of grips work well for
different shape objects based on the
vision that they’re there were
developing and they’re pretty good at it
we bought a bunch of variety packs of
toys and tools on Amazon another area
where I think machine learning has
incredible opportunities is in the area
of healthcare and I’ll give you just one
so diabetic retinopathy is the fastest
growing cause of blindness in the world
there’s 400 million people at risk who
should really get screened basically
every year but but often people don’t
get the screening they should so we
wanted to tackle this with computer
vision and you get an image like this
and traditionally an ophthalmologist
scans this to try to assess how serious
the signs are if there are if there are
any and so we got a large collection of
these eye images and we had human
ophthalmologist labeled them so if you
have a human ophthalmologist to human
ophthalmologists score these they agree
with their rating sixty percent of the
time
slightly more worrisome though is if you
ask the same ophthalmologist to grade
the same image a few hours later they
only agree with themselves 65 percent of
the time and it’s really just a hard
problem right it’s sort of an
interpretation of like how how dark are
those spots should that be a 2 or a 3 in
this rating scale so on so in work
published earlier this week by our group
in the Journal of the American Medical
station we now have a machine learning
model that performs on par perhaps even
slightly better than ophthalmologists at
this task and this is really important
because it can make ophthalmologists
much more efficient they can actually
dedicate their time to the people who
deserve the attention and not spend as
much time on screening screening most of
the people most of whom don’t don’t have
any signs of this will also be more
creative by having tools that understand
the art we’re trying to create that will
be really helpful I’ll show you another
example so Leon Gatti’s and his
colleagues from the University of
Tuebingen
and the Max Planck Institute in Germany
last year published this amazing paper
where they have an algorithm that can
take two things an image photograph and
a painting and what the algorithm does
is it renders that photograph in the
style of that painter so here you see
that same picture rendered in three
different styles automatically by those
three different artists that’s pretty
amazing and I think there’s going to be
a real opportunity for to create tools
for human artists that allow them to
really interact with systems like this
and to more rapidly get the ideas in
their head out into really new and
creative kinds of art remember how the
advance from the 1980s to now was caused
by much more computational power so the
same thing is going to be true in the
future we want more computational power
so that we can train larger models so we
can learn more and so deep learning is
actually transforming how we design and
build computers as well there’s two
interesting properties that neural nets
and learning algorithms were using have
the first is that reduced precision is
okay it’s fine when you’re multiplying
numbers for a null net to say about 1.2
times about point 6 hmm sounds about
like point 7 that’s good enough we don’t
have to spend every last detail of our
computational budget preserving all the
digits of precision that traditional
CPUs and
are designed to build and that that’s
really helpful if you just you yourself
if you needed to multiply a bunch of
things and you were able to be very
approximate like that you could do many
more operations same thing is happening
in neural nets the other thing is that
all the learning systems I showed you
and all the algorithms rely on just a
handful of specific operations they
don’t need the full generality the
general-purpose computers have they want
to be able to do things like matrix
multiplies vector operations things from
linear algebra and that’s about it and
so that allows us to build specialized
computers that can do these things
extremely well and not much else so here
is a system that we’ve built at Google
about in the last three years called the
tensor processing unit it’s essentially
a custom-designed chip that accelerates
real night computations only but because
these kinds of deep learning algorithms
are so applicable to many problems this
is great because we now have something
can speed up those kinds of computations
by an order of magnitude compared to
traditional CPUs and that’s really
really powerful it allows us to use more
powerful models in our products and get
sort of better systems out there let me
take you on a tour of some queries from
the future one of the things that we’ve
observed is that as systems get more
intelligent users expect more from up
from what they can do so which of these
I Emma jiz show symptoms of diabetic
retinopathy well I already showed you we
can actually do that today
describe this video in Spanish we can’t
quite do that we can describe still
images pretty well but not quite moving
videos but that’ll come find me
documents related to reinforcement
learning for robotics and summarize them
in German that’s pretty complicated but
imagine how productive we’ll be able to
be if we’re able to have tools that can
do that please fetch me a cup of tea
from the kitchen having robots operate
in messy environments like your kitchen
is actually pretty complicated but the
very baby steps of things like the
hand-eye coordination work I showed you
are step
along that path in closing AI is going
to help us to be healthier happier more
productive and more creative are you
excited to see what the future of AI
holds I really am thank you [Applause]