Machine Learning – A New Programming Paradigm


CASSIE KOZYRKOV:
So, have any of you had that moment
where you realize everyone is using
some word, and you have no idea what that
means but you’re too afraid to ask at this point? What even is machine learning? Well, it’s an approach to
making lots of small decisions with data, something involving
algorithmically finding patterns in data. But let me just put it bluntly,
so that we can all get it. It’s a thing
labeler, essentially. That’s what machine learning is. It’s just a thing labeler. So why on earth should you be
excited about thing labeling? Well, here’s my pet,
Huxley, and you’ve just taken in some pretty complicated
data through your senses, and you just know
that that’s a cat. And if I show you a different
image, you’re not fooled– you still know that
that label is cat. Your brain just does this. You don’t even know
how your brain does it. Now, if we wanted
to get a computer to perform this
labeling task for us, if we went the traditional way– you know this,
programmers– we would communicate with the
universe in some way, think really deeply
about the problem, and come up with a model. It’s a model. It’s just a recipe. It’s a set of instructions
that the computer has to follow to get from
the image to the label. And we would have
to hand-craft that. Now, think about what
this recipe would have to be for the computer
to correctly detect that there is a cat in an image. I mean, think about
what you, your brain, actually did with those pixels. Can you express that? Do you know what
your brain is doing? Or have you just had the
benefit of eons of evolution, and it just kind of
figures it out for you? Now, that recipe is
really hard to hand-craft. Wouldn’t it be better if you
could just say to the machine, here, look at a bunch
of examples of cats, look at a bunch of
examples of not cats, and just figure it out yourself. That is what machine
learning is all about. It is a completely different
programming paradigm. Now instead of giving
explicit instructions, you program with examples,
with data, and let that machine learning
algorithm figure it out and stitch that code that
goes in between for you. So why should you actually be
excited by this thing labeling? Well, engineers,
you know we like to get computers
to do stuff for us? There’s a whole class of tasks
where we just cannot express the instructions. They’re ineffable. AI and machine learning are
about automating the ineffable. This is a fundamental
leap in human progress. Now you can get computers to do
tasks where you cannot possibly express the instructions. This is so exciting, and seems
futuristic and difficult, right? And when you think
about machine learning, the one thing I most want you
to think about is microwaves. Who here in the
audience knows how a microwave oven works
well enough to build a new one from scratch? Raise hands. Who can build a new one? Do I see any takers? Any takers? No hands. Yeah, me neither. Who here has never used
one to reheat food? You all have. Even though you have
no idea how it works, you’re still happy
to use one, right? And anyone here feel like
you would have no idea how to get hold of one, I
don’t know, like maybe at Kohl’s, if you needed one? Now, you just told me
you have no idea how it works, so how could you
possibly trust this microwave? Well, you’re not going to
trust it by reading the wiring diagram, are you? You’re going to trust it by
checking that it actually does work, by having a
good idea of what you want it to do for you, and then
tasting what comes out on the other side. And that is exactly what you’ll
do with machine learning. And this is a perfect
analogy for machine learning. Now, what they don’t tell
you about machine learning– when we use that
term, there’s actually two machine learnings, not one. Two completely
different disciplines that are as different
as building microwaves from scratch and
innovating in the kitchen, and you need a completely
different skill sets for these. If you want to be a machine
learning researcher, you are building general purpose
tools for other people to use. And so the bad news there
is that there is quite a lot of education
that you need to get, takes a little while to ramp
up, because how on earth are you going to build a better
microwave than the one that exists already– there’s some pretty
sophisticated stuff out there– if you have no idea how
the current one works? So that’s the bad news. It takes a long time to become
a machine learning researcher. The good news, though, is
if that is your cup of tea, we at Google have a large
machining research shop where we make these
tools for you that are all powered by open source. So we give you those wiring
diagrams as a springboard so that you can go and build
a better microwave than what we have using our blueprints. And TensorFlow, a really popular
machine learning project, started at Google before
we open sourced it because we believe that
the community as a whole is so much more innovative
than just any one of us alone. But let’s face it– most of you are not here
to build general purpose microwaves. You just want to get
cooking already, right? You just want to use these
things to solve business problems, and so you are in the
completely other discipline– applied machine learning. Now, for that, you
do not need a PhD. You don’t need to know
how back propagation works in neural networks any more
than a chef needs to know how a microwave is wired. Instead, what you need is a
really good kitchen to play in, and we at Google
provide that for you. Let me show you some of
the things that we have, and what you’re going
to need kind of depends on what you’re cooking. If you’re cooking with
usual ingredients– data– and you’re making usual dishes– labels, predictions. So say you’re making pizza. Please don’t go reinvent the
concept of pizza from scratch. There’s already
recipes out there, and you can just grab those
recipes and start making pizza. And that is what our
cloud APIs are all about. These are recipes that we’ve
already built, that you can just pick up and start using. You can just plug
them into your apps to make those
intelligent applications. But say you’re cooking
with a little twist. Now it’s gluten
free, vegan pizza. Well, you still
don’t need to– yeah. You still don’t need to
start entirely from scratch. It’s still pizza. You can just adjust what
already exists out there. And so what we suggest for
you here is try Cloud AutoML. Sure, maybe we haven’t
trained our vision system on the clothing type that your
Kohl’s customers are wearing. But no problem– just feed in
your images with your labels into Cloud AutoML and
you’ll be good to go. Only a small
adjustment necessary. Now, say you’re
truly innovating. You have your own completely
unique ingredients. You’re doing something pretty
different in the kitchen. I don’t know, maybe now
you’re making an edible sock that tastes like pizza. Well, machine learning
and innovation are in Google’s core DNA. In fact, I can’t think of a
single one of our consumer products that
doesn’t have machine learning in it somewhere. And so if you are
innovating in the kitchen, if you’re building truly new
recipes, what we have for you is Cloud Machine
Learning Engine. This is access to exactly the
same infrastructure that Google uses to train our models. Incredible. So you could, if you wanted
to, after this session, grab a laptop, open
it up, and test drive our shiny,
gleaming kitchen full of those
appliances ready to go. You could feel the
thrill of a data center and all those cutting
edge algorithms rising up to do your bidding. So solving business problems
with machine learning– applied machine
learning– is far easier than what most people think. What you don’t
realize is that all those courses in university– those are all about how to
build the microwave, not how to just use it to get cooking. And when I hear people say,
oh, I couldn’t possibly get started with machine
learning until I take a course in it, or,
goodness, a whole degree, I can’t help but imagine
that same person saying, I couldn’t possibly
use a microwave until I’ve built one myself. You can just get cooking. It’s not the lack of
degree, or course, or knowledge that’s
holding you back. What’s much, much more
important than that, what you need to get
started is creativity. Figure out what you’re going
to cook because all those tools are already there waiting
for you to use them. Imagine gleaming
kitchens just waiting for you to come play in them. But since this is
a thing labeler, you need to start
by thinking, what on earth do you want labeled? And see how creative
you can get with that. You might imagine the label
being maybe the wave form that sounds like a voice that
could make a hairdresser’s appointment for you. Or you could imagine
a few other cases, like maybe you are a loving
son, and your aging parents sort vegetables by hand, and
you want to save them the trouble of having to
do that on their farm. So maybe you’re thinking
about building a system that will label the vegetables
and sort it automatically so your parents
don’t have to do it. Or maybe you’re worried
about the ingredients that go into baby food, and you
think it would be great if you could automatically check
whether the ingredient is safe or spoiled before it goes in. Or maybe you’re into
the idea of a huge bank with trillions of dollars
in assets countering fraud and money laundering
automatically across millions of accounts. If you imagined that world,
you imagined this world. Well, these are
all real examples of things that have been done
on Google Cloud Platform. The first is a real family
of cucumber farmers, the second is a Japanese
manufacturer of high quality baby food products,
Kewpie, and the third is HSBC, who uses Google
Cloud to run a better business and catch the bad guys
across 27 million customers. Pretty different chefs,
completely different applications. Why is Google Cloud Platform the
right choice for all of them? Well, think about what you
might want in a kitchen. You’d want good engineering. So you probably want to think
about getting a partner whose quality you trust. But think more
long-term than that. If you believe in your
own infinite creativity, if you believe in
the future, you would be crazy to let yourself
get locked into a kitchen where the purveyor sets
it up one time for you and then controls what you
can and can’t have in there, and every appliance has its
own proprietary little socket that you are not
allowed to tamper with. And if your business
changes, then you have to rebuild and
rewire the whole thing. Imagine, instead, the
dream kitchen that becomes what you need
it to be instantly. Complete, universal
wiring, open interfaces, the best of everything,
all the modern appliances always available, and you
can plug them in with no fuss because that wiring’s universal. That, my friends, is the
power of open source, and that is what Google and
Red Hat together believe in. We don’t want to see your
creativity squashed by what seemed like a good
idea yesterday. So we have built for you an
incredible future-proof machine learning kitchen. This kitchen can expand,
contract as you need it. You can plug in
whatever you like. All the most modern appliances
are there and compatible as you need them,
fully customizable. Because none of us can
tell what innovations the future will
bring, and we don’t want you to have to bet on the
direction of progress and lock yourselves in. We want you to be free to
be as creative as all of us can be together. So I hope that you’re
excited to dive in. This machine learning
kitchen is there, ready, waiting for you to get
started, start playing in it. Your machine learning
adventure awaits. Thanks so much.

18 Replies to “Machine Learning – A New Programming Paradigm”

  1. Is this lady in sales. If not, you should put her there. I will buy sand from her while on the beach.

  2. Hmm. I think you should reinvent the wheel. With machine learning you could invent new mathematics. This is an opportunity not to waste on cloud adaptations.
    Instead ask yourself. How can we invent ourselves out of climate change. How can generative machine learning create the next thousands of must have inventions?

  3. Wow… One of the best high level presentations I have ever seen! I wish I had such brilliant marketing people in every team.

  4. so you say, but try convincing your professor that it's okay for you to use an API rather than program the neural network yourself for your semester project.

  5. Such a graceful speaker! Google is really enabling us to transcend into the realm of ML. Thank you for encouraging the humanity to excel.

Leave a Reply

Your email address will not be published. Required fields are marked *