Introduction to Machine Learning – AUNZ Academy on Air (26.07.18)

JUSTUS MALLACH: Good morning. Hi and welcome to
everybody tuning in to another Academy
on Air livestream. Very excited to have you. My name is Justus Mallach. I am one of the automations
specialists in Dublin in our EMEA headquarters. And today I’m super, super
excited to talk to you about a topic that’s
super close to my heart that I’m working on
pretty much every day and walk you through
the basics and the stuff that we’re doing with
machine learning. I hope you’re all excited
and as excited as I am. Please feel free to use
the Academy on Air Twitter handle to talk about this and
spread the news so that also other people can
learn what Google is doing in terms of cool
things around machine learning. So let’s get started
and jump right in. So the first question is,
what is the machine learning. Machine learning is a term
that gets thrown around a lot, especially lately over the last,
let’s say, one or two years. It’s the new and cool
buzzword that’s all around. But what does it actually mean? What does it actually do? So first of all, we
have to think back where the whole, hey, I want the
machine to do humanly things, I want the machine to
make my life easier movement comes from. Actually, robots are exactly
what we thought this should be. And the term “robot” was only
coined in 1920 in the Czech Republic during a play. And from that time
until today, people became more and more
creative in, hey, I want to build a machine that
does this and that job for me and making their life easier,
eventually making our economy more automated all around. And machine learning is really
only the latest and greatest thing where we
bring this to life. Machine learning itself is
actually not that young. It’s actually something
that was invented– and the theory behind
it was invented– I think during
the ’60s and ’70s. And only now that we
have the compute power to actually bring all
these theories to life it’s becoming a thing that’s
getting applicable or becoming applicable across a
lot of industries. So it’s not a new thing. It’s actually a
fairly old thing, but it’s only now
coming to life. So let’s look at
all these buzzwords and where this comes from. So you might have seen this. We’re into a famous
company that’s helping people to manage
their expenses and taxes. They use machine learning, for
example, to make their products easier to use. Also, we have SoundCloud. And SoundCloud helps people
discovering new tracks that you might like based on
everything that you have heard or listened to before. Some other stuff
is, for example, Microsoft showcasing that
the internet of things and the Azure cloud,
their products are using more and
more machine learning to help customers win and
deal with the challenges of the internet of things. And also we have
a lot of startups coming around and
trying to apply what they can do with machine
learning to actually help them– or their customers even more– to make better
decisions– for example, here helping men to buy clothes. That’s probably
something for me. So this is really what
the whole buzz is about. It’s about helping people. But what is machine
learning actually? How does it actually work? How does it come to life? And he is a very
small nice experiment and a quick exercise. I know that you guys like
doing the interactive stuff, so let’s do this here. So I want you all to
follow along with me. This is basically machine
learning in a nutshell. So if you get the
concept behind this, you get what machine
learning is all about. Not the technicalities– you
don’t need to be a coder. But you get the idea behind it. And that’s what’s
really important here. So let’s look that the first
piece of the sequence– three and nine. So that’s the first example. Let’s think about how these two
numbers are in a relationship, how they connect to each other. Look at the second
one– four and 16. And the third– eight and 64. So now you might be able
to detect a pattern here. So if I gave you the task
of putting a number where the question mark is right now,
what would you come up with? I’ll give you a few
minutes– or a few moments– to think about this. Three, nine; four,
16; eight, 64. And you’ve probably
guessed right here. The right number is 81. So what was that? What is the idea behind that? The idea is to detect a pattern. And what you’ve seen is
that it’s a square pattern. So if you square
three, you get to nine. And if you square nine– so
nine times nine– you get to 81. So by giving you the
first few examples, I showed you a
repetitive pattern, something that occurs
over, and over, and over again and helps you
understand how something works, how two things– or
even more things; so think about having more
numbers in here, for example, more dimensions in here– how things relate to each other. So whenever I give
you the next– for example– number E
here, with, for example, 10, you would be able
to compute that without me teaching you or
hard coding you to square all the numbers to
get to the result. And this is basically what
machine learning does. It teaches you by example. . It’s the way that humans learn. And exactly that is what we’re
trying to also teach machines. We’re trying machines to learn
from repetitive patterns, from data from the past,
so they can make a decision based on what they learned
to predict something for the future. This is basically machine
learning in a nutshell. So as you see here,
you looked at a to c, learned, and then applied
that learning to example d. So it’s all about
learning from experiences. And this is exactly, as I said,
what machine learning does. Quite cool, isn’t it? So here again, repetition. The computers right now,
if we look into programming and how we programmed apps,
for example, in the past is that we had a
set of instructions that the machine ran
over and over again– if we had a loop, for example– and they just did what
was in the recipe. It’s like cooking
after a recipe, really. And you went
through that recipe, and in the end of
that recipe you always get to the same outcome. But in this case, we’re
not teaching computers exactly what they need to do. We’re giving them past
experiences and data and ask them to come
up with a pattern, to find a pattern that’s
occurring over and over again, so that if we give
them a new situation they’re able to handle it and
come up with a result that is in line with their
past experience, very much like human
beings are learning. So why is this very important? Why is the whole
concept of machine learning so important also when
it comes to online, digital, and marketing? So at first, Google
is a huge believer– and has done a ton
of research– when it comes to machine learning. It is something that will
transform our economy. It’s going to happen
online and offline. And it’s going to help us
making decisions a lot quicker without human beings
actually having to do a lot of the heavy lifting. So it’s all about
making our lives easier and enabling human brains to
actually take care of stuff that human brains are made
for– let’s say, for example, creativity. But the heavy lifting,
and the data analysis, and crunching numbers– something that a lot of
us usually don’t like– that’s done better by a computer
because a computer can make it more efficient and quicker. The second piece is that this
technology will have an impact. And this is something
I said already– it’s going to have an impact
in so, so many industries. It’s going to have an
impact in health care. We have some examples
later on about, as I said, health care, the financial
industry, and an industry– for example– that I have
been in before joining Google, in the energy economy
where we can see these benefits of becoming
more efficient, basically, with respect to the energy
economy, making sure that we’re not wasting precious
resources so we’re getting all the light that we want
and all the heating that we want while on
spending the minimum amount of resources. The third piece– and this
is something I also already touched on– is that machine
learning is something that allows us to solve very,
very complex tasks super quick. It’s something
that we don’t need to hard code a program
for or we don’t need to hard code and
program a computer for every single detail. They can figure out
some of the things by just looking at past data. And by doing that, we can
actually solve problems extremely quickly. So this is why
machine learning is one of the key technologies
for this century or even this millennium. And it’s something that we all
need to pay close attention to, especially if you’re
working in digital, especially if you
work in marketing. You should pay attention to
machine learning because it’s making our life easier, we’re
becoming more efficient, it’s going to save
us resources– money– and it will help us– eventually, hopefully–
increase our return on invest on the stuff
that we’re investing in. So this is why this
session is super important. This session is going to be
a very foundational session on machine learning
itself but also connected to some of the technologies
and some of the stuff that we’re bringing,
for example, to AdWords later this year. So, first takeaway from
this session– and this is probably something you should
all take into your companies– please make machine
learning a priority for you. It’s something for everybody. You can apply machine
learning so easily. Most of these tools
are even open source, if you think about TensorFlow. But that’s really something for
people that are more advanced. And if you don’t
know how to code, there’s still a lot of
opportunities for you to leverage machine learning
and machine learning tools. Because for example, Google is
just providing them for you. And the only thing
that you need is to apply them to your daily
life to make your life easier. So that’s the idea
behind machine learning. Let’s go a little
deeper and let’s look into machine learning. I’ve been mentioning that
it’s like a human learning or a brain learning. Let’s look into how
it actually works. So one of the best
examples to look at that is basically to use pictures. Because humans are
really good, for example, at understanding in
this picture if there’s a cat or a dog in that basket. And our brains are trained
to understand if there is a cat or dog in a basket. But if we go a
little bit deeper, we can also teach machines
how to recognize these things. And the way this
works is by using, for example, a technique
that’s called neural networks. Basically how it works is
that we use a picture– we put in a picture. And then there is
an input layer– that’s the bottom layer here
that has the most dots on it. And this input layer
is understanding what kind of edges
are in the picture– it basically looks
at every pixel and how the pixels
relate to each other, so which pixels are
next to each other. And then this input
is getting condensed into more and more
layers– that’s why it’s called a
deep neural network. And then like it’s going from
layer to layer where neurons– which are nodes,
little small nodes that recognize connections
between the layers– basically light up. And eventually there’s an output
layer that is telling me, hey, it’s very likely that
this is a cat or a dog. So eventually the only
thing that we’re seeing here is some input variables,
some neurons and layers that recognize the– for
example– color of a pixel and a relationship from
one pixel to the other one, and then some statistics
determine how likely this is going to be a dog or a cat. This is in very simple language
how, for example, a computer could recognize that there is
a dog or a cat in a YouTube video. And this is what we’re using,
for example, in YouTube. So this is the more
and more technical way to explain how
machine learning works. This is not important to know
if you’re applying machine learning techniques. It’s just very nice to know. So let’s dive into we already do
use machine learning every day, because that is actually
something that a lot of us, including me, are using everyday
without actually recognizing that we’re doing it. And what’s really important
here is that there’s nothing to be afraid of. There’s nothing that
is going over our head. It’s basically something that
is all around us already. It’s just that buzzword,
“machine learning,” that’s getting thrown around
all day that might make it come across a little bit confusing. But let’s dive
into these examples to get you more familiar with
what applications we have. So we have five examples here. The first one is the
virtual assistant– for example, the
Google Home– which is a very tangible
and Google-y example. The next piece is around traffic
prediction, online fraud, delivery services, and
unique recommendations. These are all
things that usually if we want to do it by hand– like with manual work– or even with the help of a
normal non-machine-learning computer program, they would
take a lot of processing power and a lot of very detailed
and error-prone programming. Probably not even
possible to do it through normal computer programs
without the help of machine learning. So let’s dive into it. The first thing is
a virtual assistant. How this applies
machine learning is by looking into
your prior interactions and learning from them. So for example, if you
might have a Google Home, you might have recognized
that over time with more usage the Google Home
understands and identifies your voice better and better. So it might be that the
device is in your bedroom and your bathroom is
close by but there’s still a room in-between. And if you weren’t able
to just shout at it, “play some music in the morning”
while you’re in the shower, and it didn’t realize
it or recognize it, it might actually recognize
it later on after you talked to it for a while. Because it learns that,
hey, this is how the sound and how the voice of
this person sounds like. So it actually understands you
better and better over time because it realizes
and recognizes the patterns in your language
and in your voice better. This is how the Google
assistant uses, for example, machine learning for
voice recognition. Second example–
this is the Waze app. It’s an app that
helps you getting around traffic, for example. And how it uses
machine learning is by going into
traffic prediction. So it tries to understand
from prior experiences of other drivers, for example,
that also use the app where maybe it might be the fastest
route because it’s the least distance between two points. But, hey, this street
is usually super crowded and we should really
find a way around it even though it’s a little longer. So a normal routing app would
route you through the shortest route, whereas this one would
route you around it, making you actually arrive earlier
because you’re avoiding, for example, a traffic jam. And this is something
where I feel like it’s coming very much to life. Because this is
something that we all do if we’re regular
commuters to work, for example. We know after driving a
route for a number of times that, hey, this is a crossroad
that’s always clogged. Or, hey, there’s always
a traffic jam coming up and I should really avoid this. So in our brains we
already do the same thing that this app is doing
with the machine learning. We apply the learnings
from past experiences and really deviate from
the quickest route. So this is basically
mirroring what we would do as normal human
beings, but in an app. Next example– fraud detection. So for example,
Paypal uses machine learning to protect against
money laundering or fraud. So the way it does it,
it basically looks at, hey, these are the regions,
these are the kinds of people, these are the businesses
that are more prone to being attacked by fraud. And then basically, if you’re
likely to be a victim of fraud, it basically pushes out the
question, hey, did you really spend these 5,000
pounds in Guam yesterday to just really make sure
that it’s you spending that money, not a fraudster. And again, machine
learning here is applied for pattern recognition. Because the most
fraud is probably happening in a similar way. And this way they can
actually detect it without human intervention
and notify you super quickly. Another really, really nice,
very relatable example here– for me at least– is the delivery
services, and for example how Domino’s is using
machine learning. So the way they
do it is actually by predicting based on past
experiences how long it usually takes for the pizza to
arrive at your doorstep. And I have at least
found that they are super accurate with that. And honestly, if you can
exactly when your pizza is going to arrive by looking
at past experiences and how long it
took in the past, this signal makes
you doubt even less that it’s not going to arrive. So if you’re sitting
there on your couch, you’re just waiting for
your pizza to arrive, and you know exactly
when the driver is going to ring the bell,
that gives you peace of mind. And honestly, I’ve been a lot
happier since this feature was implemented. That was a very
nice intangible way of applying machine
learning, at least for me. And something that’s also
very close to my heart is unique recommendations. I have come to love Spotify’s
Discover Weekly promotion. They put together
a playlist of songs just for me based on what
I like, what I didn’t like, what I’m listening to,
what I’m not listening to. And honestly, I’ve
found that they have hit my gusto perfectly. I find myself only listening to
that Discover Weekly playlist over and over again
because it’s so good. And then I start putting
pluses behind the songs that I like the most. And those recommendations are
becoming better, and better, and better. And sometimes they
even pick up new stuff that I probably did not
listen to in a long while. But because they’re
smart, they understand that I don’t only want to
listen to Taylor Swift– maybe sometimes– but
also, for example, to heavy metal, that is a
genre that I also really enjoy. So they put together a very
interesting mix of stuff that I’m very probable to like. And I’ve felt, as I said,
that they hit it very well. And this is also another
very nice, tangible example of machine learning. So these are very
tangible examples of where machine learning
is getting applied. And I’m pretty sure that you
had a few of these experiences as well. And takeaway for me is that
it is just super handy. It is something that is
making my life more convenient and in a business context
is also helping me. So let’s look into
other industries that are going to be disrupted. For example, health care. You know that if you’re
doing CAT scans– so, computer-aided– diagnosis
where you lay into that barrel, people are taking pictures of
your organs, and your brain, and your body in general,
it’s always coming down to one doctor to spot that
dark spot in your organs, or in your intestines
and in your body, that might be cancerous. And eventually the question
is, isn’t it very likely that maybe somebody is missing
some of these dark spots? And as a result, you might
suffer from a disease for the rest of your life. So we have these
beautiful technologies like the computer-aided
diagnosis– the CAT scans. But eventually it comes
down to interpreting the output of these scans. And machines don’t get tired. They can look into every
single pixel of that scan and every single
digit of that output and can understand if there
is probably something going on that should not be going on. So they help doctors, pointing
them into the right directions. And as a result, we all– if we’re in a situation where we
need these kinds of health care diagnostic systems, we
all– benefit from it. Because it’s more
likely that we’re spotting a disease early
where it’s treatable and we can be fine after. Whereas without that, we
might actually miss something and we might actually go
down a very different route. So this is where much
learning comes in super handy because it helps people making
these decisions super quickly. And they’re less prone
to error, for example, due to a long shift
at the hospital. As I said, machines are
not getting tired here. So health care is
a central industry that is probably getting
disrupted by machine learning in the short term because the
benefits are just so high. Another thing or
another industry that is getting disrupted by
machining learning right now is the financial
services industry. So for example, here again
we have a lot of data. We have a lot of, for example,
financial data from the stock markets that is easy
to measure and can be fed into one or more
machine learning models trying to predict in which
way the market is going based on past experiences. It’s like a very, very
experienced and unbiased trader that’s helping people
making decisions so that the wealth and
the assets of people that base their trust in
banks, those assets are developing in the
best possible way. So we can react
to market trends, we can react in real time. Because most of these
systems are actually able to make decisions in real
time and make sure that, hey, this stock should be sold and
this stock should be bought, and eventually making sure that
the wealth of people increases. So financial industry,
big, big impact on this industry through
machine learning. Retail, also a very tangible– with the example
of Amazon, a very tangible example here,
if you look into, for example, the recommendations
that you’re getting. A few years back they’ve been
not that great, to be honest. And then over time, as,
for example, Amazon– but also others– picked up, hey, we
understand now better how people actually
buy things and what they’re buying in sequence. And based on that experience
we can now recommend– or they can now recommend– better things for you so you
are having the best experience when you’re shopping online. And honestly, for retail having
that assistant that is like, oh, yeah, this is
something you’ve bought, this new Pixel phone,
and you should really buy a case for it. And you should look at this one,
because it’s super beautiful. And people like you that
have a similar profile and that we had a similar
footprint from you, they bought this
particular case. And this might
actually be something you’re interested in as well. So for retail also machine
learning comes in super handy and is helping people to
make better decisions. So we’ve covered now that
machine learning is actually something that’s very
easy to be understood. It’s something where we
basically just train machines to look at past data and
then give us an input on, hey, it’s most likely a dog
or it’s most likely a cat. So it helps us making
decisions very, very quickly. So it’s a very tiny
little but very powerful helper for human beings. We’ve also looked
into examples– for example, Spotify, or, for
example, the pizza delivery– where it comes to life and also
which industries are probably going to be
disrupted by it right now or in the near future. Until now this has not been
related to any Google products really, despite mentioning the
Google Home assistant really, really briefly in-between. Now we’re turning into
Google Marketing products, and how they are using
machine learning right now, and what kind of products are
coming out later this year, for example, also leveraging
the power of machine learning. So let’s dive right into it. So first of all,
it’s very important that everything that
we’ve built at Google is actually utilizing
machine learning. So going back to this,
all of our big places that we have all
over the internet are actually using machine
learning right now. So if you look at
the Chrome browser, if you look into the Android
platform, all of them use machine learning. For example, on YouTube,
all these recommendations that you’re getting for
new and interesting videos, machine learning is at its core. So it’s basically at the core
of everything that we do. A few examples. The Google Clips
camera is something that I have enjoyed
playing around with a lot. It’s basically machine
learning in a camera. Basically what you do
is you clip this camera to, for example, your shirt. And then you just walk around,
for example, on a party. And what it basically does,
it is realizing people, it’s realizing, hey, it’s a
dog, or it’s a cat– it’s a pet, for example. And it’s trying to get the
nicest and most beautiful shots from these people
when they’re smiling. And honestly I’ve had
great, great photos taken with this Google Clips camera. So that’s a really
nice application of machine learning that
we can all benefit from. Google Photos in general. You’ve probably seen the
assistant in the Google Photos app. And if you have
not seen it, I urge you to download the Google
Photos app and try it out. It’s actually creating
beautiful collages. It’s creating
beautiful pictures. It’s enhancing those pictures. And the most stunning
feature honestly– for myself at least– is that without tagging any
photo you can just type in, hey, I want to have a
picture from a beach. And then it’s coming up
with all these pictures that you have taken from
beaches all over the world. And it’s showing you
those standing around, people playing
volleyball, without you telling Google Photos
that this picture was taken at a beach somewhere
around the world. So it’s just recognizing,
hey, this is probably sand, and this is probably the
sea, and it’s realizing, hey, this is probably a
beach, and it’s surfacing those pictures for you. So it’s got a really, really
great recognition technology. So it’s using this
past experience to actually come up
with your photos that are matching your query. So please try that out. Also Gmail. Also a feature that I
have come to love recently while I get a lot of emails
from a lot of people, that Smart Reply function
that is looking into the email that you got and
actually recommends you a nice reply to
that specific email. It just makes your
life so much easier because you don’t have to
come up with a response. It’s basically giving
you a very good response within seconds without
you actually having to come up with your own. So this is very, very handy. So this was Google
products in general. Now let’s dive into machine
learning in digital advertising in the last few minutes. So you all take care of ads. And here are three things
that are really important and that showcase really
nicely how machine learning is part of our ads. So the first piece is
that machine learning is increasing the functionality
or gives new functionalities to what was not possible before. Best possible exam
is automated bidding. So you know that with bidding
manually on keywords what you are doing is you’re
saying for everybody, being it a person that
is interesting for me as a business or not interesting
for me as a business– as a buyer– if they type in that
keyword, they are most likely an interesting person. But if somebody is not
an interesting person, you had no way to know that. And the way it’s
working now is that we use the data that
we have from people that are putting in queries. And what we enable you to
do with automated bidding is that we say, OK, if
there is somebody typing in a keyword that is
super interesting for me as a business, we
want to make sure and we will make sure that
the ad for this person is in position
number one and that gives you the biggest likelihood
of that ad being clicked. Whereas for example,
for my little brother it’s typing in the same
keyword as I typed in before. And he is just
typing in the keyword but he’s not really
interested in buying it. He does not have any data set
telling the system that, hey, this is an interesting person
in terms of our business. We will lower it a bit and we
will probably not even enter the auction for this person. So even though he is putting
in the same keyword as I did, he’s probably not getting an ad
because he’s not likely to buy. So we’re saving
money on this person whereas we’re investing it
only where it really counts. So this is a new
functionality that we have in automated
bidding which is now driven by machine learning. The second piece is surfacing
insights from historical data. So for example, in our planning
tools, in our recommendations, in our budgeting
tools you will be able to see, hey, this is
probably going to cost this and that based on prior
experience and historical data. And we will be able to
give you that upfront so you are more certain about
which investment to make. And the last piece
is machine learning is not taking away
anything from you. It’s basically just helping
you to focus on the stuff that human brains are made for. So for example, being able
to craft the right ad, being able to identify the
right business challenge, being able to optimize
a landing site, and coming up with creative
ways to do all of that. This is what we need humans for. But for identifying the right
bid for the exact person, because there’s so many
queries coming in in real time, that’s something that human
beings are not that good at. So we should leave this stuff to
machine using machine learning. And all the other stuff
that we need humans for, that should be up to humans. So this is why machine learning
is an incredibly important part of our ads. So which areas in terms
of Google products are most impacted? Google audiences– so
basically understanding how my audience is behaving,
that’s a vital part. Automated bidding, I
just talked about that. Data-driven attribution–
so understanding which clicks are actually most
critical for a conversion, and surfacing and attribution
attributing conversions to exactly this click. And also stuff like, for
example, universal app campaigns, where we just say,
hey, give us all the creative, give us all the input that
you need, give as a budget, and we will get you
the most app downloads and the best lifetime
value possible. So these are the areas where we
should all use machine learning and apply the machine
learning techniques as much as possible because
they’re just really, really good. And eventually it will
help you increase your ROI. So let’s look a little deeper. In the audience piece we are
able to capture the interests, behaviors, and
likelihood of users to do something because we can
look into their past history with Google. This is where the whole
learning from example is really coming to life. For automated
bidding we have all of these signals from people– like interacting with
their mobile phones, interacting with the web. So we can understand,
again, from past experience and from historical
data if they are a great fit for your
business or not, and hence determine
the best possible ad position for that person. For attribution,
as I said, we are creating bespoke
attribution models for every customer that has
DDA available to determine if this or that ad click
and this or that campaign is really the one that
is increasing your ROI. And again, for UACs, for
Universal App Campaigns, we are based on historical
data able to identify users that are most
likely to download the app and convert in it. So this is where it’s most
important to use machine learning– in the
audience region; in automation, especially
automated bidding; and in the app business
and attribution. So the idea of this
is basically to make machine learning work for you. You put in a key
business objective. You determine your audience. And we will do the rest,
the heavy lifting for you in the middle, making
real-time decisions as you go. So here are some
examples that we already talked about that demonstrate
more efficient marketing. Google similar audiences,
you’ve probably heard about that before. What we do is that based on
prior experience and data we look into people that are
similar to the people that are already in your marketing
lists so we can identify them. The next piece is an
example here for that. So for example,
if you had people there or in your list that are
looking for stuff in Hawaii and for hotels in Hawaii and
they’re booking with you, we can find people that are very
close to exactly that behavior and find them for you. For smart bidding,
as I already said, we help you determine who
is a very important customer for your business and who is
less likely to buy from you. And based on that data we will
surface an ad at the very top or at the very bottom. For universal ad
campaigns, as I said, we are able to actually
help you invest money only where we see a good chance
there is a lot of value behind the customer. And last but not least,
we use machine learning in data-driven attribution. Already talked about that. But it’s extremely important
that you leverage attribution. And if you have it
available, please use the data-driven
attribution models. So this is what we already
have in store for you. And I would urge you to
use the audience signal, the remarketing lists. Use automated bidding,
please do that. Because you are not able to
bid on an individual user basis with manual bidding. This is only something
that our advanced systems can do, like maximize
conversion, like Target CPA, like Target ROAS. Those are the things you
should be looking into. Please talk to your
Google representative for more information here. And also, on top of that, please
use data-driven attribution or any non-last-click
attribution model, or UAC campaigns if
you are selling ads. So let’s look into
stuff that’s going to come out later this year. This is the very exciting
last few minutes here. And last piece
in-market search ads. So we’re using the power of
in-market segments in search ads so that the in-market
ads can help you in your performance-driven
goals and help you understanding which
customers these ads should surface to. So it helps you acquiring
basically new customers that are in-market that
are ready to buy right now. You can focus on those. So those are three really cool
and very nice product launches and changes that are coming
to you later this year. Last thing to say
here is that we will be building a lot
more things and products and giving you updates that have
machine learning at the core. We are building
everything that we’re building right now with
machine learning in mind because we understand that
it just makes everything that we do so much more
effective and efficient. So with that, we’re
actually going to the Q&A to understand what kind
of questions you guys had and see what’s on your mind. So the first
question that we have is how do you implement
on a practical basis if you’re not a coder. Excellent question. And as I mentioned
in the beginning, the most important
thing to note here is that, yes, you can go deep. You can do coding and you can
use, for example, TensorFlow, which is an open source
library for machine learning. But you can also just
dip your toes in it and make use of the stuff that
other people– for example, Google– has already built. So for agencies and for
advertising customers in general, dipping your toe
into data-driven attribution to make sure that you are
understanding, and also we– we’re understanding
which ads and which interactional campaigns
are super meaningful, and where we should double
down, and where we should not invest that much money. So data-driven attribution,
really, really important. So that is the
simplest way probably to make sure that you’re
understanding correctly what’s happening in
your AdWords account. The next thing is
automated bidding. I know that there’s
a few people that had not that great
experiences with automated bidding in the past because
these systems have not been leveraging machine
learning in the past. They were good but in some cases
they weren’t very accurate. So what we did is we basically
revamped all of them. And I would urge
you to use machine learning-based automated bidding
systems like, for example, maximum convergence,
Target CPA, and Target ROAS whenever you have
the chance to do so. Just be mindful that all of
these systems, because they work on past experiences
and historical data, they have a learning period. So make sure that
you give it at least one or two weeks to ramp up. And then also make sure you
look at the conversion delay that you have in pretty
much every campaign to make sure that
you’re not judging the performance of a
campaign that’s running automated bidding prematurely. So this is the most important
thing to keep in mind. Then you’re able to unlock
great results in most cases where you see the conversion
volume increasing and the CPA dropping if you put in a
realistic target metric, like the Target CPA or the
Target ROAS that’s achievable. The last piece is please also
leverage our similar audiences list where you can
find new people based on the data and the historical
lists that you already have. So these three
should be top of mind for every agency at the moment. So let’s look into
the next questions. We have a lot of
questions around what’s the best way to get started. Kyle Fitzpatrick is
asking can machine learning help for
remarketing lists. Absolutely. Machine learning-based
lists– for example, the similar audiences lists– perform great. And I would urge you
to make use of them. [INAUDIBLE] is asking
about the ad suggestions that are coming out later on. He’s asking what kind of ads
can be created on top of that. The way it basically
works is that we’re reassembling the ads
from ads that you already have in your account. So it’s basically not creating
new things from scratch. It’s basically
just pairing things that will have the biggest
projected conversion rate based on prior data. So basically we are
mixing and matching things that are most likely
the best possible answer to your potential
customers question that they put into Google. So we’re mixing and
matching things here. Let me look through the rest
of this sheet real quick. Machine learning will provide
the automated bidding. It means it will help
in increasing ROI. Absolutely. That’s an excellent question. What we see throughout
pretty much every campaign that has a realistic target– so the Target CPA is
set in the right way– and also it’s been given
the right time to learn, you can actually see
this learning period and see if it’s learning. If you go in the campaign view,
and if you have Target CPA, for example, running, you can
see under Bidding Strategy– in the column Bidding
Strategy in the front end– you click on, in this
case, Target CPA. And it will show you what
it’s doing right now. It will show you how
long it’s probably going to take to learn, if it’s
getting disrupted because there were changes made during
the ramp of the campaign. So it’s giving you
a lot of insight already inside of the
AdWords front end. There’s also a very
interesting question around how the traditional
advertising industry can leverage machine learning. I’d say this is
something where you can leverage online marketing. Do your market
research, for example, through understanding how
different keywords perform. So, use AdWords as a marketing
research tool on this case, and make use of the
machine learning tools that are
built into AdWords, and leverage them
for the first start. And then later on try to
learn from past experiences. Try to gather data
as much as you can. And make sure you build a
machine learning around it and tap into programs that
can help you along the way. This is probably covering
the most questions that we had so far. Also make sure that you’re
using a Twitter handle to discuss further on Twitter. And let’s see if we can get
you some answers there as well or you can get yourself
some answers there as well. This was very fun. I hope you liked it and you
enjoyed the session on machine learning. The next session will be more
specific on automated bidding and smart bidding. So stay tuned for that please
and join in the next time. This is me. Thank you so much for tuning in. And, yeah, I’ll hopefully
see you in the next episode. Bye.

Leave a Reply

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