Error message

Warning: strpos(): Empty needle in jellyfish_country_path_url_outbound_alter() (line 425 of /home/sites/jellyfish-drupal/public_html/sites/all/modules/custom/jellyfish_country_path/jellyfish_country_path.module).

Digital Journeys 2017: Phil Miles, Google (DoubleClick)

Blog | 02 Aug, 2017

Smarter Marketing in an AI First World

Phil Miles, Director at Google DoubleClick reveals what search will look like very soon and how machine learning will improve the way we advertise online.

Video Transcript

Good morning everybody. I'm going to try to and address the room, it's not going to be super easy doing that but I'll do my best, Okay? Firstly I'd just like to say it's a great pleasure to be here today. I've been invited to speak at this event a number of times in the past, never quite been able to do so. So, Rob thanks for having me today. I should just say that Jellyfish are, you know, a great partner of ours in the U.K. When we launched our DoubleClick Certified Marketing Partner program two years ago they were one of the founder members. And I'm sure many of you work with them. I think you're in very safe hands.  
 
So I'd just like to start by saying I talked to a lot of marketers in the U.K. and there's definitely a consistent trend in those conversations in terms of where they want to get to. And I would say that that movement is from a sort of more media-centric, publisher-centric relationship with digital marketing, working with the likes of Google, with, you know, with Facebook, with, you know, with Snap and others and moving towards more of a kind of data-centric, data-driven relationship where they're kind of really understanding kind of, you know, their customers through the full funnel and getting much closer to them through data. What is also underpinning that I would say and what is consistent in all of those conversations is the need to have a tech strategy in order for you to get to a really advanced data strategy. I don't think you can actually get to an advanced data-based strategy without having a tech strategy in place. That can be obviously with us at DoubleClick, it can be with other people as well, but I think that's a journey that marketers are on if they want to get to advanced data-led strategies. And I think fundamentally bringing this to the context of this presentation, increasingly that tech is powered by machine learning. You know, machine learning is really what is driving DoubleClick's innovation in the marketplace. And what we're putting at the heart of our thinking because the amount of data that you need to process when you get to that advanced stage is so great that it's really the machine learning algorithms that you need to get you through that pace and really drive the insights that you want. 
 
So, when we talk about machine learning what are we talking about? Obviously, you know, I'm not an engineer so, you know, I'm not going to give you a super-technical answer. But fundamentally, it's pointing computers at large data sets and giving them the ability to start learning. So you're not actually programming them to go and do a specific thing, you're pointing them at data and then looking, you know, getting them to look effectively for patterns and to start learning based on those patterns. 
 
So firstly let's go back in time a little bit. I studied History so I always like to use historical examples effectively to create context in the present. This is the 1980s, it's less than half a lifetime ago. Yeah? Probably what? 35 years ago, hopefully all of you are going to live to at least a hundred, and this is effectively what computing was in the 1980s. This is "Pong". Who remembers this game, yeah? Some of the older groups. This is Atari "Pong." Look at the clothes, yeah. Brown, what is that? A brown woollen dress and, you know, basic programming, what you're effectively getting here is a set output for a set input, okay. Very, very clear binary kind of like programming. 
 
So, let's fast forward to another Atari game and this is "Breakout." Anybody played Breakout? Okay, yeah. So, the significance of "Breakout" is that this is a machine learning experiment that was done by DeepMind. DeepMind is the Google end company based here in London actually that's driving machine learning thinking really for Google globally. And what they did is they set up their computers to try and play the game of "Breakout" and what they didn't do is tell them what the rules were. They didn't say what the rules were. All they said was, "Maximize your points," okay? That's all they told the computer. So, this is basically after 100 training episodes. It's kind of like, you know, missing a lot, which hasn't really quite figured it out yet. After 200 episodes, okay, it starts to get a bit better, starts to figure out effectively kind of like what it needs to do. And then after 600 episodes it figures out basically that, anybody that's ever played "Breakout" knows is the way to win is to make a little hole on the left-hand side, get the thing at the top and then basically break the board down from above, okay? And it figured this out on its own and basically after 600 goes it's the best player ever. Okay. It's basically the world champion "Breakout" player, okay? And it's a good, fun example of showing you how machine learning effectively works in practice. 
 
So, let's go to a slightly harder example now. Is anybody familiar with the game "Go?" Okay. So "Go" is like, a bit like chess. I think it's about a power to a hundred more complex than chess, I mean, 100x more kind of like moves or a million to the power of a hundred or whatever it is. So it's got more positions than atoms in the universe, okay. That's pretty mindboggling, you know, when you consider how big the universe is. So what DeepMind did is they that they trained effectively or they gave their computers effectively the challenge of playing "Go" and they did it against the world champion. So the world champion, Lee Sedol is the world champion of "Go." And in the first game they every played, the machine beat Lee after 102 moves and it was, or it reached certain victory after 102 moves. In their second ever game, it did it after 37 moves and in the third attempt it like totally smashed it effectively. So it's a really good example of how machine learning can solve really, really complex challenges and fundamentally, when we think about this in a marketing context, how the ability for machines to automate to make decisions is going to overtake our ability to do that from a human perspective. Certainly in things like, you know, optimization of campaigns as an example in something like search. I would say that today, automated bidding and those types of areas is more advanced than the human mind can do. 
 
Okay. So, let's bring this to Google. Hey, we already had a dog example from Fenton. We didn't share notes in advance, but effectively this is a good example of where Google is using machine learning to try and effectively give you what you want when you search for something. So, using deep neural networks, what Google has been able to do through machine learning is basically figure out what the common characteristics are of a dog. So, that if you put into your file, let's say you've got, I don't know, 5,000 photos in your Google Photos account but you're also go within that 120 photos of your dog, if you search for 'dog,' basically Google will kind of figure out what dogs look like and serve you up all of the dogs in your account. Okay? And that is really, really powerful machine learning bringing that to you in a really, really simplified way. And we're also doing that with other products as well. 
 
So for Google Translate, very much powered by machine learning. So a good example here is if you teach…when we taught Google Translate Korean, okay, and we then taught it Japanese, it figured out itself how to translate between Japanese and Korean without ever actually being, without having to go back into English if that makes sense. Yeah? 
 
And then in Google Assistant, do you remember when Google Assistant first came out, you kind of, if you talked to it in a, you know, an American accent you got a better response than you did if you were from Bolton as an example, yeah. Sorry for anybody from Bolton in the room, but it's got way, way better, like, 100, 1,000 times better because it's learned that regional accents, people from Scotland, you know, different really strong accents from, you know, it's figured out effectively that. And it can now really understand that without necessarily kind of like making so many mistakes. And I kind of left the YouTube one in the middle there as an example I guess of where machine learning is still learning. That's being the key thing, yeah. Obviously, we use machine learning to categorize content on YouTube and, you know, as we relatively famously, you know, we've had some challenges with that in the past few months. You know, it does a very, very good job 99.99% of the time but, you know, it's still fundamentally in a learning process and not necessarily fully perfected yet. But I think getting better and better all of the time. 
 
So, to bring this back to kind of where I talked a little about in the introduction and this kind of journey from a, you know, I think marketing is on from this kind of slightly more media-centric approach through a tech-centric period where I think clients are really figuring out what tech stack is in order to allow them to get to a more sort of data-driven marketing maturity. And that's genuinely a shift that we're on and I think on that right-hand side, you're increasingly seeing that being powered by machine learning. 
 
So, let's think about it in terms of actual kind of questions for marketers and some of that complexity. So, what are marketers interested in? They're interested in, in this constantly evolving landscape, in more complexity in terms of customer journey, "Who is my audience? How do I find more people? How do I find my audience? How do I…?" and then with that data, "How do I kind of like start to make decisions based on that audience?" And increasingly at DoubleClick kind of this is how we're thinking about things. So, you know, how to qualify your audience, how to reach it, how to make better decisions and really have machine learning powering all of that for customers. 
 
So let's look at qualifying your audience first. So, the first thing we're really, really keen to do is really help customers, is really help customers bring their CRM data, their first party CRM data together with our data to kind of start doing much more kind of like, you know, much better querying of that data, really understand their customer segments. We have seven products with over a billion users, when we bring customer CRM data to that really we can start to get deep segmentation going and real interest, really start to understand, you know, for instance things like who are the most valuable customers that that customer has? Which channels are really bringing in kind of like the sales? You know, how are things working through the journey? But also, you know, with your kind of data we can then use that to find other consumers that look like your existing customers. And we do that with our Similar Audiences feature. This again is powered by machine learning. You'd never be able to do this kind of manually, you know. This is really, really taking your data and thinking then, "Right, Okay. How can we look in the Google ecosystem?" All of those kind of like billions of users and start to think, "Right, okay, who has similar characteristics to your audience?" How can we build out effectively to your audience? But of course, you don't have to bring your data. You can, we can also use machine learning to find people within our ecosystem who are in market whether that be for, you know, perhaps travel or for credit cards or, you know, people who have strong affinities to those segments which don't necessarily need your data to be able to do that. 
 
So the second area is reaching your audience. And again, you know, this is a highly, highly complex area. So firstly we're helping people to do that across devices; across desktop, mobile, across increasingly connected devices including the television. We're doing that across 76 exchanges here in EMEA and obviously across all media. So from a DoubleClick perspective, we're not just talking about search, we're not just talking about display, video, we're talking about effectively everything. And really when you then get to that customer, how do you make sure that you pay the right price which is the last piece. And we're using 40 signals, over 40 signals to determine the actual exact price that you should pay through our automated kind of bidding algorithm. So I think, again, all of that is happening in real time at real pace and real scale and can only be done through machine learning. 
 
And the last area, is the decision making. So how do you use data to more effectively make decisions? I think one thing here is, I guess there's two slides to it. One is about measurement, really helping you to understand pathways. So, you know, across search, across mobile, across different kind of like channels and devices. And then really attribute value to the different channels across those areas. Obviously, we're able to do that because we can see the full funnel. And then the second area I guess is then putting those insights really in the hands of the marketer, you know. So giving you that information at your fingertips effectively really, really simply so that you can start to make decisions. And again, that sounds really, really easy but it's huge complexity in the background. But fundamentally machine learning that is giving, that's actually generating that and actually we can then put that at your fingertips to make better, better decisions. 
 
Okay. So this is actually the last slide. I think the key thing here basically is that our CEO, has publicly stated that machine learning is effectively our biggest bet as a company. It's the one area where we feel that we can drive competitive advantage, we can really make our products better and really make things better for our users and also for our customers. So we're putting it into everything that we do and really trying to drive that from a competitive advantage perspective. And we do that effectively in three ways. One is through computing power. So 40-plus data-centres globally really having the power from a computing perspective to be able to kind of direct at what is then a huge data set which is effectively those seven properties with over a billion users. And then using technology from deep neural networks particularly kind of from DeepMind to actually kind of really, pull all of that together. And I suppose the key thing is that if you're working with us then you have access effectively to all of this to drive your marketing and make it more effective. Thank you very much.

Contact us to talk digital today. 

Contact
Please enter a valid Name.
Please enter a valid email address.
Please enter a valid Phone.
Please enter a valid Company Name.
  • Select Service
  • Analytics
  • Brand
  • Consultancy
  • Conversion Rate Optimization
  • Display
  • DoubleClick Partnership
  • Email
  • PPC
  • SEO
  • Social
  • UX
  • Video
  • Websites
  • All Services