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: Jonathan Epstein, Sentient

Blog | 07 Aug, 2017

Evolutionary computation: The 'other' AI

Jonathan Epstein, Senior VP at Sentient, shared how some brands are using machine learning to take testing beyond A / B, fuelling conversion uplifts of circa 38-45%.


Video transcript

Yeah, hi there everyone and thanks for your time today and thanks for Jellyfish for putting on such a tremendous event. It really speaks to their embrace of technology in the marketing vision and their forward-thinking behaviours in realising these visions for their clients. 
We've talked a lot today about machine learning, and machine learning is just one form of AI, but for many people and if you read in the press these days, it's often used synonymously with AI because it's the predominant form of AI. But I wanted to speak about another type of AI called evolutionary computation and what it's doing in the area of marketing and how it might change how you think about how you apply AI today into different parts of your marketing. About 20, 30 years ago, in fact, machine learning neural networks, deep learning, all sort of versions of the same thing were viewed as unlikely to succeed. And 20 or 30 years ago, people thought symbolic systems that represent knowledge were really the future of AI, but deep learning has proven, with the advent of really strong compute and distributing computing, to be very powerful as you've seen today. It's a really powerful technology for perceiving the world. What did someone say? What did that handwriting mean? Siri is based on machine learning, as are many things, understanding patterns in data. If this, then that, how do we look at various inputs and can predict an output. The face, you know, what is that face that you're seeing and how do I then put the snap effects on top of it? It's great for detecting, you know, sort of the flip side of that anomalies in data, things like fraud detection is a very common use of machine learning, as is categorisation is that, "Show me dog pictures" would be a good example of that. 
So it's very powerful and as Alex just talked about, once you've trained and developed a model, it's very fast, right, you get your inputs in and within a second or so, you get your real-time data out. But there are limitations to machine learning as a method. I mean the first is that mainly that these models that are built are generally fixed in their purpose. They're built and designed to achieve a specific decision or goal, but they're not really flexible or adaptable. They're expensive and they're time-consuming to build. 
Now, a hundred data scientists, King is a very profitable company. I play a lot of Candy Crush Soda, I'm currently on level 1272, something like that, really and spend a lotta time and a game's also time-consuming if they're done right and designed well. But in San Francisco, I don't know about here, a hundred data scientists would cost about $50 million a year in salary, so that's just to...and that's what you need if you're really gonna do it at scale. So it's not necessarily within the reach of everyone to hire up an army of data scientists to solve these problems. It also takes time to train 'em all while neural networks work very fast when they're trained or built...sorry, to build and train 'em all when they're done, the actual architecture of sophisticated deep learning networks can take something like four or five PhDs a year to do, so you can do the math, half a million a year per data scientist [SP] is very expensive. They're also a bit of a black box, you can extract rules out of neural networks but in a lot of applications, you don't really know what's going on, and while that...I actually think that's okay for many people in marketing. In other businesses, that's not okay. You wanna understand what are the mechanisms that are, you know, from these inputs I derive these results. And in general, they're used for predictions and not so much decision-making.  
But AI is many different things. It's many different ways to simulate human intelligence. And there are at least five different ways to do it. Some people would say more, things like symbolic reasoning, bayesian statistics And then there's evolution. And evolution is a little tough for people to wrap their arms around with...wrap their arms around because it doesn't sound like intelligence, but yet, if you really think about it, it is the forces of evolution that brought us all here today. No evolution, there would be no brains that we can then copy and model with using deep learning networks. 
So about 45 years ago, a guy named John Holland created something, you know, the first foundations of what are called evolutionary computation, and you also hear evolutionary algorithms, genetic algorithms, these are all sort of in the same bucket. And he looked around and he saw that, you know, if you look at life, there's maybe two to the 1.8 billion know, combinations of DNA, this unfathomably large number, but there's really only nine million plants and animals and they're all adapted and ideal for a given niche, whether it's a baby or a penguin, they're all designed well. That something about natural selection is really good at finding optimal answers or optimal designs if you wanna think of it that way, for a given problem. And so he adapted that into math, something called schema theorem.  
So I actually couldn't tell you what all these variables are, but fundamentally, schema theorem defines the principles of natural selection and talks about how through the active natural selection, things...good traits are propagated through successive generations, bad traits are culled out, and so by the process of evolution and breeding, you can create these sort of optimal answers, optimal designs, optimal solutions to problems. And there's three core concepts that really...are at the basis of them, it's much more refined now, but in Holland's day and still the basis of evolutionary algorithms are three key things. One is a notion of fitness. The fitness goal, what makes something better than something else? Right? So in websites, that might be more revenues per user or games. In financial trading, that might be trading the stock market better. If you're an animal, it's really about getting...having children, getting to childbirth. So that's the goal. 
Combination says if you have two successful designs, ideas, solutions, better than average, whether it's a race horse or a website and you combine them, or when we're feeling spicy, we say breed them. If you breed them together, that some of the children will be better than the parents. It's a way to advance along the performance curve. And then mutation, like in nature, is a really critical function in evolution to make sure that we're not just marching up a hill and getting to the top and calling it a day. Mutation spreads guesses, if you would, around the map of potential solutions for a problem and climbs all hills simultaneously, if you wanna use statistical terms, and finds the best answer, not just a good one. 
And Sentient is one of the masters in evolutionary computation. Now we use deep learning as well, I don't want this to be that we're not pro-deep learning. We have another product that does visual product discovery for large fashion retailers. People like Skechers and Sunglass Hut and Zappos, you can find our technology in use in that site, and it's an ideal application for deep learning, how do products look like and relate to other products. 
Since we're the sort of smallest tech company here, Sentient is maybe not a household name like Google, or IBM, or Apple, or Amazon, or others are, but in the world of AI, we're pretty well-known. We've raised quite a bit of money. We do quite a bit of research in AI, a lot of patents. And we also are distinguished for having run the world's largest instance of evolutionary computation ever run. There are certain papers on that. We've solved some interesting scientific problems, and we use evolutionary computation for many different things. This is a very powerful form of AI for optimisation. How do you get to the best result? We run a hedge fund, we do health work, we do agriculture work, and later, I'll talk about how are we applying this to marketing. But a few of the examples, we run a hedge fund called Sentient Investment Management. This is one of the things that use that world's largest instance of evolution. We evaluate 40 trillion different virtual trading desks a year to find the one, maybe two that trade the best. Forty trillion trading desks. And those trading are combinations of rules that were randomly generated initially but through a process of evolution. And one important point I wanna make is that evolution is not a black box. Evolution, if you know how to...if you're a trader and you knew our scripting language, you can understand exactly the trading rules that are being applied for this trader, which is important by the way if you're in the trading business. But it gets us past that black box problem the same way we can sequence the human genome now. As long as you can sequence it, you can understand it, so it's an advantage there. 
We're using evolution to create a predictor for sepsis, one of the leading killers in the emergency room, and more recently, if you Google Sentient, you'll see a lotta stuff about cyber ad [SP], which is this concept of how do you grow anything anywhere in a closed container? And Sentient's role in this is determining initially for the ERB [SP] at basal, what's the right combination of water, nutrients, light, and shocks to achieve the best yields of basal inside this fish tank like...this fish tank size container. Not just the most though, it has to be flavourful as well, it's a multi-goal optimisation. 
So tremendously powerful and versatile technology. So we're sitting around with our prowess in evolutionary optimisation, we said, "All right, we're saving lives. This is good. We're solving hunger. This is good. How are you gonna make some money here?" And so we decided all these things take a while to achieve. We decided to focus on this whole world of testing and optimisation, something that you've seen in our recent presentation, some others today, is really core to defining the best user experience, right? And if you look at companies like Amazon and Peter Diamandis, if you don't know him, well-known futurist, great newsletter, its experimentation that really has made Amazon the success that it has and that companies that want to compete on that playing ground have to be able to experiment at mass at scale, and we don't all have 400 million monthly active users, right? 
So the problem we're trying to address is that AB testing is slow, right? Even at scale, I think we heard it takes a couple weeks to finish an AB test on these vast amounts of traffic because, you know, individual lists can be low to get to significance. It's laborious, right? You have to set up tests and determine things and build long lists of ideas and prioritise them. And generally, it fails, right? Most tests fail. Now it's different when you have a professional conversion optimisation team, such as Jellyfish has, the hit rates are much better but the industry average for AB tests is something like one out of four, one out of five actually succeed in moving the needle. So a lot of big companies, a lot of companies in your business don't do it because it doesn't generate insights fast enough for the speed of your business that are material against the effort that you have to muster for that. 
So we saw an opportunity to fix this as well. By the way, multivariate testing, people say, "Well it just tests multiple things." It's just slower. You're just time-slicing it more. So we want something called Sentient Ascend. Sentient Ascend's been out since September, so it's been 10 months. It's been a very exciting 10 months because it's a very disruptive product that's changing the face of testing and optimisation in the market today for both small, scrappy companies and some of the world's largest enterprises in travel, in telecom, in retail are using the product now. And using this approach, I'll describe how we do it, you can test many more things at the same time than you ever thought possible on the same traffic you have today. 
So if you take a page like this, a design like this. You all have ideas. Your team has ideas, how do I improve my performance, right, and you have these long lists. And in AB testing, someone's gotta decide which one to do. It's usually the boss, right? It' a certain extent, it's opinion-driven because you gotta pick one. But take this page, there's 13 different things that maybe you've come up with to improve your conversion on this page. And yet, if you look at the math, there's, you know, two to four variants [SP] and three [SP] to get 1.2 million designs. You'll never AB test it. You would never be around long enough. The robots will have taken over and we would have reached our demise long before this test ever would've been done. 
But, you know, using AI, you can solve it differently. So using evolutionary algorithms, as long as you follow the rules of evolution, I'll talk about how we do that, you can solve this problem in just 600 tests, and those 600 tests in turn actually only take 15 different rounds of 40 tests running at the same time, each of which takes less traffic than an AB test. So how is that possible? How is that statistically relevant? People sort of wonder. Here's how we work real quick, I'll pace through this and this, Sentient Ascend. Let's say you had nine ideas for improvement on your website, or your email, or your mobile, you know, your mobile website. Ascend generates nine different versions like most testing platforms, but not just one other one, it generates nine other ones, and you've set a goal. That goal might be revenues per user and we assign users to test groups, similar to other testing systems, and see how they do. 
And as soon as we can, as quickly as we can, we want signal, but it doesn't have to be strong signal, doesn't have to be definitive statistical anything. Needs to be enough to say, "Hey, these five changes individually worked better," and then we do what's called an evolution and we create all the children of the successful parents. So maybe this headline and this layout, this headline, this image, this layout, and this copy, all those combinations so we have a second generation of running around with two variants each. And once again, we test them and the surviving parents still in the game, competing against their young, and we see which ones are best. And once again, we combine and create a new generation. 
Now generations might take a week, they might take a day for really large sites, it might take hours for something like King, might take a couple weeks for smaller websites, but it doesn't take that long. And around the third generation, we start doing mutation. And mutation brings back changes that didn't succeed in the first round. And we do this to avoid false negatives, but we do this mainly to find combinations that you never would've found on your own, and to make sure we're casting guesses all around the solutions map to find out what's best. And then typically what happens, evolution is never done. Evolution will always try to get better, to do better. But you reach with something that's called convergence, assuming your audience doesn't change. So assuming your audience doesn't change. Obviously not true, let's assume it doesn't. You'll hit a point where you've put in your ideas and you get get the best combination, keep trying, you're there. No one ever makes it that far, not yet, not in a year, because the gains they see on the way there are worth harvesting and taking down at any time. For the final 10%, we'll start a whole new set of testing. 
So, you know, how have some people used this? Cosabella's a high-end lingerie retailer, and they distinguish themselves really for testing some major brand elements, right? Ascend has also proven very good at taking lots of small changes and turning them into big results, but in their case, they wanted to look at fundamental message ideas. Are we made in Italy or are we family-owned and operated since 1960? What should the colours be on the button? What should the cart, you know, copy be across the funnel? And in less than 7 or 8 weeks, sorry, in this case, let's go back, in 8 weeks, we achieve this result of about 38% more conversions. They went on to test their email sign-up, and they achieved a record in what can only be viewed as an outlying gain of 2,000% percent in email sign-ups, but generally, 38% is sort of more normative.  
And then you's another example of a company called Above. Now Above is an example of how you can use this kind of vast velocity for testing to achieve high conversion gains without testing anything large. A lotta things you'd never wanna try to test in your AB testing. If you're doing it, a button colour or a font, it's too small. It's not gonna move the needle and it won't. You shouldn't in AB. But when you can test 10 or 20 of these things all at once, it can add up to very significant results. So Above is a next generation sort of publisher built on the affiliate lead model. And they wanna get your name, right? They publish content about careers and how to achieve it. So they set up 42 different variations for 7 items on this, and Ascend runs through this process where it's testing... Get the video starting here. It's cycling through different things evolving. And each time there's a generation, we see two things: one is that you find newer, better designs and then the average results of the test is also going up as you go. So at the end of all of that, they saw, you know, rather than say, "Search for this," or "Search for that," they're using choose and find and sort of things that promise reward was much better, these colours were much better, 45% gain. And this is a kinda consistent response, you see this. We haven't failed yet. We can't say we'll never fail, but when you have 20 or 30 ideas to try at once, you probably have some good ones, evolution will find them, it will find the right combination from them, and it will deliver 10% to 40% results usually in a month or two.  
In the interest of time, I'll skip past this. I wanna talk a little bit about neural evolution because there's no one truth to AI, and the real fun is coming from combining evolution to deep learning. What can we do to solve the problems of deep learning, super powerful technology that takes a long time to train and requires expensive staff, and that's really the exciting work we're doing here. How do you use evolution to architect the deep learning networks themselves, which can be by the way, millions of notes laid out in multiple layers. And so that's what we're doing at Sentient, and a lot of our research is there, and what this will bring about soon is the full personalisation of marketing and ecommerce. It does many other things as well. 
So when you look at how this applies to Sentient Ascend, and by applies, I mean roughly in January, rather than doing testing across an audience or your segment, we throw out your segments. Your five segments, your nine segments, good, but they're like 8-bit video games compared to a PlayStation 4. And now instead you'd say, "What are the characteristics that I wanna consider for each user about themselves, about their visit, about their device, about the weather where they are, whatever it might be?" And Ascend will evolve a neural network that considers each of these signals and then produces the output from your ideas. It's not writing your site, it's taking your ideas and find the one that is best for each user. So it puts us in an era, we're not sure whether to call it personalised optimisation or optimised personalisation, but it will break through the challenges that we heard earlier are limiting us from achieving one-to-one goals. 
So, you know, in conclusion, evolution, very powerful alternative technology, it's changing the other side of marketing. We hear a lot about sort of the acquisition side, but as Jellyfish knows, and they're really leading with other agencies in their field, if you don't control both, the acquisition and the landing page and the optimisation, you're only in charge of half your story, and so we're working with Jellyfish and their conversion team to really build out and bring these capabilities to clients through various trials. So that's it. Thank you very much. Appreciate the opportunity.


Contact us to talk digital today. 

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