 Thank you. Good morning. A few months ago, the organizers of Big Data Spain asked me to come up with a talk about some of the things that I've done in my career, but especially could I speak about F1? And it's funny that every time I talk to someone about my, you know, the things that I've done in my career, they always end up asking about Formula One. So I was wondering why this is, you know, why do people feel such need to ask about that particular area, of all the other things I've done, which I think some of them are possibly more exciting. And it got me thinking, can I put together some sort of presentation that actually ties up two or three of the industries I've worked on and the commonalities between them and what methodologies I've learned in one that I've carried across to the next. And so I think that one of the things that unites them all in this particular area of Formula One, energy, environment and healthcare, is the driving force behind all of them. And I do not mean money. So passion. Passion is the one thing that actually makes people who work in Formula One or people who work in energy or people who work in healthcare do what they do. It's passion for the race, passion for the fans, for the sport. It could be passion for the environment to have a climate that does not change and we can have an environment that our children, their children can't live in without fear. Or the passion for the people who look after others, who make us feel better when we're poorly. In general, all of the people in these sort of industries have a mission and that mission is either make people happy, make people healthier, make the environment cleaner and all those things are the ones that actually make these sort of industries move together and achieve challenges that possibly couldn't have been achieved without the dedication of all those people. And when you're trying to build something complex and you're using your passion to drive you forward, you will always come across challenges. And these challenges are common in all industries. I'm not claiming these are exclusive to these areas. In fact, I'm sure you'll recognise all of them. I've picked sort of four representative ones that I think will be quite interesting and I'll sort of put some examples or anecdotes on some of these different industries and see how those challenges have actually happened. The first one is what we call the human factor. The human factor has different dimensions to it and the first one, probably the most common ones is that knowledge takes time, experience takes time. It's quite a hard thing to, for example, make it in F1. You have to come in as an apprentice, you have to go from team to team, learn the skills, become either a race engineer or a vehicle scientist or all the different things that you can do within F1. It's quite hard to come in from the top. It's normal to go and take years and years to get to the point where you actually can make a difference. And similarly, we all know how hard it is to actually become a clinician, to become a doctor or a nurse. You have to study for years, then you have to practise for years before you actually let alone in front of a patient to deal with them without any other support. And in energy it's the same if you imagine what we call the network design scientists. So these are people who have studied physics for years and they work out the power electronics and everything it takes to build a massive network that conducts electricity across the whole country or even a continent. That is a very hard job. It takes a long time to learn it and it takes a long time to trust you enough to design a country-wide network that carries power through. The other two are quite related. The people have a natural aversion for machines. Possibly nobody in this audience has that problem, but people who are not like us, they're not scientists or engineers or technologists, find it quite difficult to trust machines. And some people think that if we automate everything, then what's going to happen to jobs? That's the threat of AI everybody's talking about at the minute. What happens if I automate away this particular role I'm doing now? Where am I going to do next? And this is quite an interesting effect that we're seeing as technology advances and it moves across. And I think that one particular case in health tech, a lot of people have this sort of distrust for AI with reason. I mean, we haven't validated that any other decisions that artificial intelligence making on someone's health can be trusted. And there's two camps in this area. One camp thinks that AI will replace doctors in a few years. We won't need them anymore. And you will be able to be diagnosed and treated by a machine. I'm on the other camp. The camp that says that AI, machine learning, and all the technologies that we're learning will empower doctors to do the right jobs and will empower patients to fight the right treatment. And that's not necessarily seeing the doctor. It could be something else like stay home and take some tablets or go and talk to someone about mental health, things like that. So this sort of divide between the two areas is quite telling of what the two sort of groups of people think of what's coming next. The other challenge that I want to talk about is complexity and complexity manifests in lots of different ways. It has three examples of them. And I'd like to tell you a little bit about a story in Formula One that tells you a bit about the complexity space. We think that the complexities in the... Well, most people think the complexity is in the algorithms and how complex they are, but actually sometimes complexity comes in the environment around the algorithm that makes the algorithm almost useless. To give you an example, a couple of years ago, when I was in Formula One, we were the smallest team in the grid. We had tens of times less money than the big guys. And so the money we had, we wanted to spend in technology rather than vehicle engineering. This is because we didn't have hundreds of millions to spend in a car. We had tens of millions to spend in a car. And so even so, there was that difference. We were literally fractions of a second behind the fastest car. The reason that this was possible is because we were using technology to make some decisions that other teams were more traditional about. But the example I want to tell you about complexity and how it makes it very hard for any predictive algorithm to actually work is the race in Brazil in 2016. I remember very distinctly because it was the one time I actually decided to invite my family to come over to the factory to go into the op-center and watch the vehicle scientists and the strategists and everybody run around the screens making decisions, shouting things on the radio. I thought it would be quite exciting. At that point, we were number 10 in the 11 teams. And for those of you who don't know about Formula One, only the top 10 teams get the prize money. The 11th team gets nothing. And so it can be a game changer if you're not in that top 10. We were top 10 for most of the season. Then Brazil came. And it was a horrible day. It was raining all the time. We thought the race would not happen. We had prepared for it. We had simulated all the different scenarios. We thought we were ready. And then the race started. We were in a good position. We thought it was going to be cancelled at some point and we were in a position that we would retain our 10th position. The race stopped because of an accident and we thought that was it. Then the stewards decided to restart it. The race started again. Again, we retained our position. We were in the money. If we stayed where we were, we didn't let the bad guys behind us take us over. We would have retained the top 10. And so there was another crash and another stop. And at that point, we thought, right, we're in it. And now the stewards decided to start again. And this third time, literally almost the two hours of the whole race were up. We thought, you know, it's now or never, sort of thing. And in the last few laps, our direct competitors, which are number 11 at the time, over two casts, the race was over, the team was over, the money ran out. And so that's why I'm not in F1 anymore. But that example is just to tell you how complex an environment can be that doesn't matter how much simulation or how much AI you throw at it, you will always come across situations where it's very, very hard to predict what you can do and to fix it quickly enough. The second condition, the physical limitation is actually a very interesting problem and it happens in all industries. There's things that you just can't do physically. We don't have the technology or we don't have the knowledge to solve a problem. However, if you think hard enough, you can come up with solutions. And one good example when I was back in energy after Formula One, one of the big problems that we all have, we all want, I think, 100% green energy running our grids, right? We want wind, we want solar, we want to get rid of coal and gas and all those things that generate carbon. And so the problem we have with the national grids in every country is that they were designed 100 years ago. They were designed for these power sources that we call inertial. And you can imagine an inertial power source to be this big power station running coal or gas which heats up water. The water makes this big weight spin very slowly and that spins actually an electromagnet that's generating electricity down the grid. And that's what we call inertial because it takes a long time to start and a long time to stop. So the way we see this wave, 50 hertz to 60 hertz depends on what country you're in of energy going through the grid. When that 50 hertz to 60 hertz gets a little bit off, like 49.6 or 51 or something like that, the grid is at risk. The grid can actually black out. It's not designed to be outside of this very narrow band of tolerance. And so you'll have these guys in power stations literally twiddling knobs to actually say oh, we're getting a little bit higher, let's change the load and balance to keep it straight. Then along came solar and along came wind and those are what we call non-inertial. The reason why it's non-inertial is because there's no big way to move. They just switch on and they switch off. A cloud comes over a solar panel, everything stops. A wind gust comes over the wind turbine and a lot of power is generated. So this generates a lot of noise from the grid and so that nice smooth curve becomes really noisy and that noise is dangerous if you get outside that narrow band. And so general country-wide networks limit the percentage of green energy that can go into the grid to preserve its stability. And so I think in the UK I mean it's about 40%. So when you get more generation from solar, which is really what you want, they actually won't let you put into the grid because you're risking the stability of the grid. So a few of us got together, how can we solve this problem? This is a physical limitation of the grid. Can we do something to actually avoid this situation and increase 80, 90, 100% of green energy into the grid? And so we came out with the concept of a smart grid. And what a smart grid is, it's a bunch of devices that you put everywhere in the country, wherever you can actually consume load, for example, a fridge and air conditioning, a heater, a radiator, things like that. And then the things that you can switch off and on very quickly without the user actually noticing. So if I switch off your radiator for two seconds, you won't tell the difference. The room will continue to be warm. And so we have this really complex network now of millions of devices out there that are making these decisions really quickly to actually try and balance the grid. So when there's a lot of noise in the grid, these devices will switch on and off very, very quickly to try and do the noise cancellation like you were doing headphones, for example. So that's a good example of using algorithms to actually try and solve the sort of physical limitations, limitation of a system. And the third one is like the more understood, I mean, Formula One is incredibly regulated. The rulebook is bigger than, you know, and the energy markets, you know, how to actually generate electricity, how to put into the grid. All those things are very regulated. But healthcare is by far the most regulated industry. I think that it deals with people's lives and so we want it to be regulated. We don't want some cowboy to come along and say, oh, yeah, just try this chatbot. It'll save your lives, I think. And then they walk away with something that was wrong. So healthcare is probably the one that's lagging behind in technology development purely because we want to be safe. And sometimes we create technology that we're very excited to use, but at the same time we have to make sure it's safe. We have to validate it. And regulations move slower than technology. Everybody knows that. So we're trying to lobby the regulatory bodies to actually try and catch up with us, but at the same time we don't want to jump ahead too much because people will know us playing with our lives, really. The third challenge, scale. We all deal with it. Three examples because scale can mean many different things. In Formula One, there's only two devices. If you think about it, there's two cars in the track for a team. And those two cars have hundreds of sensors in them. And those hundreds of sensors generate thousands of streams of data. Some of those streams of data are in the kilohertz scale, so we have thousands of samples being taken at any given second for a particular metric. A good example of that is we have these cameras that take 32 points of infrared temperature sensors against each tire thousands of times per second. So you can literally draw the heat map of a wheel of a tire as it spins. And so you can predict degradation, grip and all these other things. So here we have just two devices generating huge streams of data that you need to process really quickly because if you want to make a decision as to whether to pit the car in the next lap, you have to make the decision in about 30 seconds. And so we're talking about huge amounts of data. This is something that Professor Roddy mentioned yesterday actually in his keynote. When he said aerospace challenges to make decisions to like 40 milliseconds, read lots of data, pull the cord, if the rocket is going to blow up you want to save the crew. We're in a similar situation, but at the 30 second scale. So again, lots of data to be processed very, very quickly. There's huge amounts of devices in the example I've given about the smart grid. We have millions of them trying to control what the grid does very, very quickly. So you need the infrastructure to cope with millions of connections, 24-7, sub-second response time so you can regulate these noise corrections and things. We're also a very, very interesting one to deal with. And in healthcare we have huge amounts of people. If you want to provide a health system that serves the world, I think at the minute we're about 7.6 billion of us in the planet. So if you want to provide a system that deals with them and can keep the medical records of 7.6 billion people and the interactions with doctors, the interactions with other clinicians, the tests, the insurance claims, all these things, we're talking massive scales. So that's an entirely different problem in itself. And the one that I think everybody's concerned was really good talk yesterday from Spotify about what they talk about the sensitivity of data, the privacy of people's data. And that's a good example for healthcare, for example. That's our number three there, how do we protect data so nobody gets access to very private information you might not want your employer to know about or other people. The priority information is a very good example of Formula One. Formula One is incredibly protective of their design and their systems. And so when you build a new car or a new part for the car you do not want your competition to know about it because obviously it makes your car slightly be faster than them. And so people who quit the team to go to another team actually get put into what we call gardening leave, which is you will not work for anybody else for a year. So they get paid to do nothing for a year so that the knowledge they have when they go to the competitor is already too old. So this is how paranoid they are about this proprietary information. In the grid you can think of we have platforms now that can control the whole grid of a country. Can you imagine what if a bad actor, a nation state that's trying to attack the country would break into that system. It could completely collapse the economy. It could do all sorts of damage by switching things off or even creating a resonance wave in the grid that would actually break the devices attached to it rather than just switching them off. So it could be catastrophic and so security in the energy space is huge. And the last one, the gaming system is actually quite an interesting one because I'll give you an example where recently Dr. Ling, where I work now, we just released a new version of our diagnostics engine. It's basically an AI chatbot where you can go in and if you feel bad, you feel poorly or something, you'll go in, it'll ask you a bunch of questions and it'll actually work out what's wrong with you. If you need a doctor's appointment it will actually allow you to book right there and then make an appointment. And that if you don't really need to be seen by a doctor so if you could go to the pharmacy or you could just go and stay home and take some paracetamol, something like that we'll try and do that. Some patients really want to see the doctor no matter what, so as soon as we release the new system just a small portion of the population just to test it. We included this feature which at the beginning of the chatbot it would ask you are you playing with the system or is this a real case? Just so that our stats were nicely partitioned a few people in one day that chose I'm playing with the system and they did lots of different chatbot traversals until eventually they got to one that gave them a doctor's appointment that day and then they switched to this is a real case now and they did the same traversal to get the doctor's appointment so within 24 hours of releasing the new version they already worked out a way to cheat the system and get that appointment that actually they didn't need so we now have systems in place to do anomaly detection stuff to actually rectify the situation very quickly but security is not only a concern of people coming from the outside but also people coming from the inside that have access to the system trying to cheat it now that was the preamble to what I want to talk about now which may or may not be controversial depends on how you think about artificial intelligence machine learning but I think this quote is quite quite interesting because it sort of sets the mood for what I'm going to talk about next so how do humans learn there's lots of ways of explaining the human learning process but one way of representing it is this we dream or create or build something we use it for a while we get some feedback as to whether it works or it doesn't work we find the issues that we had and then we just go around the loop and make it better a very good example that occurred to me the other day because I was looking at some old photos a while ago photos when they were still filmed we didn't have digital then the photography companies wanted people to have easy access to photography and to make these nice cute little point and shoot cameras and one of the features they had was the flash everybody uses the flash now and they wanted to encourage people to take photos in parties at night in dark conditions and so they released these cameras and everybody loved them they all took photos and obviously they were in digital so you had to wait a few days to get your lab photography developed and your film comebacks in pieces of paper and then they realized that everybody's eyes were red like demons and this is like what's happening why is this happening and so the manufacturers realized that the human pupil dilates in the dark and so when you shine a flash into it it reflects it back and it looks red and so they thought about this and thought oh I know what we'll take an initial flash and then a fraction of a second later we'll do the proper photo with the flash and so they changed the cameras the released new version people started using it and now everybody's photos were wrong because everybody was looking away after the second flash and so this sort of cycle then realized humans had to learn to actually respond to the change in technology say okay now every time I get photo taken in the dark I wait for the second flash if I don't want to look like an idiot so that sort of thing is the theme I want to run through next is how humans learn, design, build the machines tell us what's wrong with things and then humans learn from that and develop better machines but also we change our behavior we realize that we also have to change and so I can't I can't claim I came up with this concept I just thought of a name for this thing it's a machine assisted learning this is not the machine learning you used to hearing and you've heard a lot about this few days this is the process whereby humans design an algorithm and that could be a mathematical algorithm it doesn't have to be an ML model this is an actual physical algorithm then they run it in production and let's say the example is I do a model of a car and then I want to run it in a simulation and I want to see that that car looks realistic then I gather data that could be telemetry from the car and then the machine learning model this is where I sort of added this bit that I call machine assisted learning a machine learning model will then do anomaly detection, pattern detection and other things and work out what the differential between my model and the real world was and then the human learns how to actually tweak their models and become more accurate and learn things that they didn't see before because the machine is much better than us doing that sort of thing and so you refine your algorithm and you do the same loop so this is the concept I want to run through and I'll give you an example of each of those three industries just to see and I probably could think of a lot more many more examples like this but I'll give you specific ones which will make it a bit clearer so formula one one of the things that when I joined formula one it was quite interesting that because they've always done ways in a certain way they didn't quite favor technology as much as the favor engineering as you would imagine so they spend millions of dollars in a car they spend a few grand on computers right it's not quite the same as we're used to in tech and so I sort of had a look and said well we have all these data that's been created by the car at the minute there's really old slow processes to deal with it can we make that faster first so my first three weeks I just wrote some code add the data out of the car get it out in a couple of minutes do a lot of analysis and so within a couple of laps I could do something with it so then we realized well if we can our vehicle scientists which are I guess the equivalent of a data scientist but specialized in physics of a car could model mathematically the car and I'm talking to excruciating detail like the model of every spring the model of the aerodynamics of every part of the body that can be changed to change the aerodynamic performance of the car the stiffness of the springs the angle of attack of the wings all those things that the grip of the tires the type of tires the type of tarmac whether it's wet or not whether there's wind coming all those different environmental simulations mathematically and I could generate a lot of simulations so when we started we only did two or three simulations before a race by the time I included cloud and massive parallel computing HPC and stuff we were doing about race simulations in the weekend before so that we can then have this massive array of parametrized simulations with all the different configurations that we thought would benefit that particular race so say Barcelona race you would want more drag force you would want more downforce to get the car to stick in the fast corners you knew it was going to rain so you would do something with the tires that sort of stuff would run all these huge models and you get this n-dimensional very hard to visualize map these possibilities and then we had some sort of parallel coordinate system where you could highlight a few of those and it would massively massively reduce your sample of simulation runs and so we come to Friday those of you who don't know Formula One on Fridays we do what we call free practice free practice one in the morning free practice two in the afternoon that's where you put the car on the track and you do configuration changes and see what works because you know this is the real world all the simulations in the world are going to happen until you try it and so the race engineers which are very expert in this will tune the car as they think will fit that day those conditions that track and they will run the car a couple of times and then we'll identify some of the run simulation runs that will fit closest that performance and so we'll get rid of all the others and stick to those then do a few more changes and eventually we'll iterate within a few laps in free practice to the simulation that looked the closest and here's the simulation so you can see that there's a combination of machine and humans working together then you get to the simulation and you run another set of different simulations we call the strategy simulation so these are the ones that actually do gamification they simulate all the other cars they work out what happens if this car crashes here what happens if there's a yellow flag or a red flag what happens if we have to stop the race because of rain that sort of stuff and it gives you all the different possibilities because your competitor is probably going to pit you a lap later and you can actually overtake him that sort of strategic decision very quickly generated with this one simulation that we narrowed down so now you have almost a perfect simulation of your car and then you can run that model into the strategic simulation then we run the race and then we see whether it works or not and strategy as I said is very difficult to predict sometimes the conditions so most of the other times it would work in years the vehicle scientists everybody would get together look at the data which was lots of them machine learning was used to actually help work out what happened what was predicted versus the actual and it would go and tweak the models and you know we go to the next race that's a very nice sort of loop example of what machine assisted learning can do in terms of humans designing algorithms that then get modified or improved by what we learn from machines in energy this is a short example but in energy particularly in the smart grid we did a very similar thing we had data scientists and physicists modeling people's rooms how do radiators hit up your room can we tell that your window is open and therefore you're wasting energy and the room is getting cooler faster than normal can we model the grid can we model power generation what happens if I switch this nuclear power station because there's been a problem then we would come up with algorithms how to control devices out there to make those changes respond to those changes very quick and so we could enable more green energy in the grid and so then we would run this in production the devices will send some telemetry to say this is the power I'm generating this is how hot the room is all these different things and obviously get millions of those and then we use anomaly detection pattern analysis and other things to actually work out how close was our model in a way that we hadn't predicted and so we learn which of our models we then consider ok what if some other power station goes so far we have a massive gust wind in Scotland that's going to generate a massive rush of energy all those different things we'll learn, modify and do the loop again so again another example of actual human design algorithms being reinforced with machines and in healthcare this is very close to my heart because we're actually building this as we speak we have probably the only system probably can't claim that so I won't say that but we're one of the few systems that actually have a set of algorithms that can actually diagnose almost I believe that we're now up to 95% of all conditions and Dr. Ling we really believe on clinicians being in charge of that being able to design algorithms and those algorithms are done with using Bayesian logic using natural language processing using a bunch of different techniques we know about your family history we know about your lifestyle, do you smoke, do you run do you have someone in your family with diabetes all these different things so we have a picture of you and then we create algorithms that will actually tell you when you're feeling poorly and we ask you there are questions can we work out what's wrong with you and tell you how to fix it so we create the algorithms patient comes into the app starts talking to the chatbot and it'll be asked a few questions because we already know about you so we want to ask you are you diabetic if we already know you are that sort of thing so it sort of makes the chatbot a little bit more streamlined and the UX is a little better for the user so we collect all the data we work out what we think is wrong with this person at the same time we're analyzing how long did it take to traverse the questions did it hesitate in any of them could it have gone in a different direction what is the actual conclusion what do we come up with who should fix it and by what time frame and then we capture the data as that patient goes through the pipeline and that patient has to go to the doctor and we see him appearing in the doctor because we put a beacon in the surgery and we can see them showing up then we can start collecting all that data so much so that we now have huge amounts of data that can give us population level analytics so we can tell when there's a flu outbreak about to happen in Birmingham for example there's a huge condition happening in a particular area which could be something with the water so at the same time we're analyzing how did the algorithm perform did they actually predict the conditions correctly did the patient do what we suggested or did they do something else and why can we change the algorithm to make it better so that whole process then gets presented to the clinicians the insights from the machine learning models gets presented to the clinicians on their toolkit design where they design the algorithms and he says right this is the algorithm you start with the best you make to it but it's up to the clinician to then assess whether they're safe and then merge them so it's all like a development process you get your pull request you assess it, you merge it, you put it in production and then we'll go through the loop again so you can see that the process works really well in this case because we don't want to lose the human interface that would decide whether something is actually clinically safe the machine alone would not be able to make that assessment so I think this quote is really appropriate because from the CEO of IBM it basically comes to say that as valuable as artificial intelligence is actually it's augmenting us and we shouldn't forget that the human in the equation is actually very, very important and we're the ones that will improve AI in turn or being improved by AI so just a final note just say we've talked about what's in common with all of this but something that I sort of blocked about years ago was the concept of cross-pollination one thing I normally do if I change jobs is to change industries I don't like to repeat industries if I can avoid it the reason for that is because when I go to a new industry I'll learn a lot that I didn't know because I didn't need to use that methodology or that technique but at the same time I see things they're doing that maybe another industry has solved years ago in a much more efficient way and so that cross-pollination of different industries coming together and solving the same problems because as you see in the methodologies are very similar you can actually resolve issues quicker so Formula One is a great example as I said it's very hard to get to F1 unless you're getting from the bottom and if you get from the bottom then you're going to be taught what F1 does, how they do it and why they do it and the answer is because we've always done it that way which to me is not a reason, it's an excuse so when someone from the outside comes in it's an eye-opening experience for both because I saw, wow, technology solved this so much quicker, so much easier why don't we try this but also I learned about process I mean the way that they manufacture things they run them through the pipeline, they know when something is faulty and they're already ready with their replacement on track, that was amazing I thought, wow, we could use that in a hospital if we could do that with patients moving around different wards and different beds and all that sort of stuff what Formula One does is amazing in that system, so this cross-pollination I think is a fantastic concept that we should encourage and if not by changing jobs by coming to places like this and actually talking to people that are in different industries and see how they solve problems and whether that applies to something that you might have been thinking about that there's a better way to do so just to conclude passion is what makes us what we do what makes us do what we do humans are the better part of creativity from nothing just from our knowledge and experience we can build great things but machines are much better at spotting issues crunching huge amounts of data and then give us that feedback straight away and with this machine-assisted learning concept we close the loop we learn from the machine the machine learns from us and it keeps on going until eventually probably we'll all get much better off and that last concept I just talked about the cross-pollination concept I think is very, very important I encourage everybody to after this go out and network and see who works in our industries and what you can get out of them and that's it, thank you very much any questions happy to talk to anybody outside of there don't want to ask a question in public for the talk I have a question, you mentioned that you used physical mathematical models did you also use statistical models? some of these were Monte Carlo simulations there were a lot of different models put together did you use a combination as in did you use feature extraction using physical models and then statistical models on that? I can't say much but yes we did a lot of overlays of different models and we learned about how to re-order them and how to do them in the right way so there's a lot of combinations of models we don't have any more time for questions that's fine, thank you everybody