Can AI make the world a better place? – transcript

Professor Hugh Durrant-Whyte, NSW Government Chief Scientist and Engineer.

---

Professor Hugh Durrant-Whyte:  It's a great pleasure to be here today to talk about AI. I have to say … you know … I'm getting a bit long in tooth. I've been involved in the AI field for, heading towards 35 years now and in fact I think the interesting thing is it's not really a new area.

About five years ago I went to the 60th anniversary of AI that was held at MIT. So it's sort of been there for a while. And I was reading originally … you know … some background on where and how AI was created. And essentially what happened was, computer scientists got together and they said, "I know, we've going to find a new field", and what academics do is they find some good buzz words, and AI was a good buzz word at the time. And they went to the defence department and said, "Give us all this money and in five years we'll make you computers as intelligent as people", and this was in the 1950s.

So what they should tell you is that AI is much harder than people think it is. Alright. And I have to be honest, we were having a conversation earlier … we have been through a number of cycles of AI about to take over the world. I remember the one with expert systems in the 1970s. Okay. And we've been through the 1980s, 1990s with neural networks and so on. And while I do think the bar has been raised, and AI is a very important thing and it's seriously impacting everything at the moment, I'm going to talk about that in a bit, it is nevertheless on a cycle. So do not worry your jobs are not about to be taken over by AI in the near future I suspect. Okay.

So I'd like to start by just giving you some idea of what AI is, and I think we use the AI probably over much in terms of what's going on and I want to refer back to, initially at least, Patrick Winston's book. Patrick Winston was the first head of the first AI lab at MIT back in the 60s and 1970s and he wrote the very first book on AI, Introduction to AI, and I still have it and underlined in the preface, because it was the first book I'd ever read about this area was "What is AI?". And what he said is "AI is anything we currently can't do when we know how to do it it's called an algorithm". And I've got to be honest, that definition holds true today. Okay.

If you can show me a piece of AI, I can show you some data and an algorithm and I can write you down exactly what happens and how it works. Alright. So in some sense there's no … well can I say … singularity out there where there's stuff we don't understand that's "pixie dust" that suddenly transforms things that 'we don't know how to do' into things that 'we do know how to do' using a black box. Okay. AI is basically an algorithm and data.

Let me also compare that to a slightly more recent set of quotes. So Vladimir Putin, who you probably … is a world renowned expert on research and AI, has said, (and this is something we used to bandied around in Defence in the UK), "AI is the future for all humankind ... whoever is the leader in this sphere will become the ruler of the World" and I think he fancies himself in that category, as do China, as do the US, as do everybody who's investing massive amounts of money in the area. Intriguingly I always find, probably the one big country that is not investing lots of money in AI is actually Australia but let's come back to that later on.

So you take that at one level and I like the riposte from it, Andrew Ng (for those of you who don't know him, [he] is arguably one of the best known people in AI globally). He's a professor at Stanford. He was CTO of Alibaba and a number of other companies. I love his quote, he says "I worry about AI super-intelligence in the same way I worry about overpopulation on Mars". It will happen but not in my lifetime. Okay. And I think that really does summarise, I think, where AI is. It is coming, it has transformed lots of things and it is important but it's not about to take over your lives. Alright. And I think we kind of need to make sure that we understand where we are in this area.

And just to clarify that, I sometimes like to think about where we are today and these sorts of things and there's this notion of what's called "weak AI" and "strong AI". And weak AI is really about structured data, simple problems, typically problems where you've got a lot of data and a very simple question. For example, "Will this person buy a yellow dress?". Okay. Well I've got lots … I've got a million people who've looked at this advert and the number of people who buy this is this and so it's an easy problem. Okay. That we know how to solve and we think of that as AI but the reality is those are trivial algorithms, they are algorithms … there is nothing that we do in that area, in AI today, that researchers in the 50s or 60s would not recognise the algorithms. Okay. Things have changed, we have much quicker computers, we have lovely smartphones, we have other things like that and lots and lots of data that we did not have in the 50s and 60s but generally the algorithms are the same.

Then there is another form of algorithm, what I think of as strong AI which is really trying to solve what I think of as hard problems, for example, "Will this person, age 15, die of cardiovascular disease before they're 65?". And why? And what should I do to intervene? Now that is a hard problem. There's lots of different pieces of data. There are very difficult questions to answer. There are lots of moral and social and ethical issues going on and so on ... and the reality is, although we are tackling those problems today, and indeed we are going to talk quite a lot about this sort of problem, it is not a problem we can currently solve. Alright. It really is quite hard. It requires quite different thinking about data, about algorithms, about the way we put information together, about the whole way that we model humankind. Alright.

And these are important questions at the forefront of not just AI but of medicine, of all of these different areas. And I think one of the things that's very clear is you are genuinely seeing a merger of science and data science in a process that really is helping both sides. It is helping us think about algorithms and mathematics but it's also thinking about how we transform fields that are … you know … that are important to humans and to societies and to economies and so on. And I think that really is where AI is ultimately going to do good. Okay. So it is not your Facebook feeds, it's about solving medicine and other sorts of problems and I think that's one of the key things that I want to get over today.

So all of these are going to have lots of future impacts. There's this economic one - we are genuinely … the whole thing is being disrupted. There are the cyber worlds now, the way businesses work, the five biggest companies are no longer ones that make anything, they're about using data and so on. And I think there is a huge adjustment going on in that process. At the same time there … and I'll come to this towards the end of the talk ... there is this impact, and in Australia I'm responsible for a lot of it. If you go out automating things then what happens to the people that were being employed in those sorts of occupations? What's going to happen to the future of society and so on. And there's this whole thing also, which again we're all grappling with which is: you can use AI to do good but at the same time there's this cliff edge around inequality that results from the way you manage data, use data, apply it to things like automating jobs, nutrition, whole health, crime, you name it basically. There are all these sorts of issues which information and data has never really resolved in a useful way and people are thinking about it.

So the summary of that really is the following. Is AI … it's about data … it's about algorithms and it's about applications. There are lots of great applications out there and really that's what I kind of want to focus on today. I want to show you the kind of, where Australia is good in this area, where there are things perhaps you've not heard of. So I'm not going to talk about Google and Facebook and people like that, I'm going to talk about geology and medicine and all these sorts of other applications where I really think our data science and AI is really making a difference.

So, the opportunity for using information has always been out there. There is a deep seated thing where, in truth, what we're talking about in AI actually, fundamentally, is about mathematics. It's about the way that we describe systems, about the way we describe information, the way we understand how to build algorithms out of that information, the way that we understand how we're going to visualise that. And it's interestingly ... it's core ... I mean I would say in truth, if people want to get into AI and I often get asked by school kids, "Hey I want to get into AI as a career, should I go and do computer programming?". I usually say, "No, don't do that because anyone can do that", in fact more than that, I know lots of people who've already got companies automating coding AI. Alright. So it's one of the easy things to disrupt. What you should really get into is mathematics because the core through AI in the end is applied mathematics, "It's about statistics". It's interesting I saw a number last year and I think this is good, given I'm in this thing … statisticians now get paid more than lawyers. Alright. Because it is the one field that is really in demand because it underpins everything that's going on in the whole AI and algorithms area. So if you want your kids to do something useful tell them to do maths. Okay. That's where the future is because we're all automating code writing.

The other thing is really where you get to apply that. It's in this kind of, domains which really impact the physical world, the life of the world, the societal world. You know, they're generally much broader than the kinds of Facebook/Google things that perhaps you're used to. But it is genuinely about trying to transform the physical environment, that medical, that physical life environment and also the way society engages. And I think there are very big transformations happening in this area and interestingly a lot of them are actually happening in Australia for lots of different reasons. Okay. Because of the type of economy we run, because of our dependence on things like minerals and so on and so forth.

So there are lots of great applications that are out there now in lots of different ways and I'm going to try and unpack some of these things and I divided it into three categories because I think this kind of makes some sense. I think about physical systems, you know, and these are just some of the projects that I've been involved in over the years. We do things like, for example, apply AI to do mineral discovery and I'm going to say a little bit more about that in a minute. We apply AI to running power systems, so that picture down in the bottom corner down here is actually Snowy Hydro, so we use a lot of methods for example to predict weather and therefore to predict the value of a kilogram of water and energy in the future. Alright. So there are lots of interesting ways of doing that. The one in the middle there is about water processing which is becoming a critical issue in Australia and the way that you manage water through that entire supply chain. Then there's things like materials design - new materials. We can no longer just run experiments because there are just too many possibilities. But we can use machine learning to learn from past experiments what future materials might be good in a "bio-compatibility" sense or for stealth if you're in Defence. All of these sorts of things.

So AI has huge applications in lots of physical domains and I'm going to do a bit more of a detail, and we're going to a deep dive around one or two of those in a minute. But these are truly transformational. They will change the way these industries work and I'll come to the mining one in a minute. You know I've been heavily involved in the automation in the Pilbara and we've gotten to a point of where … you know … nowadays places like Rio can run 14 different mines across the globe from one room in Perth. Alright. So things are really changing in a fundamental way that genuinely affect the economy and what people do and the way they think. So that's a big part of it.

The other huge area is in life and health sciences, this is unbelievable. The science and things that can go on in this area, you know. And again I'm going to unpack some of these in a later example. But simply, you know, if I pick one trying to understand, you know, what genuinely it is through people's entire lives that will affect their … the possibility of them getting cardiovascular disease or getting cancer or, indeed these days, obesity, diabetes, etc... etc ... And a lot of these processes are very individual to people. So, you know, your problems and your problems will be completely different, and it's your lifestyle, as you go through in terms of the way you add information together about what you do, what you eat, the way you behave, your genetic background. And all these things that need to come together to produce in some sense a predictive says … you know … you should be doing this. You should be doing resistance exercise, aerobic exercise, or you should be eating that and not eating that and so on. And away from these generalities, like I'm sure you've seen in newspaper, red wine is good, coffee is bad etc., which generally don't apply to everyone, they just apply to a few people. Right. I'm sorry to say that. Okay. But equally in other areas.

So another huge area of application in Australia is agriculture. This is an area that has a lot of data. It is very poorly used. How do we genuinely breed good cattle? How do we genuinely … if they're milking cows ... eat the right piece of grass, get milk at the right time, get supplemental feed that will really actually help the outcome and so on. How do you join all that information together? Or in terms of ecology there's again a lot of data out there around water aquifers �� you know … predictive analysis of what's happening in floods and droughts and these sorts of things. How do you pull that information together and build a sustainable agriculture, a sustainable ecosystem around that. So these are huge data science problems, huge AI problems alright which really generate a substantial difference in terms of outcomes, and really will change again the way that these businesses engage.

And arguably one of the trickiest things in my book is the impact that AI and machine learning is going to have in the social science area. There is already lots of work out there saying for example, and you read it in newspapers and I get real worried when I read things in newspapers and I'll come to why in a minute, about … you know … "X causes Y". Alright. There is no such thing. Okay. One cause of Y could be X but there are a billion other things that could do it as well. Alright. And the issue is how you actually understand all that interplay. So how to use data science for example to reduce criminality. Alright. So it's not like criminality suddenly happens and it's a crime and you send that person to prison or otherwise. What happens is that person gets born, they get brought up, there are social interactions, there are points of intervention, there are whether, you know, that person has been in a particular household, in a particular region, whether they have been … you know whether they are subject to poverty and all these sorts of things that come together to really determine what the propensity of that person is to commit a crime. Okay. And how do you manage that and how do you intervene and I just picked one example there. In the social sciences area it's immensely complex, it's immensely complex, because on the contrary to what you get in Facebook and Google it is not a big data problem. Usually there is so little information about that sort of problem that you genuinely have to think hard about what those influences and outcomes might actually be, and I'm going to unpack that again a little bit more in detail.

So you can see there are lots and lots of applications of AI that you're probably less familiar with than the kind of Google and Facebook and so on and so forth. But these are the areas where I think AI will have much more of a profound impact than what you see on your iPhone. They will genuinely change economies, the way that we live and the way that we engage socially with everyone. Alright. And there's a lot that's going on in these areas and I want to, again as I said, try and unpack a little bit of that and show you a little bit of the detail as to what's going on and give you a sense I guess as to what AI really is. And I want to come back to what I said in the beginning: yes it's only data and algorithms. But I think this power of data and the power of algorithms and the increasing power of computing has indeed transformed and indeed disrupted many of these areas.

So I'm going to start actually and say a little bit about geology and minerals and mining. And the reason I'm going to talk about it, in fact all the examples that I'm going to show you here are projects that I've been involved in. And I just think that's a good thing because I actually know what I'm talking about. Okay. Good.

So here's a really great one. You know, we've been involved now for about a decade in the following problem. This is a picture of Australia and the underpinning picture is actually a picture of the surface geology of Australia and you will notice that it's got a multi-coloured out in the West and it's slightly multi-coloured out on the East Coast but in the middle it's all green. Okay. And the reason for that is that when Australia was formed that green bit was all underwater. Okay. And it was all underwater and so it got covered with what's called a regolith, so there's about 100 metres to five kilometres of covering over the top of what's going on in Australia. The other thing that's on the picture is those little dots and those little dots are where the current mines are in Australia and the astute reader will notice that all the dots are where there is something that's on the surface. Okay. And the big $60 trillion question is what is under the green stuff? Okay.

If we really could find out what was under the green stuff we would transform the entire mining and mineral business. And there is a big opportunity to do this, this is known as the "UNCOVER" project and what that basically says is, let's take all the data that's ever existed, and there turns out to be lots and lots of it, you know. And let's take that data and let's really apply AI and modern mathematics and statistics to it, and let's try and uncover what sits underneath the regolith. Okay. If we can genuinely do that in this country we would transfer the way that mining companies operate. And that's a real live project. It's really interesting. People are making lots of good progress and it genuinely is starting to change the way people do exploration and minerals discovery.

So this is a really transformational kind of problem. Going into a little bit of detail here, this is kind of a little bit of an idea of what happens, so we build AI techniques here that are not about discrete things like buying a yellow dress but actually a continuous problem. What is underneath? What kind of structure does it have? What sort of fractures, zones? Does it look like it's going to conduct copper and iron ore? Or does it not look like it's going to conduct like granite and so on. And how do we combine different types of data like seismic, gravity, magnetotellurics, all of this sort of stuff to really get a deep picture for what's going on. These are kind of critical things that really will change what goes on in the industry.

At the same time there's lots of local stuff. This is work that I've been involved in as we talked about in the beginning. A lot of AI is already in use. So in the Pilbara now there are 300 autonomous trucks, there are 150 autonomous drills, all the trains are now automated and it's not just what's in the Pilbara but the mines in Mongolia and in Utah and so on are all controlled from a single room outside Perth Airport. Okay. This is work that's been going on now for about 15 years. Okay. But what it shows you is that AI, it may still be algorithms and data but when it's applied can be truly disruptive. This is changing the way, firstly that mining companies work because now you don’t need people to drive trucks, you still need people on the mine in lots of different ways for lots of different things. But equally it's going to change the way mining practice actually happens, and it impacts safety, it impacts a whole range of different things. It is also disruptive because the people who used to earn $150,000 a year driving trucks, no longer earn $150,000 driving trucks. Okay.

So these sorts of things are coming and I often like to say to people there's this and the ports and so on, there's more autonomous vehicles working in Australia then there are anywhere else in the world. So forget Uber, Waymo and everything else, these are all operating commercially in Australia. It may not be on the road but nevertheless they're out there executing. And so again it's kind of interesting, Australia leads in a lot of these areas because of the types of industries it has.

It's much more general than that and again I'm going to repeat a bunch of pictures here. But there are lots of industries out there in Australia that are, particularly in the primary industry sector but also to do with power, water and these sorts of things that are again using AI to automate a lot of their processes, to automate things like maintenance, to automate operation, to prevent, to optimise and all that sort of thing. And again it kind of fits with the kinds of industries that work in Australia so there's dozens of companies now that are really probably on the cutting edge of what's going on in this space. So again, data and algorithms plus the computing power plus a decent application is where AI is at this point.

Going to talk a little bit about ... unpack ... about one of the applications … one or two of the applications around health and life sciences and this is a problem that I … these are areas that I … I'm not typically, I like kind of, you know, I like talking to robots and rocks as my wife likes to say rather than people. But I have to say some of the health challenges are really interesting in terms of an AI and data kind of problem. And the one that stands out, and I've actually just got quite a big grant in this area, is to try and use machine learning and AI techniques to understand human metabolism. And metabolism is a really complex thing, it's probably arguably the most deeply rooted genetic thing that exists, you know. Organisms when they were just bacteria develop a genetic code alright, in order to be able to input energy and use that energy to grow and replicate and we still inherit that, okay. And so there's this huge underpinning genetic understanding of what goes on and yet you add that to what people do and we, as I'm sure you're aware, are not the cells that we think we are because we have lots of cells in our guts, in fact we have more genetic code in our gut than we have in the rest of the … 90 per cent in fact of the genetic code in the human body is in the gut, Alright. And it interacts with all that kind of thing. And then we have like … we are no longer hunter-gatherers, we go out and have Big Macs and Coke and so all of these things kind of engage in different ways. And the challenge in here is that it is obesity and it's comorbidities are really hugely impacting every economy. You know, obesity arguably costs Australia $20 billion a year … $21 billion a year and I'm sure that's an underestimate. And the issue is how do you generally … how do you deal with diseases that are not like, you know, a rare cancer in a genetic code but how do you generally get all that data together and help people manage that kind of disease. And the same sort of thing happens in cancer, cancer is one of these things, it's not really a disease, it's something that happens in everyone and you have to manage it. Alright. And you have to bring all this sort of stuff together.

So we've been building this kind of process of bringing together things like cell biology so this is all the proteins that are generated in the cell every time it ingests a glucose and you want to understand how those proteins vary with time and we have these huge machines that basically measure that. You want to combine it with a general understanding of what is it in metabolism that really works. I have to say I'm not doing this, I work with a bunch of biologists and medicals who kind of pro-generate the data and we run experiments. I hate to say it, with you know, 10,000 mice and we feed some of them carbohydrate …. Big Macs "sorry" ... and some of them Coke and some of them … no, it's not quite like that ... but you get the idea. And we try and understand what happens to them. Do they live longer? Do they live shorter? Do they get fat? Do they get thin? And then we've got people and we observe their lifestyle data and that's a really hard piece because it's so hard to get data on people … you know … that's truthful and they wear wristbands so we know how much they exercise. And then we've got all these data sets and what we want to do is try and come up with a model for metabolism that means that I can go to you and say this is what you should be doing tomorrow and this is what you should be doing 10 years from now and so on if you really wanna manage that process through your life as an individual. Because I will say one thing that's very clear early on in metabolism is everyone is different. So when you read in the newspapers about red wine and coffee and all these sorts of things actually, no one is average. Okay. So it may be good for you but it won't be good for you, and you need to do this and you need to do that, and that's because we're all unique and individual in the way that we're brought up.

So solving these sorts of things are kind of critical and there are good progress, again, being made in this area. I'll pull out the Charles Perkins Centre at the University of Sydney. Over a 1,000 researchers working in obesity and it's data that basically is driving the progress in this area, how you use this data, how you use it to really make clinical outcomes. So I just pick one area of health where data will make a huge difference in terms of understanding it and AI is the difference between good health in the future, not just for individuals but for populations and bad health.

There is also, as I'm sure you're aware, a lot going on in this whole well-being space. So it's been recently established this Digital Health CRC which I've been involved in, which is looking at this thing. But this whole process about how you really understand the dynamics of, you know, when people … there are so many things that are going on in health, how do you manage people who, you know, present with a particular disease versus how do you manage the fact that there is a population issue around, you know, people having poor lifestyles in particular regions versus the issues of pollution or, you know, other elements that might be in contaminated ground water and how do you pull that information together and generally manage costs in the health system. Alright. Because if we're not careful health is basically going to be the entire state budget in the next 20 years. Alright. So, again data is a kind of common thing that kind of makes a difference in terms of what you want to do.

And then this one, I love this one, this is actually a project that my wife did, not me, she's a statistician. And we've got this farm and we have these robot milkers. Has anyone seen these robot milkers? They're great. So what happens is the cows decide when they want to get milked and they come along to the gate and they push the gate and they have little tags on them so they can't come in too frequently. Alright. And then they come in and then the robot recognises the udders and puts the things on automatically and if they get a good amount of milk then we feed them some extra stock. Alright. And then they go out and we know which area of the grass they’re grazing on and we also know what the grass is going to look like two days from now because we've got the weather and so on. Alright.

So now you have like a completely automated dairy farm and the challenge to the data is how do you use that to make … what it turns out actually is what you really need to do is to make the cows happy. If they're happy it turns out they produce lots of milk. Okay. And so we've done this analysis that actually figures out how to … what yield you can actually get in terms of … from the cows and what you have to do to the cows in terms of the areas they feed, the way you behave with them, how much extra yield you give them but actually on average yield is about 4 to 16 per cent extra milk yield from the same cows. Alright.

So I just want you to think about how you apply that across the rest of agriculture, alright. Bringing all that kind of data together, using those things … using those modelling techniques, data and algorithms and everything else to make that sort of thing happen. So there is a transformation that's possible in agriculture using these sorts of things, so again it's not just health and the great thing about cows is that, you know, there's nothing to do with, you know, personal data guarantees, they don’t need to sign anything if you want to use the data so that's what I like about them.

Now we come, in my view, to probably one of the trickiest areas which is around social sciences. Again, you know, these are areas we often get asked about and what we realise very early on is these are ... the current … let me try and be careful here ... and I'll go to this. So, we've been involved for a while in this problem that talks about disengaged youth, so-called not in employment or in education. Okay. And it turns out that, you know, once someone has been in that category for a while not only are they almost never able to get out of it, so if they've been in that for like two years then it's almost impossibly hard to get someone out of that and also they begin to suffer from lots of other things like mental health conditions, they revert to crime, a whole range of things.

And so the issue is how do you stop people getting into that position in the first place? How do you intervene early enough on, make a prediction about someone that says you're not going to get in that position, if you do get in that position, how do I intervene to make something happen? And it turns out this is much harder than anyone thinks. So … and I'm trying to give you an example here, sorry about the mathematics. So it turns out this is a very hard problem. If you go to the American bible on mental health and disengaged youth there turn out to be 1,300 factors which might impact what's going on. So if you think about it mathematically that actually means a model space that has the dimension 2 to the power of 1,300, so that is a model space that has 10 to the power of 80 dimensions. Alright. So this is ... unbelievably … has more dimensions than there are atoms in the universe.

And then you go out and you run a clinical sample for 1,000 people. Okay. And you try and sample that space, and the answer is you have no idea what you're doing. Okay. Because there are so many possibilities you could never sample every possibility. You see what I mean? And so as a consequence what happens is anything you think should happen is provable with the data you've got. Alright. And this comes back to this sort of irreproducible results that tend to happen in medicine. I take a sample, it's a huge space and my samples predict exactly what I said they'd predict. Okay. Therefore fund me or do this intervention or whatever. Okay. And what we're … and I guess there's a little cartoon here that says this is what happens right. You've got this bunch of doctors with their funding and they say, "It's this, it's that", and so on because I've got this data, I've got a hypothesis and in a very large model space I can prove anything I want. Okay.

Whereas the reality is, what you really have to do is to understand if I have that data what are all the possible explanations for that data. Alright. So you think about the problem another way round. So if you think about someone who's disengaged then you would … you know … discover for example that half of the people that I have surveyed don’t have a good family support alright. Yes, that's a factor but it may be other factors that are part of that as well, you see what I mean? And so there are explanations as to why that data might be like that because in fact what you really need to say is "no, it's not just the fact that half of these people don’t have family support" but if you look at the counter example all these other people don’t have family support and there's no problem with them, what happened? Do you see what I mean?

So the mathematics of this problem, it turns out, is actually very complex, right. It's not just a problem of, "Can I buy a yellow dress?". A million people say "yes", half a million people say "no". Do you see what I mean? It's a very complex problem and so in any of these social areas there's a real challenge making people … you know … trying to, if you like, get across the idea that humans are complicated things. Okay. Many things got to happen and when we talk about people we've got to understand every possible data set, and how every possible data set impacts every possible outcome. Alright. And on the basis of that perhaps come back and say these are the things that are most likely to impact it. So it turns out, much to the horror of the people we were working with in social science that actually … you know … there are all these interventions they could do like: get this doctor to do this, get this nurse to do that … you know … family support, fund this, do that. Turns out the most likely thing that will stop youth engaging is to engage them in team sport because then they engage with people. And when they engage with people they don’t tend to do things like that. Cheap. Simple. It works. Alright.

So in some sense rethinking the data, rethinking the models really gives you a very different insight into the way these things happen. So we've also done this kind of thing which I think is also very interesting, is the whole issue about using data analytics to protect criminal outcomes. This is a very contentious area. Okay. And I'll give you an example. The police, I won't name the particular police force, were using some analytics that they'd got from … that some company had got … you know … using big data techniques etc. And what they realised is, yeah, crime happens a lot here so we'll go to this area and we'll deal with the crime. Of course what happens is they go to the area. Okay. And because they're in that area they detect more crime. And because they detect more crime, they put more police into the area and because they put more police in they detect more crime. Do you see what I mean? And all what happens is they confirm the bias they originally started with. Alright.

And so, again if you like, this is a much more complex problem than that. It's a complex problem because not only do you need to see what's going on in the rest of the space but you genuinely need to understand people's paths through that space. Why is that person committing a crime? Is there an opportunity? Is there a region? Is there an area? What is it that actually makes … prevent ... if you're doing policing incorrectly, how do you intervene a priori to stop that sort of process happening? Do you see what I mean? And that is a slightly more subtle way of going about these sorts of things.

So there is this notion, you know, and I'll come to it a little bit later as well, about bias and all the rest of it. But bias is usually just stupid mathematics. Alright. It is not an ethics problem. It is because you don’t understand the statistics. Okay.

So let's talk a little bit about impacts and disruption and I will wind myself out together here. So as you probably realised, I mean, I'm not a big data fan. Okay. I use it very advisably in that sense. I think big data's good. It's nice that we've got lots of information and the two pictures there, and certainly this picture over there is the readership of The Australian and what their political beliefs are. Alright. So this gives you an idea of what sort of nice things you can do. Over here is actually someone's social network, alright, and this is the company that I helped found and they use their social networks to predict what videos you're going to watch, and then pre-loads them, and offers them to you. Alright. We made a lot of money out of that. You can make a lot of money out of big data, okay, but in some sense what I'm saying to you is, I'm not sure that that's where AI is going in the future. I think we've kind of mined that everyone's off doing that and so on and so forth.

There are huge issues as we all know around privacy, bias, inequality, disruption the whole piece and I don’t disagree with all that. So, you know, I think one of the big things we are seeing and I'm going to come to it in a minute, is that many jobs are being disrupted by the fact that we can now automate them. And, we automate them using algorithms and because we have data and I'm going to go a little bit more into that in a minute.

There is an economic thing as we've been seeing so, companies and genuinely get algorithms and apply algorithms appropriately can generate good commercial outcomes whether it's an industry, in health or in social media and I think that's important. I will say I worry that one of the challenges here is that we've gone gung-ho into this without really understanding some of the impacts so, as I said, bias is something that keeps on coming up particularly in a university environment. The philosophers say, "What about the ethics of AI?".  Etcetera … Etcetera .. and actually the problem with the ethics of AI is generally, as I like to say them, algorithms are not ethical or unethical. And people are not ethical or unethical in this sense. Data is ethical, that's a different question. Alright. So if I use your data or I use a biased sample set, I will end up with a different answer. But, algorithms per se, if they're implemented correctly at least are not biased and they're not, you know unequal and so on. But, it does require that you do your statistics properly because of you don’t do your statistics properly then you'll end up with the wrong answer but hey we knew that a hundred years ago. Okay.

So I think there are sort of put points in there. I think, the privacy issue is interesting … again I think, sometimes these things get configurated. Our privacy is not about algorithms, privacy is about data. Okay. And, I think, we need to make sure we separate out the two things and there are well established techniques for dealing with data privacy. People may not use them, that's a different matter. Okay. But, we may not have the right legal process for ensuring that we use the right thing but, there are in principle processes for managing data. But, it is not about algorithms. Okay. And again I want to make sure people understand the distinction between these ... these sorts of things.

And I'll talk a little bit about this disruption because it is very topical. Actually I wrote the first paper on this back in 2014 for CEDA. And, what's intriguing is, I even went for a thing this week that McKinsey's ran at UTS. I don’t know whether anyone else went to that. And they're still using the same data set I generated. So, I'm not sure there are, you know, we've learnt a huge amount in the time. But what was interesting about this was the following, what we did is we actually went out and we did an analysis of all the ANZSIC codes essentially around job categories. Okay. And, we looked at them and for a handful of those we said, for example, here is a surgeon. Is what the surgeon does going to be automated in the next 10 or 20 years and we give a probability to that. And, also we say what does a surgeon do? Well he's dexterous, there is thinking and so on. So there are the qualities that are likely to be automated, you see what I mean? And, then we took those and we basically learnt, using machine learning algorithm, what happens in all the other job categories.

So, I think, there's 710 different job categories in the ANZSIC code. And, what you end up with is a picture of what kinds of jobs are going to be automated in the next 10 or 20 years and what are not. And there are a couple of things worth saying. The first thing is if you do the raw numbers, then 40% of jobs that currently exist are going to disappear within 20 years. And, I'm amazed, I keep on telling people but that was not the point because if you go back 20 years, undoubtedly 20 years ago 40% of the jobs that existed 20 years ago have already gone. So nothing has changed. Okay. We are changing jobs at a reasonably fast rate and we always have done. Okay. Even though that always tends to be the headline, but the real headline is actually the bottom picture which people seem to ignore which is the following: there are very different types of jobs that have been automated by AI are then going to be automated by things in the past.

So in the past we would automate blue collar jobs. We would automate manufacturing. We would do things like that. But this type of job that is being automated by AI is very different. Its jobs if you like that are in the middle. So as I was just talking to someone earlier, the person serving coffee out there is not going to be automated in the near future because we have no idea how to build a robot to make coffee. It is incredibly hard. Okay. Or waitressing, or anything that involves social engagement, healthcare, you know, age care, that sort of thing, these people are not going to be automated. The other people who are not going to be automated are the people, frankly like me, who design robots or do other kinds of things. But there's a large quantity of people in the middle who do analytic things who basically manage data, analyse things, produce reports. So you know, proof readers a lot in the legal profession that's being automated, there's I'm sure many of you know. But equally – doctors, they're going to need fewer doctors in principle because a lot of the diagnosis is now being automated. The kinds of things that used to happen in banks, we're used to that, they've gone to teller machines, they've gone to this and so on. So what's intriguing here is that many of the jobs that actually we produce graduates for out of universities, they're the ones that are going.

And if you look at the picture what's happening is, the big take home is that this is becoming very polarised, so we're getting people at one end who do one thing and people at the very other end who serve coffee and then nobody in the middle. So the big shift is not that we're losing jobs but the fact that the jobs that we are doing are becoming very polarised, alright, And that is a huge challenge, because we genuinely need to worry about what's happening in society and Australia's behind the curve on this in many respects. If you look at the data out of Europe 50% of graduates are unemployed. Okay. If you look at what's going on in the US, it's not … in fact it's not the blue collar workers who are suffering, it's the people who paid out all that money to get degrees who are suffering. Okay. And there are lots of weird things happening in that space, so there was a recent article in The [Australian] Financial Review, only 40% of people with degrees in science are getting jobs in science. Okay. So a lot of the analysis that would appear in that area is going.

So the big challenge with AI in automation in a societal sense is to understand what that polarisation is doing and understand how we're going to fix that, and I'm not the only person by any means saying this thing now. I think the other big worry for Australia is that this whole polarisation process is very localised. Okay. So this is a scale of where we think the impacts are likely to be, based on where those jobs are and the different codes in the different local districts. Okay. So you will see the places automating in the Pilbara are disappearing in terms of jobs. In fact the really worrying thing is the tendency, not in cities, is to lose jobs. And you lose jobs just because a lot of what's out there: agriculture, mining and so on, is all being automated. Okay.

Even if you look in Sydney: what's happening is the CBD because that's where people come when they want to do creative things, the jobs there are going to grow, but the jobs out in Western Sydney are disappearing - manufacturing, automation, food processing - all of these things. Okay. And so this … the other serious thing I think in Australia is this geographic polarisation of jobs, that again is going to cause some serious social disruption. And it's not just a case of politicians offering some particular thing out in the west or out in the sticks. We genuinely need to start creating different kinds of industries out there that will support different kinds of jobs given where they are in the AI and data society.

So let me try and finish up.

So I think there are a couple of things I wanted to make sure everyone got as a take home from this. The first is strong dystopian AI - frankly is not going to happen in my lifetime. Alright. So if I don't think it's going to happen, you guys shouldn't worry about it. But I've been working on autonomous vehicles for most of my career, just so everyone knows, and I don't think there will be driverless taxis in Sydney without a steering wheel in my lifetime. Alright. It's just too hard. Okay. I can tell you that for starters, so, no matter what everyone's telling you.

The other thing though is that weak AI is clearly disrupting stuff and we genuinely have to worry about it. It's disrupting business, economics, society, applications and there are predictable analytic jobs that are going and there are other areas that we need to start growing I think that are different. And I think although we're going to get increased productivity, and I think it's working for Australia this whole automation data area, we are generally going to have to worry about inequality and how that works. It's a really pressing problem in my view and it's already happening.

And I think the other thing that worries me a lot about … Australia's increasingly … is focussing on urbanisation and we're already a very highly urbanised society and I just feel that Australia is going to get depleted in its regions if it doesn't fix what's going on.

And then just to finish … this is one of my favourite cartoons in the area just to show you that … I mean … this isn't going to happen but I like Dilbert for this: I compiled a D&A test kit with big data to predict a person's future health issues, and that depressing knowledge caused every member of the test group to make risky lifestyle choices and now half of them are dead.

At the risk of bragging that's exactly what I predicted.

There you go. Thank you very much.

[Laughing and clapping].

---

Tim Brookes, Partner, Ashurst: We've got time for some questions from the audience if anyone would like to ask Professor Durrant-Whyte. … If you can just say your name and a mic will come your way.

Michael Coonan, SBS: My name is Michael Coonan and I'm from SBS. And I think you said if you'd worked on social science projects which had proactively addressed that bias concern, which I know you mentioned a lot of people have, but I take the point that the algorithm doesn't have bias on the other or the data doesn't. But when you were talking about the fact that you can seek out answers and see your name when you've got huge data sets, surely it is some policing in the design of the assumptions that go into the algorithm.

Professor Durrant-Whyte: There's a big argument going on at the moment in communities. Right. Because medicos will go off, you know, they've got a hypothesis in effect and then they gather data, not necessarily to prove it a priori, but there's a kind of intrinsic bias in the way that you take data to prove a hypothesis. Whereas, if you were a proper statistician what you would do is: you would try and disprove the hypothesis. Do you see what I mean? And then, if you try and prove it with this then you end up with so-called "P values" which is the way that, you know, roll off statistics you would test the thing. You test whether the data matches the model and the answer is, "Yes, it matches the model". But I can produce a billion other models and it would also match. Alright. So the challenge … and this comes back to the big argument in the field which is a lot of medical researchers ended up in the so-called area of non-reproducible results. People produce a result no one else can replicate, and it's not a problem with the experiment. It's a problem with really understanding what, you know. I was trying to say here is, many of the problems in the social sciences involve enormous complexity. And we generally … these are what we actually talk about now, is a small data problem. The amount of data we have is so much less than the size of the model that we genuinely cannot go from that to infer anything about the model. You know. In an absolute sense. And so we are actually kidding ourselves about what's going on. So it requires a different approach. It requires, I think … you know … let's be clear here … it requires people in AI to stop hiding behind … you know … the big data analogy and actually trying to get to grips with people who are experts in the domain. You know. And work with them closely to genuinely understand what new mathematics is actually required to genuinely give us the right answers or the right guidance about it.

One other thing, because in my role as a policy person in government right … in my role … the other thing is to really articulate to people the uncertainty that exists when you do that so. You know. Nothing is certain. So if I predict something I should genuinely try and explain to the stakeholders the ambiguity and uncertainty that comes from that because if we don't we're not being honest.

Tim Brookes: We have one question here.

Glen Frost, FinTech Summit: Hi I'm Glen Frost from the FinTech Summit. You mentioned that Australia was a bit behind the rest of the world in some areas on AI and innovation. Given your role as Chief Scientist, what's you key message to government both as state and a federal level to, how things can be improved or more sort of policy changes? What's your comment to them on that?

Professor Durrant-Whyte: Well, I note first, that we're in caretaker mode. Okay. So I'm making any comments - I make as a private person in this talk. Okay. My view is, and I hope I show that in the talk, is that we have … you know … you should focus on where you have your comparative advantages. Right. We do not have a comparative advantage in the manufacture of smartphones. We do have comparative advantages in say FinTech and indeed if you go to York Street, we have several thousand people there in FinTech start-ups who are genuinely doing interesting applied AI. Right. But equally we have advantages in mining and mineralogy. We have the best most technical mining companies in this country … in the world, sorry, and this is an area that we should genuinely focus on. We are generally good in some areas of health, we're genuinely good in. And that's where I think the effort should be. We shouldn't try and replicate what everyone else is doing, you know. And I think that's genuinely what's done. But it happens naturally, people gravitate to where there are strengths so - yep.

Garry Taylor, Croesus Project Services: Hi Gary Taylor. One of hats I'm currently wearing is I'm on two committees: one to establish an Australian standard and one to establish an ISO standard in AI ethics and the law. And you made a statement that concerned me. You said that bias is not an ethical problem, it's a statistical problem. Could you expand on that

Professor Durrant-Whyte: Yeah. So I guess we tend to have these arguments with philosophers and other people in the area. Right. And they'll come up with an example so the usual one is … you know … there's an ethical bias is … they say there's an ethical bias in this algorithm. First thing is there is no bias in the algorithm. Okay. It's in the data. Okay. It is a very different prospect. Right. If you go out and you make multiple samples which do not actually span the space as in its poor sampling practice. Alright. Then you will end up with the wrong answer. Do you see what I mean?

So if I go out there … here's a very good example we are faced with actually … we go out and sample for our crime rates and it turns out that we go and sample, you know, in Redfern and there is an ethical bias even before we start. And the answer we get about crime rates is ethically biased because we've just taken the wrong sample. Do you see what I mean? And my challenge is, people say that's an ethical problem, it's not. If they knew something about how to get better … how to get the right samples … it would not arise and I think that's important.

Garry Taylor: But the data that's collected is analysed by algorithms and algorithms are written by human beings who have seminal biases.

Professor Durrant-Whyte: Yes but, let's be clear here … you know … in fact I would say every algorithm that I can think of that someone would use - clustering, regression, you name it - is statistically correct. That is, if it's given the right data, it will give you the right answer. If you give the wrong data to that algorithm though … okay … if you give a biased sample and things like that, it will give you the wrong answer. It's not … I don't think you should imbue it with anything more than it isn't. I mean I can talk to you a little bit offline if you like, but there's a real misunderstanding I think in the ethics community about what algorithms do and don't do.

Garry Taylor: So algorithms cannot be misconstrued?

Professor Durrant Whyte: No, I didn't say cannot be. I can clearly make a programming error and make it do something, but if I genuinely write down a standard algorithm with standard statistical properties what you get is known. Okay.

Garry Taylor: Thank you.

Tim Brookes: I think we've got time for one more question.

Carol Austin, HSBC Bank Australia: Carol Austin HSBC. Can you talk a little bit about the implications for the structure of industry. So if you take the dairy industry for instance, the small family farm can't do the research that you described. It's more likely to be done by a large corporate entity. When you look at what Rio's doing, you're talking about one of the biggest mining companies in the world. So what does it say about the capacity of small business to operate in an environment where AI is really transforming the productivity of the industry. And if it does result in a dominance of big business, experience suggests that innovation is coming from the FinTechs, it's coming from the small players. So how do we maintain innovation if we are going to have a concentration?

Professor Durrant-Whyte: Yeah, that's a very interesting question. I mean … I better not overstep the line here. I will say that … you know … it's an interesting area because most large companies really fail to do AI well. Alright. And I have a definition of a failed AI project, "It's anything that run by the CIO". Right. 

[Laughing].

Professor Durrant-Whyte: I hope there aren't too many CIO's in the audience. Right. And genuinely with a lot of the companies that I've worked with, I said to them, "Don't try and do this internally". Build a platform. Build an ecosystem in which you can engage externally with all these companies and really try and be different about the way you approach it so … you know … mining companies are good at spending 10 years figuring out whether they're going to spend $4 billion right. They're not very good at figuring out … you know … here is a thousand companies and I'm going to spend a few hundred thousand, and I want a project in eight weeks, how to do that? Do you see what I mean? And I think that is a real disruptor.

And I think …you know … I think we're at the beginning of very different business models in that space so … you know … I could see a thing that basically says agriculture is a service. Right. So I'm going to provide a system that manages your cows in the same way that you're beginning to see in other areas, so mining as a service. You know. I have an ore body, please mine it for me etc.. So I think they'll be very disruptive around the way that that works and I can foresee that happening in the not-too-distant future as in decades rather than in my lifetime. Maybe decade … that's the same thing these days … I don't know [laughs]. But do you know what I mean? I think there are fundamental disruptors going on in those areas. I think the other real interesting one to watch will be health 'cause I wonder whether the monolithic way that we do health is sustainable.

[Applause].

Professor Durrant-Whyte:  Thank you.