And k.q. We i.f.n. North Highlands Sacramento the time now is 8 pm. Welcome to world affairs the weekly broadcast of the World Affairs Council I'm Jane Wales c.e.o. In this week's episode we'll hear from certain Nigel Shadbolt about artificial intelligence and open data he's Professor of Computer Science at Oxford University in England this event was recorded in San Francisco before our world affairs audience and that broadcast is made possible by the generous supporter Chevron and now to our moderator Quentin Hardy he's head of editorial at Google Cloud. Thank you very much good evening and thank you for joining us tonight I'm Quentin Hardy and it's my great pleasure to introduce tonight's guest says Nigel Shadbolt is professor of computer science at the University of Oxford and a principal of Jesus College he is also chairman and co-founder of the Open Data Institute which will be speaking about tonight in 2009 Serge Nigel was appointed information advisor to the u.k. Government helping transform public access to government information including the widely acclaimed data dot gov dot u.k. Site please join me in welcoming so Nigel Shadbolt for a conversation on open data and artificial intelligence I think the best way to begin is to just try and set a couple of definitions what is artificial intelligence let's start with a really easy question there. Is a hard question because when we think about our own notions of what it is to be intelligent psychologists and flosses a struggle over this question for generations and it's also the case I think when we think about automation intelligence what are we trying to build and so for some people it's kind of the art of building computers like the ones in the movies you know the kind of self-aware mad bad and dangerous to no computers in other cases it's about building a machine to. Build and behave very adaptively to learn from the information that's go again to be flexible one of the hallmarks of intelligence is a kind of a religion question in the self and as we explore aspects of that whether it's adaptability whether it's self-improvement whether it's actually being able to explain the reasons for a conclusion we get different facets of that question I think about what is intelligence we might know what the artificial is the intelligence piece is a bit more challenging it seems to me too often that people in the industry are victims of their own language using terms like intelligence and brain and mind. Most of the time from what I've seen it's a pattern finding its pattern finding in pattern matching which is something intelligence does but it may or may not be anything like human intelligence what is artificial intelligence not. Well you'd be thinking if you were to listen to some of the news and marketing that it's going to be everything and is going to solve all our problems and possibly even eliminate Assad although he's telling us yeah it's kind of like. Very very substantially over egged or hyped at the moment I think they're releasing reasons why we're at a transition point in some ways but it's not for the reasons people think it's not about the machines or about to replace or about to wake up and become self-aware it's much more about understanding what they can and can't do as you say and the narrow nurse of those capabilities the striking thing about an ai program is when you open it up and look in the code base what's not that in terms of what is there when we think about our own social cognition our own emotional states what makes us human and capable but there's no doubt that we're going to build classes of system that have this fantastic pattern finding capability this adaptability this ability to feel responsive that will occasionally spook us out it'll feel a bit uncanny and I think that's another characteristic but it's often smart but not smart like us right. Oh you got to say more about that right away I said Right and then I realize. It's not original to me that was a very I mean and professor of it mit called Patrick Winston and he had this phrase you know there's lots of ways of being small that isn't small like us when he was asked the question about what is a I and he was saying imagine a space of possible intelligent systems we occupy one part of it but there are lots of ways of being flexible and adaptable and I could build you up. Game playing genius programs that we just have alpha go deep mines Alpha go to beat the best players in one of the most difficult games we thought in terms of pattern recognition. But it's not having a discussion in philosophy and it doesn't write good poetry you know this is the whole issue around this is what are the extent and limits of the systems that we build and how narrow is that compass and maybe what we're doing in ai really is building lots of narrow task achieving capabilities we when we put them together seem interesting but individually they still don't amount to a general intelligence Peter Drucker said a mouse is good at being a man a mouse and an elephant is going to be enough and you don't expect a mouse to be an elephant you know and similarly you expect these things to solve patterns but now and increasingly spooky ways things have happened in the past 15 years say that makes for this revival of artificial intelligence and you've been in the field since 1908 so what happened now and it must been quite exciting for you to see it take Yeah it's been very exciting I mean to the point you made earlier though it's interesting that when we do get these triumphs in ai And if you remember about some of you remember back to 996 when I.B.M.'s Deep Blue beat Gary Kasparov you know that was a high watermark we thought at that point what's happening in ai are they going to wake up and take over remove all our jobs and they. Didn't But at that point Kasparov was also rather unnerved by the program you know and in some sense it was easy to endow the program with more with different capabilities than it actually had what was doing was searching hundreds of millions of positions a 2nd you know it wasn't playing chess like a human does but it was able to manifest extraordinary performance in that area and likewise when it wanted jeopardy the real advantage was it didn't hesitate going to the buzzer know when it had an answer it immediately and when you saw that it kind of felt spooky you know it's in reasonable speed and accuracy of response and i.b.m. a Trained a great games player what was different about Watson I.B.M.'s Watson and what's different now is the availability of huge amounts of data and I think that's what's making a difference it's simply the case that we wouldn't have imagined. 30 years ago when I was 1st in the field getting into the field and we're looking at computation linguistics How do you get machines to understand language we reach for grammars we reach for very elegant ways of describing how language structure worked because we didn't have databases full of billions of examples of sentences and pages we have that now we can build models that translate language and understand the implications of answer questions in a way that is entirely different than the methods we had to use 30 years when the data was much sparser So I think data data everywhere is the thing that's making this latest phase of Ai very powerful what was extraordinary was the browser turned out to be a kind of sensor and a sensor for all sorts of things. Obviously the content of web pages in different languages but also links time people spent looking at things where they went next what they related to a sensor of desires in a sense and now we've blown that on the cell phones and all sorts of other things that model is in the physical world so there's 3 dimensions to it yeah well that's right and the genius just getting richer the genius of the Google result the original kind of search engine result was to realize that those hyperlinks all those links on all those pages that humans were making represented an expression of interest and relevance and at the same time computing became so much cheaper so that you and your colleagues could suddenly do things revive old algorithms and write new ones it is the case that since I kind of completed my Ph d. And I and to about now there's been well over a 1000000 fold increase in the power of the machines 6 orders of magnitude you know we talk about exponents in the world but actually very few areas and I would say that computing is one of them these rates of change are so remarkable that methods brute force methods with a little insight take you a very long way yes now people think. Artificial intelligence is something on the horizon and it may look like Arnold Schwarzenegger when he gets here with some malevolence but in fact it's really quite in the world right now isn't it I mean how many areas do you see Ai and today well if you open up your smartphone you know you've got about 6 or 7 right there you know if they're running essentially products from Ai labs whether it's speech recognition or face recognition or the question answering that's coming back on the interrogation of the of the large amount of data on the web to answer your questions or it's in your engine management in your cause or it's actually in the recommend the systems that tell you what you might like to be interested in buying or purchasing so but of course that's present a very narrow kind of capabilities in a sale that's not a I Where's the Terminator Where's the kind of self-aware x. Market when will it you know when I was lead to me and I'm saying I and some people have a strong view that that's going to be not too far in the future I think it's going to be a very long time in the future is a matter of fact and that notion of general intelligence General artificial intelligence is very tough we have very little understanding of what would constitute such a thing but in the meantime we're busy building these extraordinarily narrow autistic capable task achieving system yeah a computer that plays chest is quite an achievement a computer that feels like playing chess is an almost unimaginable achieved so that's the game we're not really there to the gap but that said where do you think it will go what is the you can probably see a path out at least a few years what will come next well I think we just again the exponents of the field it's a perfect storm in a way it's increasing power of our computers it's a spin ensure rates of increase of the data it's device engineering I often say that as an ai practitioner as a computer scientist. Most of my progress I owed to electrical engineers who have kept on kind of putting doubling the power of the computers and doubling the capacity of the memory storage on pretty well a 15 month basis and that has allowed us to build huge farms of machines that can hold the Tallahassee of this information and gather more and more and I just see that the consequence of that will be will be a kind of a ubiquitous distributed set of capability from everything from automated systems that advise you on your recipes to help drive your car through to translating languages from one to another but these isolated capabilities will just move into the background in the same way that we don't think of all phones as carrying a are around on them or the products because they're not self in the not this idea that we have of what an ai could be or should be doesn't mean the technology isn't doing just a host of things for us and that it's going to I mean I would say if you want to take an area pick toys pick leisure pick the fact that Spielberg's ai is an interesting film the most plausible thing in it was the teddy bear Ok and that teddy bear which was. Capable of recording and responding in an effective way to its owner but in fact you know it's not going to be self-aware but the teddy bears of our children's and our grandchildren's future will be effectively life logging systems they'll be companions that the child will take. Effect relationships but the teddy bear won't be looking back out having an affectionate relationship with the child the child may be the other way around but they will become objects which embody large amounts of our capability and I can't think of an area that won't have these Ai capabilities embedded in them somewhere is going to change human consciousness. A child growing up with something like that will have expectations of what from the world well I think our tools and our technology of always changed as I think the remarkable fact of our current state is that you know we've been making tools for 4000000 years probably I mean those very earliest hand axes those very earliest ones they began to shape our cortex as well as our motor responses and they gave us mastery over the environment and I think that has always happened. To mentor us they extend us and in just the same way we have more information at our fingertips than Leonardo De Dora or Crick or Darwin or any of these individuals as the saying goes we make tools to shape the world and in turn they shape us a person they do indeed and I think that that how we will develop those notions of companion billeted and what's sociability will mean the things that matter to us now we have an infinite capacity as a species to to answer for more files into objects I remember my sister used to drag around on a piece of string a brick that was her imaginary pet friend and he had a name for it and rest of it and it was a brick you know but it wasn't for her and so I think this that we will project into these systems extraordinary amounts of affectation and they will become important parts of our environment text most pernicious proposition is will create a future so perfect that we won't have to be good. And we're still going to have to be good well it'll be a jewel use technology right I mean so that's one of the challenges we will face with all of this and it's always been the case if you think of artificial intelligence and computing while Think of all the other great technologies chemistry chemical warfare biology biological warfare you clear your plate Wolf you know what we do with our technology is chiefly important like it's not the machines waking up and deciding to do away with us that I worry about it's the fact that. Our artificial intelligence plus on Natural Stupidity these things in mastery over us and we'll give them to make decisions that are genuinely worrying and if you know we can make mistakes in scale like never before so the technology will just do the most amazing things but the thing we will have to co-develop and really do it fast and hard at the ethics and the regulator restructures around the deployment of the systems let me just spend a minute crawling up the chip a bit things like deep neural networks deep learning transfer learning these more advanced systems What do you make of them were they real breakthroughs in this process. They're very interesting illustrations of a couple of things I would just say the ai in its history has at different times taken very different ways and developed very different methods to try and produce adaptive and flexible behavior in the good old days as we rule based systems logic based systems. Rules and approaches that are still out there running systems today then we realized that maybe out to did introduce uncertainty into the system so we develop the footy logic and Basie and reasoning and all sorts of families of methods to try and develop more of a toolkit of ways to build our systems were really into systems now that are almost incapable of explaining themselves well that's a good point so I think one of the one the instant things about earlier generations of Ai and we need to build systems called planning systems planning systems would work out how to construct your house or how to supply the do the logistics planning to get all your material from one place to another we still use them at scale all the time routinely but right from the earliest days in ai there was a different game in town which to say let's look at how brains of Bill let's look at building mural networks Let's see what we can compute if we have networks that change in a doubt that structure on their weights in a way that responds to inputs and gives us interesting outputs that method your own networks that they're around for a very long time and we explored them very thoroughly in the ninety's 1990 s. And again in the 2 thousands they were doing interesting very interesting things but what was happening in the meantime is that people were thinking well what happens if we could build. Arbitrary deep neural networks layer on layer Well that would take a lot of computing it take a lot of power hey we've got that you know and at some point the ability to take these your own networks a load them on architecture that could run them reasonably fast and then to be fair do a set of technical developments which recognise that if you've got these big network inputs you can do a certain kind of computation on one sub area of the network and then feed that through to another layer of the layer so you've got lots of complicated mass being performed to look for the features and coalition in very complicated inputs at a very low level you're doing elementary pattern finding and then more refined pattern finding but then you're doing patterns on the pattern finding and they were thinking 3 layers now people are what $78.00 many layers the Only in any facts and indeed the that is one of the things that is really shifting the game the point about those layers when I when I open up the box and I look inside the kind of program I often will see this brilliantly adapted connection graph with lots and lots of weights and it's doing stuff how do I get it to explain why it's doing what's right it's not even clear that if you took the machine fairly well into a game of Go or a game of chess and ran the patterns that and it could expose itself afterward that it followed the same tree up both times and how many Now if you really count ovaries you know how would you counter it so this this area is now being called algorithmic accountability or the demand for explanation or transparent machine learning is a big hot area of research so as you get this new set of capabilities you suddenly say well you know what if I'm going to use this method to run a call or make decisions on insurance classification I may end up in court one day I'm being asked how do I exactly where it's going with this x. If this outcome they can't explain themselves how do you regulate them. Cept they're incredibly powerful in that state so that will become I'll live with that I'll become an you. Is it is an air of such an act fair research right now let's change gears for a moment and talk about what feeds this that is to say this enormous amount of data and your particular passion Open Data talk a little bit about what open data was to begin with and where it exists now and why you're so passionate about it well this is one of those things where it was kind of a necessary ingredient for a number of us if we were going to build one of our interests was this whole idea of building up a web of connected data what if you could get this data just as we have it with a web of documents connected up in some way and cheat the web as if it was a large distributed database that could be powerful where you had to get the data on there in the 1st place what kind of data would you have possibly data about when the trains are on or where the bus stops all. What hospital procedures are being produced or what prescriptions are being what drugs are being used by your health system. Tons and tons of information which public services Initially we were thinking in those terms collect use isn't to do with personal information isn't to do with me necessarily but it