Podcast transcript

Jump to the sections in the podcast that you are most interested in.

• BR: Prof Benjamin Rosman
• TS: Tim Spira, head of digital content, Investec
• DM: Devina Maharaj

• 00:00: Intro

  Benjamin Rosman: So, we're kind of at the point now we can take any reasonably focused task and probably automate it to better than human levels and this is being done like all the time now, as in anything from chest x-rays to playing video games.

   

  Tim Spira: Hello everyone and welcome to the latest session of Investec’s Focus Talks, a series of candid conversations with leaders, influencers and change makers. I’m Tim Spira, head of Digital Content at Investec, and I must begin by declaring an interest in today’s talk. It’s a disclosure I’m very proud to make. Benjamin Rosman is my first cousin.

   

  Now I remember engaging with Benjy over the dinner table on some of the trickier issues in philosophy of mind when he was barely a teenager. Having studied philosophy, I no doubt held forth with a bit too much swagger, but even then, I was disarmed by his insights and questions. And when I tried the same trick a couple of years later and I quickly realised that I was way out of my depth.

   

  Fortunately for me though - and for many listening in – Benjy is not only brilliant but also has a knack for distilling the complexity of his field — artificial intelligence and machine learning — into accessible clarity.

   

  Benjy’s also able to engage with the ethical and even existential issues thrown up by this relatively young science. And, perhaps most impressively, he has made great strides in building a community of experts in AI in Africa.

   

  In 2017, he started the started the Deep Learning Indaba, an annual gathering of professional researchers in the field of deep learning— the study of algorithms that learn by trawling through gigantic data sets. That gathering is now the largest of its kind not just in Africa, but globally.

   

  In March 2018 Benjamin received a Google Faculty research award — the first in machine learning in Africa. He also got behind the critical imperative to introduce more women into AI research – a cause he remains passionate about.

   

  Benjamin has been lecturing in the school of computer science and advanced mathematics at Wits University since 2017 where he was recently made a professor.

   

  His work is published in leading journals both locally and internationally and he has presented at conferences at such institutions as Brown university, Stanford and his Alma Mater, the University of Edinburgh, where he earned his PHD. He also recently became a faculty member at the Singularity University.

   

  Now lets listen in to professor Benjamin Rosman in conversation with Investec’s Head of Digital in South Africa, Devina Maharaj.

   

• 2:19: What is Artificial Intelligence?

  Devina Maharaj:  So I must say even as the Head of Digital, when it comes to AI and deep learning and machine learning and artificial neural networks, this topic scares me and I was quite intimidated I think, just hearing Ben's, like achievements just it's, it's mind-blowing I think just to have him here today. So welcome. 

   

  BR:  Thanks, it's good to be here.

   

  DM: Let's just kick off and start off around what is Artificial Intelligence. I think we hear the words Artificial Intelligence, machine learning, deep learning, can you give us a brief summary, in like I often say in like the Davina's dummy guide. What does that look like?

   

  BR: Cool. Yes, I think AI is everything we haven't done yet. So if you looked at like this crazy sci-fi idea of a phone that you could just talk to you and tell it what to do ten years ago that would have been AI, now we're like naah it's just SIRI.

   

  But really we tend to think of it as like a broad umbrella term for a whole lot of different sub-disciplines and it's really just about making machines do intelligent things, but the kind of modern incarnation of this which is all the ideas around machine learning, which I guess we'll talk more about, is trying to be able to do that from data. So, I want a system that as it gets exposed to more data or more experience with the world or something like that, it gets better at some tasks that it's trying to solve.

• 3:45: Supervised and unsupervised learning

  DM: And in my research, trying to sound very smart and in this conversation, I discovered that machine learning has different aspects to it like supervised learning and unsupervised learning. Can you tell us a little bit of, a bit more about that and what does it actually mean?

   

  BR: In AI that there's this thing we now call good old-fashioned AI and that's this idea that you've got a whole lot of formal statements about the world, and it's very powerful, you can do a lot with that, but it breaks down with the kind of irregularities that you find in the real world.

   

  So the challenge was really that, if, like essentially if you're trying to write these kinds of programmes to solve difficult problems, like whether it's recognising a face or translating texts or anything like that, it's really hard to come up with the rules of what, like defines that problem.  But on the other hand, it's quite easy to come up with examples.

   

  So if I wanted to do face detection, right, the kind of old way of doing this would be I'd have to come up with the rules like a face has two eyes over nose over a mouth and then it breaks if someone's head is on the side or if they're a one-eyed pirate.

   

  And then you still have to come up with the rules of like what makes an eye, like how do you describe an eye in an image? But it's really easy to collect millions of pictures of faces and millions of pictures of things that aren't faces, and this was like the first core idea of machine learning and this is, going back to your question, what we call supervised learning.

   

  So the idea here is I can give a whole lot of examples of the thing I care about and of other things and essentially instead of me writing the code to solve the problem, I get the machine to figure out the code itself that does that. 

   

  There's a few other areas, and the main one there is this other idea which is called unsupervised learning and there you don't know the true labels so you might have for example a whole lot of images and I don't know what's in them, but I want my system to do is look at this data and try find some sort of structure in it.  

• 5:44: Reinforcement learning

  BR: And then the third major category, which was a very much a niche thing out of the way, nobody cared about until about a year or two ago, which is actually a thing I've been working on for a while, is this area called reinforcement learning and that's about what I think is really the interesting AI problem, and that's about making long-term decisions.

   

  So, this area, reinforcement learning, is about asking the question of, if I get to make a whole lot of decisions, moving into the future - what's the right thing to do? 

   

  Now, this is a bit further away from kind of widespread application, but the kinds of questions you can ask there would be around getting a robot to navigate in some environment, or getting a computer to play chess or video games, or making long-term business decisions over the next 10 years, or treating a patient whose got a chronic condition coming into a clinic every month and you want to ask questions like should we change treatment or, or get different tests and that sort of thing.

• 6:42: Deep learning

  DM: And is this where it kind of leads into that area of deep learning where it’s solving much more complex problems?

   

  BR: So you got this quite a deep structured model that, and we call it deep learning because it's got lots of components to it, lots of layers, essentially, and what effectively you're doing here is solving what's been a bit of a thorn in the side of machine learning for a while and that was feature engineering.

   

  So if you wanted to classify the difference between say a dog and an elephant and the things you knew about it were the number of legs it has and the number of ears it has, you can take the smartest algorithm in the world or the smartest person in the room and they'll be useless, right, because you don't have like the right features to make the distinction you want. But if you just knew it's weight, right, then like you don't need to be smart at like, any dumb algorithm could tell the difference between those things. 

   

  So, the challenge is finding the right features to answer the questions you care about. And the easiest way to visualise this is with images.  Right, so again in kind of slightly older-fashioned ways of doing machine learning to recognise faces you would have needed to go and manually write little bits of code that maybe look for ears or look for eyes like and maybe they don't even do it very well, but then you get a whole string of these features and then you take your image and pass it through those and now you've got a whole lot of information you can use to make your decisions.

   

  But here what you do is you just give it the raw pixels and it will learn the kind of intermediate level representations that are required to make the decision.

   

  So you learn these representations automatically rather than, from like as raw data as possible, and that's really the big idea behind deep learning. But it requires lots and lots of data which is why everything's data hungry these days because the models are very complicated.

• 8:28: AI in education and healthcare

  DM: Can you tell us a little bit about that in terms of some real-life practical examples of how this could change the way we kind of live in the world, which I think is super exciting.

   

  BR: We've been putting a lot of emphasis lately into particularly looking at problems around education and healthcare. We're building systems that, to learn to solve difficult technical problems need to be taken through a curriculum, that they can learn these skills, we should be able to customise curricula for students as well.

   

  So we're looking at questions around personalisation of education there, which says if you know someone's engaging with some educational material and you can learn something about the way they're doing it you can actually make recommendations for the order in which they should traverse the material to get a better understanding of what's going on.

   

  And we're also looking at similar kinds of questions in healthcare, as I said long-term treatment questions like diagnosis of different say, for example, lung conditions from chest x-rays, from blood smears, and also looking at ways to not just improve the diagnoses but also better understand why those diagnoses are being made and maybe find places where it would be useful to ask for other information, from like either the history of the patient or like other tests and that sort of thing.

• 9:48: The impact of AI on jobs

  DM: What would be the thing that you would say excites you most about this field and what frightens you?

   

  BR: Probably the same thing. So, there's a big question generally about what's called artificial general intelligence or in some hype sense artificial super intelligence. So, we're kind of at the point now we can take any reasonably focused task and probably automate it to better than human levels and this is being done like all the time now, as in anything from chest x-rays to playing video games.

   

  And now a lot of the challenge is around broadening what our systems can do. So if we can build a system that can solve one problem, can we, you know, without changing anything in the underlying infrastructure, now get it to solve a completely different problem and then a third one and hopefully as it does this it gains more just general experience and expertise and it finds it easier to solve other problems. This is kind of the big goal at the moment, is this idea of a broader or more general intelligence. 

   

  And that's just a fascinating question because it really ties back to who we are and how we work and how we learn things, but it’s also, I think, a potentially frightening thing I think to be working on for a number of reasons that mainly because humans might mess it up.

   

  DM: I know when we chatted, I had cited off this research that, that I had read about that said the, the, actual net effect of AI on the workforce would actually be a positive one. You looked at me like no, not really, tell me about that and what, what, what are your thoughts?

   

  BR: I mean the whole aim for a long time has been to try and automate what we used to call in robotics the dull, dirty and dangerous jobs.

   

  Obviously, it's a different kind of problem in a place with a high unemployment, but there are a lot of things that people shouldn't be doing for various reasons. And like if you can free people up from certain things, they can focus on stuff they really care about. An example always like is someone in HR probably goes into HR because they're interested in working with people, not interested in doing infinite paperwork.

   

  I also think a lot of the big challenges we face as a society are the kinds of things that work better when you throw more intelligence at them. So for example there have been projects on trying to optimise things like truck efficiency and there have been huge gains in these sorts of things that you can completely bring down the amount of fuel emissions, the amount of petrol these things need, just by running certain kinds of optimisation which is effectively a kind of machine learning.

   

  And so I think there's a lot of potential there and the flip side I think there's a narrative that's told a lot about how for every job that gets automated away another 1.74 jobs get created and that might, if that's true, which I'm doubtful of, then it's true just in the very immediate future.

   

  So, the two things people do are physical labour and cognitive labour and the whole physical labour thing, we've had robots in factories now since the 60's. 

   

  There's a lot of things obviously that are quite difficult and that takes longer, but a lot of the cognitive work we do is actually in a way much easier to solve. There's this thing called Moravec's paradox, it's basically this idea that stuff that we think is difficult is actually easier to automate, so things like some crazy complex thing you do in excel we can actually probably automate pretty easily, but walking is quite challenging.

   

  I mean, it's very exciting to be working in these things. But like fundamentally if you automate out a whole lot of these jobs, then the question is, you know, there's just fewer jobs and what should people be doing? And so personally I feel that this is a longer-term discussion we need to have as a society about, what is the value of jobs and my personal radical academic view on this is jobs are immoral and we shouldn't have them.

   

  And so, to me, actually jobs are a problem and we should be solving that and not the other way around.

   

  DM: What does that mean, jobs are a problem?

   

  BR: The reality is most people in the world do jobs that are terrible, like either physically, emotionally, psychologically and if you don't do this thing for like, I don't know, 8, 10, 12, 16 hours a day, at the end of the month you and your family die.  Like, that's pretty shit. Like, this is not a thing we should be proud of as a society.

   

  We’ve got this massive problem with unemployment and we've been struggling to figure out what to do about that may be the correct thing to do is find some way of saying like look, we've got machines that are coming along, maybe it's 20 years away, maybe it’s 50 years away, but we're pretty sure this is coming.

   

  And anything we care about we can get considerably more efficient at whatever it is ranging from farming to doing taxes and maybe instead of being scared of that we should embrace that and say, okay, so if there's a whole lot less work that needs to be done in society, right? What do we do about that? Do we wait for the day when you suddenly get 60% to 80% unemployment happening in the US and riots in the street and then go, ‘Oh we didn't see this coming’ or do we start preparing for it now and you know thinking about what is the world we actually want to live in, that is the best possible world we could build.

   

  If you look at for example, so I think the biggest employer in the U.S. is driving and we all know like autonomous driving is something that dozens of companies e working on, they keep promising it’s three years away and they've been promising this for about five years now, but, like it's very close and at some point you get, I mean when you get autonomous driving, then it doesn't make sense for people like Uber to keep drivers around, it doesn't make sense for big trucking companies to keep drivers around right?

   

  So now you're going to put like one of the biggest sectors completely out of work overnight because it makes financial and economic sense, and it's also safer, you've got something like 1.2 million people dying in car accidents every year, right? That's barbaric, but somehow it's okay because it's driving, right and again, so it shouldn't be like an ethical thing to let people drive cars, but now what are you going to do with all these people that make a living from driving? 

• 16:21: AI in Africa

  DM: What do you see as the biggest challenges for us living in an emerging market, dealing with kind of the same challenges as you would in a developed country where it's almost, you've got to, the leap is just far bigger, because for me that is where I think the rubber starts hitting the road in all of these conversations living in South Africa.

   

  BR: I think that a challenge with this kind of automation and bringing AI into the workplace is that it's really your kind of middle-class desk jobs that are most at risk, which means we should be very motivated to try and do something about this.

   

  So, having said that, I think there's also a lot of opportunities to make things better in Africa. So, I've been involved in setting up this kind of network across Africa and what's been really inspiring is just seeing the kinds of projects people are working on.

   

  So, we already know, like the kinds of leaps that like local societies have made with say cell phone technology and not even having to rely on landlines and that sort of thing.

   

  But also, apps that will, you can take a photo of your cassava plants and diagnose if there's diseases, that sort of thing. I think that kind of thing is very empowering as well, so I think there's that which is possibly different to the kinds of situations you would see in developed countries where there's a number of kind of immediate problems that people are dealing with on the ground that you could just basically solve straight away.

   

  DM:  You are doing some amazing work with like the Deep Learning Indaba and tell us the story and how that became this global forum that actually is leading the way in this?

   

  BR: So I did my masters and PhD in the UK and so reasonable size network of us that were abroad and at the time that was really my only option if I wanted to study further in these topics I had to go abroad and then came back with the vision. But we got chatting soon after I came back to a few other, many South Africans that either studied abroad and were now working here either at the CSIR or local universities, and then a few other expat South Africans that were now working mainly for Google. 

   

  And at the beginning of 2017 a few of us jumped on a Skype call and we thought no we need to do something. Let's have a little workshop and will like use our own money and fly the other ones that aren't in the country, fly them out here and maybe over a couple days, we probably find 30 to 50 people who actually would care about this in the country and just have a chat for a couple days. 

   

  And then we decided we would lose momentum, so let's do it right now and commit to doing this, this year. This was in 2017, and so we decided would host this thing at Wits and you know, it turns out a few months later we had an event of 330 people, we had over 750 applications, so we ran for six days.

   

  It was a hardcore technical week, like a long week, every day was jam-packed. We had a number of speakers coming out from all over the place. So, it was aimed at you know masters students, PhD students, post-docs, academics and faculty, and people in technical roles in start-ups and in industry.  

   

  So, Google gave us a lot of money and a few other big tech companies to completely pay for, I think we paid for 60 students from around Africa to come, to fly out here, we paid for everything.

   

  We've just had the third one, which was our first one outside of South Africa, it was at Kenyatta University in Nairobi. And there we had about 700 people and now we've got like UNESCO involved, we've got like parallel sessions on ethics, on policy, we're trying to get government's involved and open up discussions with governments about the way they should be doing these things.

   

  We had attendees from over 40 African countries and actually there's a number of other programmes we've now attached to this. The one that I'm very excited about because I run that, is what we call the Indaba X Program. It's kind of like the Ted-x thing.

   

  So, we had 13 events last year and then this year we had 27 events, which is half of Africa, and this included events running in Somalia, Sudan, DRC, Burkina Faso, like name it, we got stuff going on there.

   

  Partially on the back of this Google opened their first research lab in Africa, which is in Accra in Ghana. Google and Facebook are funding a, what they now call the most prestigious master's programme in the world, which is running in Kigali, and I was lucky enough to be able to go teach there for a bit and these students some of them have three Masters or put their PHD's on hold.

   

  Now a lot of them are getting jobs all over the world or into PHD programmes, the calibre of students is exceptional and there's a lot of interest now from a lot of the big tech companies.

   

  So where we actually trying to push harder now to get our local companies involved in this because otherwise you're never going to hire a good person ever again with the interest that's being shown and the support that's been given by all the big tech companies and that's been just incredibly inspiring. 

• 21:31: Existential issues raised by AI

  DM: I think that's super exciting to hear from an Africa perspective and well done on that. When change happens people often ask what does this mean for me? And, and, I think if you were sitting in the audience and had to ask yourself that, in terms of what does that mean for me, if I work at a bank or I'm working in a job that could potentially have some level of automation or have a machine doing this better than I do. What does that mean? What would your answer be? 

   

  BR: There's things to think about on a lot of different levels. Like you know, what should you be thinking about for your kids? And what should like, what should you think about in terms of your own trajectory through your career? And these are hard questions to answer mainly because the time frames aren't certain.

   

  I mean for me personally, one question that I think is quite useful is what do you actually get value from in your own life?  The two questions that we always get when we have these kind of radical, pseudo-socialist type discussions, is you know, what do you do for income, but what do you do for value? And I think a lot of people defend jobs by saying this is what gives people value in life and I think that's wrong.

   

  I think I mean it does give people value in life, but I think that's probably not the way to be living and maybe there is things to be thinking about like what would you actually do if you had this free time?

   

  There are more and more companies that are doing things like trying to reduce work hours and that sort of thing which I think is the right kind of direction to be going in.  But not very many people have the ability to enact those sorts of changes, but you know, thinking about what you'd like to be doing and actually keeping track of some of the changes that are happening in technology is another very useful thing to do.

   

  So we often advise when we're talking to kids that want to go into various areas of tech that if, at least the present state of the world is everything is driven by computers, they're not completely autonomous, but they do everything, and if you can't tell computers what to do, then there's questions about your own relevance there.

   

  So, like being able to communicate with what's happening in technology and with computers I think it's a very important thing to be able to do, but as I say, for the longer term, I think there's more, the questions are more philosophical. 

• 23:54: Singularity’s South Africa chapter

  DM: So, let's jump there because we, I know you're a faculty member of Singularity University. Do you want to talk a little bit about that and how it relates to AI and what it might mean in the future.?

   

  BR: Sure. The idea is, there's now chapters around the world, so South Africa is one of the newer chapters.  And there's a number of faculty members, I think there's 12 of us at the moment in, with different areas of specialisation.

   

  And really the idea is to act as a bit of a bridge between industry and various kinds of disruptors, whether it's in academia or start-ups or various different places and actually talk about some of the technologies and different ways of thinking that people are playing with, to, in a sense, try and help people prepare for the future, to think about ways they might embrace different kinds of technologies and actually just kind of foster better integration of some of these ideas into the into the real world.

• 24:53: Conclusion

  DM: I know one of the burning things and I'll ask the question and leave you thinking about it for me, but I asked Ben last week when I met him is that, in all the AI, machine learning, deep learning thoughts that we have, is there anything still magical about human beings?

   

  Is there something that makes us special as human beings and is it something that we can replicate? Consciousness, empathy and I know you have interesting views around it and all of us I think are waiting to kind of see where this goes. So, thank you. It was really, really interesting and definitely got our creative juices going.

   

  BR: This is great to be here, and this is a very cool format so thanks. 

   

  DM: Okay. Great. Thanks.

   

  TS: Thanks for listening to this Investec Focus Radio podcast. If you enjoyed this conversation, please take a moment to rate us and subscribe to Investec Focus Radio wherever you get your podcasts. And for more insights and expert views, please visit focus.investec.com

• Disclaimer

  Focus and its related content is for informational purposes only. The opinions featured on the site are not to be considered as the opinions of Investec and do not constitute financial or other advice. The information presented is subject to completion, revision, verification and amendment.