:00:00. > :00:07.Imagine a future where machines permeate every aspect of our lives.
:00:08. > :00:10.That reality is frankly already upon us.
:00:11. > :00:14.Artificial intelligence is in our homes, in our cars,
:00:15. > :00:20.But, how will artificial intelligence or AI changed the way
:00:21. > :00:25.we do business in Asia bastion Mark and what challenges will it bring?
:00:26. > :00:51.All of that and more on this edition of Talking Business.
:00:52. > :00:57.There is so much talk about artificial intelligence these
:00:58. > :01:03.And deep learning, another phrase that is often thrown about.
:01:04. > :01:06.We'll be speaking with a panel of experts in just a moment
:01:07. > :01:14.Artificial intelligence is a blanket term for any kind of technology
:01:15. > :01:20.or software that enables machines to perform tasks on their own.
:01:21. > :01:23.It is very much a part of our everyday lives.
:01:24. > :01:27.From something as simple as automated factories
:01:28. > :01:29.to intelligent personal assistance like Google Assistant,
:01:30. > :01:39.But AI is slowly evolving into something much more complex.
:01:40. > :01:47.Deep learning is one branch of machine learning that teaches
:01:48. > :01:50.computers to process information on several different
:01:51. > :01:54.layers at the same time, much like neural networks
:01:55. > :01:59.Part of that is for machines to learn for themselves
:02:00. > :02:02.by the accumulating and extracting patterns from huge amounts
:02:03. > :02:14.But the eventual goal is for machines to incorporate
:02:15. > :02:16.aspects of human intuition and emotional intelligence
:02:17. > :02:21.As we teach machines to be more independent of human beings,
:02:22. > :02:24.we are also asking questions about what could potentially go
:02:25. > :02:30.wrong, and more importantly, what all who the smart machines
:02:31. > :02:36.of the future will be held accountable to.
:02:37. > :02:46.Artificial intelligence or AI is already being used
:02:47. > :02:49.in many industry and business processing today.
:02:50. > :02:51.Deep learning within AI is really the next frontier.
:02:52. > :02:56.But what kinds of possibilities does this open up for businesses?
:02:57. > :03:00.With us we have Modar Alaoui the founder and CEO of deep learning
:03:01. > :03:10.It makes, amongst other things, Eyeris is a deep emotion recognition
:03:11. > :03:12.software that reads facial micro-expressions in real-time.
:03:13. > :03:18.Vasilios Vonikakis, a research scientist
:03:19. > :03:31.with the Advances Digital Sciences Center.
:03:32. > :03:34.And the chief technology officer of Coin Tech, dealing with deep
:03:35. > :03:36.learning and automation in the transport and
:03:37. > :03:38.Thank you, gentlemen, for joining us today,
:03:39. > :03:41.to talk about artificial intelligence and deep learning.
:03:42. > :03:44.I'd like to start with you, Modar. What is deep learning?
:03:45. > :03:46.How would you describe it to our audience?
:03:47. > :03:48.In image recognition, deep learning enables algorithms
:03:49. > :03:54.to recognise the same face from different angles,
:03:55. > :03:59.and in text recognition, it would try to identify the text
:04:00. > :04:06.In, for example, speech recognition, it would try to identify
:04:07. > :04:08.the same keyword or phrases from different accents.
:04:09. > :04:12.So, it needs a lot of data to be able to do this.
:04:13. > :04:15.Typically, a lot of data but a lot of quality data as well
:04:16. > :04:17.so not just quantity, but also the quality of the data.
:04:18. > :04:21.Yuris, in your business, can you explain how you use deep
:04:22. > :04:27.So, we do a couple of things, basically.
:04:28. > :04:32.One of the things we do is we do recoup addition of passing trains
:04:33. > :04:40.and different kinds of things about it, what is normal,
:04:41. > :04:42.what is not, starting from very simple things like numbers,
:04:43. > :04:45.which was initially how we started it.
:04:46. > :04:53.And, then, moving on to identifying different kinds of issues.
:04:54. > :04:56.Lately, we also use this to identify things which, from video feeds,
:04:57. > :04:59.what is happening in the city, like how many pedestrians, cyclists,
:05:00. > :05:00.cyclists with helmets, cyclists without helmets,
:05:01. > :05:03.with child seats, without trial seats in the back, transport
:05:04. > :05:05.all the information you can get from any video stream,
:05:06. > :05:07.transferring this into actually valuable real-life data.
:05:08. > :05:11.Vasilios, how do you see deep learning being useful in Asia?
:05:12. > :05:14.Because you are working here at the advanced science
:05:15. > :05:29.How far along has the research and the technology become advanced?
:05:30. > :05:31.Basically, there is a lot of research in this field
:05:32. > :05:34.going on in the local universities and in many different research
:05:35. > :05:37.I have seen a lot of start-ups, actually, utilising this type
:05:38. > :05:40.of technology, start-ups that have to do with observer recognition,
:05:41. > :05:45.in order to help shoppers find similar looking items from a big
:05:46. > :05:51.database and going directly and buying this item online.
:05:52. > :05:57.It has also been used in autonomous vehicles, a lot.
:05:58. > :06:02.This is a very ongoing research in this direction.
:06:03. > :06:07.Modar, where do you see the most advanced usage
:06:08. > :06:15.Image recognition still holds a number of challenges,
:06:16. > :06:22.simply because we cannot map the entire world in a very
:06:23. > :06:27.short period of time, nor can we do it that efficiently.
:06:28. > :06:30.So, I do believe that is an area that is also ripe for research.
:06:31. > :06:33.Not in Asia, but in the world in general.
:06:34. > :06:44.What I'm interested in is the deep learning that is trying to break.
:06:45. > :06:53.Because I understand having developed from machine learning,
:06:54. > :06:55.as a result of the limitations of machine learning,
:06:56. > :06:58.would it be fair to say that there are limitations
:06:59. > :07:01.I would say that deep learning is just the next
:07:02. > :07:03.step, not even close to full artificial intelligence.
:07:04. > :07:13.The limitations are set there are regarding the ideas
:07:14. > :07:15.about what are the tasks that we want deep learning
:07:16. > :07:20.Do we want to recognise images? Do we want to recognise speech?
:07:21. > :07:22.What kind of images are we searching for?
:07:23. > :07:24.So, it is not about making decisions.
:07:25. > :07:27.It is actually about doing specific tasks at a certain level,
:07:28. > :07:29.so precision is also important, to understand what is relevant
:07:30. > :07:31.and what is an acceptable level of precision, for example.
:07:32. > :07:35.So, to follow up on that, how is it being used in real life?
:07:36. > :07:38.If you can talk specifically to some industries, what kinds of functions
:07:39. > :07:40.or what kind of people is deep learning replacing?
:07:41. > :07:44.Yes, some examples of the things we are doing.
:07:45. > :07:48.Like, when you talk about on the railroad,
:07:49. > :07:51.passing over the train or the wagon between two companies or two
:07:52. > :07:56.motocross and points or something like that,
:07:57. > :07:59.we actually have a bunch of people running around doing
:08:00. > :08:09.different kinds of checks, visually checking it,
:08:10. > :08:21.doing this stuff 24 hours a day, seven days a week at all the time.
:08:22. > :08:24.So, what we do, we actually managed to replace them,
:08:25. > :08:27.to take the necessity of way, and this is a very simple
:08:28. > :08:29.and but there are still thousands of people who are actually
:08:30. > :08:32.So, Vassilios, currently, deep learning isn't taking over
:08:33. > :08:34.the entire supply chain of the labour force employed
:08:35. > :08:46.Do you see that changing in the future?
:08:47. > :08:49.Do you see more businesses employing deep learning and replacing people
:08:50. > :08:56.I think that I can see in the future a lot of jobs being displaced by AI.
:08:57. > :08:58.So, many people are talking about new
:08:59. > :09:04.Yes, for sure, but I think the rate at which this happens will not be
:09:05. > :09:07.as great as the rate with which the jobs become obsolete.
:09:08. > :09:10.So, for example, if you take autonomous vehicles,
:09:11. > :09:15.actually went this bull happen, you can say that a lot of jobs
:09:16. > :09:23.will be displaced there, like truck drivers, taxi
:09:24. > :09:30.Yes, I can add on top of this as well, a lot
:09:31. > :09:34.You talk about security guards in the malls and everything else,
:09:35. > :09:36.that is another industry where there are a lot
:09:37. > :09:38.of people working for it, they are considerably not
:09:39. > :09:41.the highest level of jobs, they are not that technology
:09:42. > :09:44.intensive, and these are very likely to be either replaced or optimised
:09:45. > :09:47.as part of applying more and more technology solutions.
:09:48. > :09:49.Modar, do you see this becoming a sort of long-running industry
:09:50. > :09:57.trend for the future or will deep learning, as we have been
:09:58. > :10:00.discussing, hit a barrier and another breakthrough has
:10:01. > :10:04.As in everything, technology will definitely be a barrier,
:10:05. > :10:15.Just like what we have seen with machine learning and other
:10:16. > :10:19.But I do have to say that deep learning is still in its infancy,
:10:20. > :10:23.there is still a long way for it to mature, and to disrupt a number
:10:24. > :10:26.of industries and lives before we can see it being potentially
:10:27. > :10:41.So, as it is now, deep learning requires a lot of data and a lot
:10:42. > :10:43.of good quality data, as you mentioned.
:10:44. > :10:48.So, as it is now, deep learning requires a lot of data and a lot
:10:49. > :10:51.of good quality data, as you mentioned.
:10:52. > :10:54.So, many people are working towards making this technology able
:10:55. > :10:56.to generalise better with less number of data,
:10:57. > :11:01.How can you train in the neural network with fewer examples in order
:11:02. > :11:07.to generalise as well as it will be with a large area of data assets?
:11:08. > :11:10.Another direction is how to make these deep learning networks learn
:11:11. > :11:13.by themselves unsupervised without having to tell them exactly.
:11:14. > :11:17.But on that point of unsupervised learning, we have seen incidents
:11:18. > :11:22.already just recently of when unsupervised learning
:11:23. > :11:29.or an attempt to make this kind of technology more independent,
:11:30. > :11:36.One example that comes to mind, for example, the hate tweets.
:11:37. > :11:40.That is an example of trying to educate a chat bot, if you will,
:11:41. > :11:47.to do a certain thing, has ended up having the adverse
:11:48. > :11:55.Supervised is here to stay they because at least we know that,
:11:56. > :12:03.you know, ground truth data is being labelled basically by human
:12:04. > :12:12.beings, therefore we are stirring up the level of accuracy there.
:12:13. > :12:15.Unsupervised is still going to have a certain number of question
:12:16. > :12:17.marks around it at least for the foreseeable future.
:12:18. > :12:19.There is a lot of effort being conducted there.
:12:20. > :12:24.One thing I wanted to add to this also is that it is not just that
:12:25. > :12:27.deep learning technology which has to advance, we also have to add arms
:12:28. > :12:41.The ability to interact with people, the ability to see the world,
:12:42. > :12:44.so we talk about all sorts of vision development which include also
:12:45. > :12:45.sound, image, video, whatever you are processing that
:12:46. > :12:50.So, it has to be hand-in-hand to become better, so it is not just
:12:51. > :12:54.Hold that thought, gentlemen, we are going to be talking
:12:55. > :12:57.about deep learning in just a few minutes again but before that,
:12:58. > :13:00.let's take a tongue in cheek look at deep learning
:13:01. > :13:07.Deep learning, it is the latest phrase that I'm going to impress
:13:08. > :13:15.But it will have applications that might actually surprise you.
:13:16. > :13:18.I've come to the seat of learning here at Trinity College Dublin,
:13:19. > :13:21.to find out about one of the surprising ways in which deep
:13:22. > :13:28.Deep learning, it is a great phrase, isn't it?
:13:29. > :13:31.It combines two wonderful words, learning, what
:13:32. > :13:36.You take away key learnings from the meeting, use it
:13:37. > :13:51.at the foot of the master, learning lots of wisdom and deep,
:13:52. > :13:53.Deep Thought was the famous computer in the Hitchhiker's
:13:54. > :13:57.Deep Blue was the first computer to beta human at chess,
:13:58. > :14:00.so combine the two of them, it has got to be good, right?
:14:01. > :14:01.It certainly sounds better than careful, now.
:14:02. > :14:04.Let's meet Dr Ahmed Salim at the Connect Research Institute
:14:05. > :14:09.He's developed, along with Dr Mohammed Elgary,
:14:10. > :14:11.a portrait painting technique which is based on deep learning.
:14:12. > :14:19.This is how Frank Gough would paint me.
:14:20. > :14:25.What we have here enables you to photograph and the van Gogh
:14:26. > :14:28.painting, the model by itself, produces these results.
:14:29. > :14:30.Look at me. A masterpiece.
:14:31. > :14:33.We are using these models to do feature extraction.
:14:34. > :14:36.Then we upload these features into our mathematical model.
:14:37. > :14:41.It will generate the painting that we are looking for.
:14:42. > :14:48.Professor Linda Doyle has a view on the bigger picture.
:14:49. > :14:53.Actually, it is true, the eyes do follow you around the room.
:14:54. > :14:55.Artificial intelligence is already actually embedded
:14:56. > :14:58.in the world around us, and people don't realise that.
:14:59. > :15:02.So, when they think of artificial intelligence, they tend to think
:15:03. > :15:05.about robots and science fiction and things they have seen on TV,
:15:06. > :15:08.whereas in fact in everything that you do, in Google,
:15:09. > :15:10.in Facebook, and in all the applications that
:15:11. > :15:12.you use, there is some kind of artificial intelligence there.
:15:13. > :15:17.If you have an algorithm that shows you how to avoid having an accident,
:15:18. > :15:25.there has to be some decision-making process that says A is more
:15:26. > :15:31.important than B so I'll bump into B rather than A.
:15:32. > :15:33.That have to be put in by humans somewhere.
:15:34. > :15:36.Yes, there is a lot of to think about when it comes to artificial
:15:37. > :15:38.intelligence and this iterative machine learning.
:15:39. > :15:42.So, when you think about artificial intelligence, what you are saying
:15:43. > :15:44.is that the algorithms running in the background that are learning
:15:45. > :15:47.behaviour and making decisions, but they don't do that in a vacuum,
:15:48. > :15:51.If you believe that human beings are a finite number
:15:52. > :15:53.of cells or whatever it is, trillions, eventually,
:15:54. > :15:56.we are going to make artificial intelligence that will be smarter
:15:57. > :16:00.We have to grapple with the ethical issues now because artificial
:16:01. > :16:02.intelligence isn't something of the future, it is
:16:03. > :16:09.Mm, where to immortalise myself on the wall?
:16:10. > :16:15.Of course, when you talk about anything like AI,
:16:16. > :16:18.there are still so many questions that remain because so much
:16:19. > :16:28.One thing I can say, now, though, is a rebuttal to anyone who ever
:16:29. > :16:34.If you like that, you can catch more of those videos
:16:35. > :16:43.Now, back to some serious business with our panel of experts.
:16:44. > :16:45.Today we have with us Modar Alaoui from Eyeris.
:16:46. > :16:48.Vassilios Vonikakis with the Advanced Digital Sciences
:16:49. > :16:58.Before the break, we were discussing the possibilities of deep learning.
:16:59. > :17:01.Now I want to talk about some of the fears that have cropped up,
:17:02. > :17:11.There is always this discussion about how machines
:17:12. > :17:17.Who is to be held responsible with the current technology
:17:18. > :17:20.that we have in place today for example, autonomous cars,
:17:21. > :17:22.if there is an accident when an autonomous car
:17:23. > :17:26.Maybe, Modar, you could pick that up.
:17:27. > :17:29.Yeah, well, there has to be a certain, you know,
:17:30. > :17:36.I do potentially think that there will have to be some sort
:17:37. > :17:42.I also leave that companies and governments and authorities
:17:43. > :17:51.will have at some point to decide how the AI should respond
:17:52. > :18:00.For example, the autonomous vehicle accidents, recently an announcement
:18:01. > :18:06.was made by one of the automotive companies, they made the choice
:18:07. > :18:11.for the AI that is going to be inside the vehicle to protect
:18:12. > :18:18.the driver at any given time, whether that protects the pedestrian
:18:19. > :18:25.or the person that is in front of the vehicle.
:18:26. > :18:27.the driver at any given time, rather that protects the pedestrian
:18:28. > :18:30.or the person that is in front of the vehicle.
:18:31. > :18:34.Who to blame and how to think about how AI would respond
:18:35. > :18:39.But doesn't this bring up a whole host of questions about how this
:18:40. > :18:42.kind of technology is going to be applied in daily life?
:18:43. > :18:47.Given the fact that currently it is on track to replace so many
:18:48. > :18:52.jobs, and we were talking about it before the break, it is going to be
:18:53. > :18:55.something that we are going to see more of in the future,
:18:56. > :18:57.shouldn't we be trying to answer these questions
:18:58. > :19:02.From my perspective, I would like to use a very
:19:03. > :19:11.One of the Ps would be for precision.
:19:12. > :19:14.That will differ from case to case, from application to application.
:19:15. > :19:15.Another one would be personal data protection.
:19:16. > :19:21.The third thing would be predictability which would be
:19:22. > :19:23.actually to do with the ability to predict something.
:19:24. > :19:27.And the fourth one would be the productivity issue.
:19:28. > :19:29.Where we say that machines can be more productive
:19:30. > :19:33.If you look at this perspective, then the question of ethics
:19:34. > :19:37.Because it is not just about autonomous vehicles,
:19:38. > :19:39.it is also about health care, it is also about diagnostics,
:19:40. > :19:42.and so, the way people perceive trust is actually changing
:19:43. > :19:49.And I think that is also one of the changes which isn't
:19:50. > :19:52.going to be about technology, isn't going to be about science,
:19:53. > :19:56.it is going to be one of the changes which will be about the perception
:19:57. > :19:58.of people, the way they see the world, the way they see
:19:59. > :20:01.the tech ologies they trust, and then of course we can talk
:20:02. > :20:11.principles about decision-making. into the deep learning,
:20:12. > :20:19.But the more complex systems get, the more demands there will be. How
:20:20. > :20:27.do we make the technology fair for everyone? Click for example Google
:20:28. > :20:33.along with Amazon and Facebook and Microsoft, if I were to mention
:20:34. > :20:40.these as the four or five horsemen companies, they have the ability to
:20:41. > :20:47.time they recognise that using it time they recognise that using it
:20:48. > :20:57.for the greater good of advancing humanity is of paramount role than
:20:58. > :21:03.to use it otherwise. What happens if or when an entity like this comes an
:21:04. > :21:08.independent freethinking entity? I know we are delving into the realm
:21:09. > :21:11.of science fiction here but, you know, it was far-fetched about five
:21:12. > :21:18.or ten years ago to talk about deep learning, who is to say it is deep
:21:19. > :21:23.-- far-fetched to talk about this? If you look at where we are now, we
:21:24. > :21:29.are still miles, not even Miles, millions of miles away from having
:21:30. > :21:33.an entity like this. Unless there is some hidden research which nobody
:21:34. > :21:37.knows about. Most of the research which is open is still focusing on
:21:38. > :21:46.specific tasks, it is still about being better in a very specific,
:21:47. > :21:52.narrow task. It is about being good about decision-making, being
:21:53. > :21:58.cognitive, making decisions on a larger scale, we are still very far
:21:59. > :22:01.away from this. Do you think we are getting closer to the stage where
:22:02. > :22:07.machines will be able to make decisions that currently humans do?
:22:08. > :22:13.Gradually, yes, but it will take some time. I expect that the things
:22:14. > :22:17.that we already have will become better and better over time.
:22:18. > :22:23.Usually, I think that weekend to overestimate what we can do in the
:22:24. > :22:30.short term. We tend to overestimate what we can do in the long run. I
:22:31. > :22:35.think, generally, AI will become better and better but it will always
:22:36. > :22:40.be applications specific, at least for the near future. Thank you
:22:41. > :22:44.gentlemen for joining us for this discussion on deep learning. That is
:22:45. > :23:03.it for this edition of Talking Business in Singapore.
:23:04. > :23:11.After a great and misty day, a murky night on the way tonight. Be wary
:23:12. > :23:12.and the Woakes, fog would