We sat down with Syen Nik, Head of Machine Learning at Jade, to learn what we could about machine learning and were duly impressed by his passion for the work he’s doing.
How would you explain what machine learning is to a 5-year old?
Animals. I would start with animals. Why? You can recognise animals because you've seen them many times in your books and in real life. Your parents no doubt read you book after book about farms and zoos and wildlife, not to mention the many renditions of Old MacDonald. It’s been amazing to watch my 2.5-year old doing just that. From not recognising anything to now being able to tell rough sketches of elephants or giraffes. Sure, he sometimes gets it wrong, but that’s all part of the learning process. He recognises the letter "M" as "W". It's quite funny when we bring him to McDonald’s.
So we train our machines just like we would a child, but we do it with far greater repetition. We start off with the basics, feeding it data to show what’s what, telling the difference between this and that. When it gets things wrong, we correct it, and look on like a proud parent when we get the results we wanted. After this, we add to its understanding by getting more and more complex with what data we feed it and the patterns and tasks we want it to learn.
As any parent will tell you, learning takes time and patience.
How can that five-year old help a machine learn?
Interesting ... in some respects a machine can be as intelligent as a toddler but it's only good at doing very specific tasks. On one hand, I think we can have machines see how toddlers learn over time by watching their behaviour and reactions.
For example, when kids fall down they get up and hopefully remember not to trip in that same spot again. Some parents are teaching their kids that if they act cute they get treats. So on one level, we could use computer vision to aid such machine learning. But where it gets really interesting is with something you and I often do on a daily basis, and this has something to do with passwords.
Passwords? That’s right. When we plug in our passwords on Google, for instance, we’re asked to select all the boxes that have storefronts in them. The comedian Michael McIntyre has had a lot to say about these. But there’s something else happening here. People are actually training ‘the machine’ to recognise storefronts from images. There are also traffic lights and dogs. It’s a smart play, and five-year olds can definitely help machines learn this way. It’s just finding the appropriate forum to do so.
What are some of the biggest issues facing machine learning?
I think the best way to answer this question is to quote Andrew Ng: "Worrying about AI evil super-intelligence today is like worrying about overpopulation on the planet Mars. We haven't even landed on the planet yet!" (Disclaimer: Andrew is my idol in the field).
One issue is that we're over-worrying about the threat of AI, which leads to reluctance when it comes to letting machine learning algorithms guide human decision-making that will affect human life. There are two ends to this. Machine learning practitioners like myself can do a better job of making it less scary. We can let others know we're not making decisions for them but helping them make better decisions.
The other end is that the general public can also be more open to the inclusion of such algorithms, essentially learning to trust the experts more. It’s a chicken and egg scenario but if we can pull it off, we’ll see more machine learning applications for the public good, rather than just for commercial use.
If you could get your hands on any data set in the world, what would that be and what would you do with it?
I love sport, and there are two questions I would love to have answers on: What makes Roger Federer evergreen and why is Rafa Nadal so dominant on clay? If I could get data on every ball hit, plus health data collected from the likes of fitness trackers and dietary intakes, I would analyse things like their footwork, their choice of shots, and much more. Similarly, there are some badminton players I would like to analyse. So if you’re listening out there Roger or Rafa or even the ATP, we’d love to look into this for you!
How is machine learning helping with digital employees?
Machine learning can help people using digital (next-gen chatbot) employees get to the end point more seamlessly. After some initial conversations, the algorithm can predict what the user is after and give the answer earlier.
We can also integrate machine learning to personalise the service offering for each user. With user profile and some initial chats, digital employees with a machine learning algorithm behind it can recommend which service is the most suitable for the user. And this can be based on as many variables as required.
What are the best examples of machine learning you've seen?
Where do I start? The ETA prediction for DB Cargo, because it’s the first piece of ML that will go live at Jade ;) Seriously though, DB Cargo is really happy with this and we’re seeing some great initial results. In the coming months we’ll be able to speak on this in greater detail.
Going global, it will have to be the big companies that have been really impressive. I like what Google has been doing, particularly Google Maps recommending restaurants based on my wife's past ratings (we use this a lot), and Google Photos being able to categorise photos into albums of things and people. Google is also getting into the whole idea of gamification to get data from people, which is becoming more popular.
I studied medical imaging so I'd also like machine learning applications in the field, such as those developed by Stanford University, which can diagnose diseases based on chest X-rays better than certified radiologists. The machine can do it in sub-seconds rather than minutes. Some people are worried this will cause loss of jobs, but can a machine deliver such weighty news with empathy and compassion? Not at all.
Where to from here?
So far we've just been discovering regular patterns within the data. It's amazing what we've been doing with it. As Judea Pearl sees it, the next big (10 years?) milestone would be for machines to understand cause and effect, rather than just correlations. Who knows what we can achieve with that?
I’m pretty excited about the future for artificial intelligence and the role we’re playing at Jade. So if I have one last piece of advice, it’s watch this space. Because if you don’t, the machines will.